Overview

Paper	Medical Objective	Categorization	Description
2024 - Zhao - Application of Physics Informed Neural Network for Breast Cancer Detection	Simulation of temperature distribution in 2D breast tissues to identify breast turmos.	Data: Radiologic Imaging (thermography) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	This paper used PINN to model the temperature distribution patterns of breast cancer thermography. The loss function was designed based off the heat conduction according to the Fourier heat equation. Initial conditions were assumed to be temperature distributions before the tumor manifestation. The PINN results were compared with FEM simulations which showed good accuracy.
2024 - Zarei - A Physics-informed Deep Neural Network for Harmonization of CT Images	Reduce variability in chest CT images by harmonizing the images.	Data: Radiologic Imaging (CT) Representation: Scientific Knowledge - Mathematical Models Integration: Feature transformation; Knowledge-regularized objective	Computed Tomography (CT) quantification is affected by the variability in image acquisition and rendition. This paper aimed to reduce this variability by harmonizing the images utilizing physics-based deep neural networks (DNNs). While there exist DL algorithms to harmonize images aposst-acquisiiont, these models require to be trained on ample data covering diverse imaging conditions, which is often not available. The authors simulated vritual CT images acquired under various imaging conditions using a vritual imaging platform with 40 computional patient models featuring lungs with different levels of pulmonary diseases. Then, an adversarial generative network was trained with a single CT slice and the modulation transfer functions (MTF, which quantifies the spatial resolution characterisitics o fthe CT system) of the scanner as inputs, and the harmonized CT image as output. Results showed that the harmonized images significantly enhanced image quality and quantification accuracy in CT imaging.
2024 - Yuan - Cell-ontology guided transcriptome foundation model	Construct a Transcriptome Foundation Model with integration of existing ontology knowledge.	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective	Existing Transcriptome Foundation Models (TFMs) use pre-training methods analogous to NLP like masked token prediction, treating genes as "tokens" and cells as "sentences". However, existing TFMs treat cells as independent samples and ignore their cell-type lineages. This paper proposed a way to leverage the existing ontology knowledge. The cell ontology knowleddge is downloaded from Open BIological and Biomedical Ontology databases as an unweighted directed graph, where edges denote hierarchical lineage relationship of the form "is a subtype of" between cell types. Personalized PageRank is used to estimate pairwise node structural similarities from the oncology graph. This foundation model is pre-trained with three levels of tasks: (i) gene-level masked gene prediction, (ii) intra-cellular cell type coherence, which is contrastive learning based on cell type classes, and (iii) inter-cellular level ontology alignment, which uses the pairwise node structural similarities computed from the ontology graph to contrast pairs of cells. The model is evaluated on a wide range of downstram tasks including zero-shot cell clustering, novel cell type classification, marker gene prediction, and cancer drug response prediction.
2024 - Voigt - Biologically-Informed Shallow Classification Learning Integrating Pathway Knowledge.	Develop classification models to distinguish Castration Resistant Prostate Cncacers (CRPC) from primary cancers	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	This work proposes an approach to build biologically-informed neural networks that is slightly different from the commonly used masking node/edge approach. This work uses the Generalized Matrix Learning Vector Quantization (GMLVQ) model as a the backbone shallow classifier. LVQ models identify a prototype set that minimizes the overall classification error with respect to the prototypes. The GMLVQ variant provides additional model interpretation possibilities beyond the standard vector quantization by providing a classification correlation matrix that indicates strength of correlations between features that contribute to the class discrimination. To incorporate the knowlege, the authors adjusted the matrix structure in GMLVQ according to the pathway knowledge encoded in a series of knowlege matrices. This shallow biologically-informed network provides a (i) even easier interpretability compared to standard biologically-informed NN, and (ii) allows shortcut paths achieved by inserting additional vertices between connected layers.
2024 - Sainz - Exploring the potential of Physics-Informed Neural Networks to extract vascularization data from DCE-MRI in the presence of diffusion	Learn pharmacokinetic (PK) model parameters by fitting the voxel-wise contrast agent concentration time curves of DCE-MRI imaging data to provide diagnosis and monitoring of tumors	Data: Radiologic Imaging (DCE-MRI) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	The authors focus on an extension of a standard PK model called the diffusion extended Tofts model (D-ETM). This differential equation is based on the concept of effective diffusivity applied to biological tissues. The differential equation is incorporated as a soft constraint in the loss function to ensure that the PINN solution obtained by NN satisfies the physical laws described by D-ETM. To mitigate the issue that error in the PDE parameters were high, the authors used the residual based adaptive refinement method, in which the PDE residual is evaluated at a new set of collocation points randomluy sampled after a certain number of epochs. The PINN is tested in various experimental settings, including homogeneous vs. heterogeneous parameter distributions, noisy and incomplete data, showing that PINN overcome typical convergence issues when fitting the D-ETM to DCE-MRI data.
2024 - Ruan - Magnetic Resonance Electrical Properties Tomography Based on Modified Physics-Informed Neural Network and Multiconstraints	Develop deep learning models to retreive the distrbution of electrical properties of scanned tissues from the measured transmit radiofrequency in MRI	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	A fully connected NN is proposed to learn mapping from a measured transmit radiofrequency (RF) in MRI to the spatial distribution of electrical properties (EP) of scanned tissues. The PDE considered here is the Maxwell's equation to describe the relationship between the magnetic field and spatial distirbution of EP. In contrast to convention PINN that directly learn mapping fomr spatial coordinates to EP, the authors propose a PINN that learns the ampping from the residual of the central physical equation of convection-reaction and its spatial gradient and Laplacian to the spatial distribution of EPs. Multiple constraints were added to avoid the ill condition of the problem, including one soft similarity constriant between permittivity and conductivitiy, and another gradient constriant to mitigate reconstruction artifacts introduced by noisy measurement and numerical computation of spatial derivatives.
2024 - Rohani - Predicting transcriptional outcomes of novel multigene perturbations with GEARS	Predict transcriptional responses to single and multigene perturbations in single-cell RNA data from perturbational screens	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Model design - Graph models	The problem is formulated as: given unperturbed gene expresssion and perturbed gene expression, the DL model predicts the gene expression outcome. For each gene, the model generates initial embeddings for unperturbed and perturbed states. A gene perturbation similarity graph is constructed leveraging the pathway information contained in gene databases. A GNN is used to to process each separately into combined information. The output is combined across all genes in cross-gene MLP layers and fed into gen-specific layers, and finally prediction layers of perturbed state.
2024 - Rodriguez - Using Physics-Informed Neural Networks (PINNs) for Tumor Cell Growth Modeling	Predict the growth pattern of tumor cells	Data: Clinical Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	This paper focuses on two mathematical models used to describe population growth over time: Verhulst growth model and the Montroll growth model. Standard PINN was used to learn parameters of the two models on a small hamster dataset.
2024 - Raeisi - Mathematical modeling of interactions between colon cancer and immune system with a deep learning algorithm	Identify different immune patterns and cancer cell behaviors through estimating key parameters of the cancer models.	Data: Clinical Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	This paper used PINN to estimate parameters and numerical solutions for the fractional differential system that captures the dynamics of the interactions between colon cancer and the immune system. Then, the PINN was used to perform simulations to confirm the. effectiveness of dendritic cells in the control and treatment of colon cancer. The authors also looked into the equilibrium points of the model and their stability to drive further insights for cancer treatment.
2024 - Podina - Learning Chemotherapy Drug Action via Universal Physics-Informed Neural Networks	Predict drug toxicity or efficacy and individual response to the drug	Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	This paper focuses on quantitative systems pharmacology (QSP) models that are used to assess drug pharmacokinetics and pharmacodynamics before they go to clinical trial. First, the PINN is applied to learn the drug action of a cancer growth ODE. Then, PINN where used to obtain an approximation of the drug doses in three ODEs. Lastly, PINN is applied to learn the net proliferation rate of doxorubicin from in-vitro data after validating the method on similar synthetic data.
2024 - Perez - A Transformative Approach for Breast Cancer Detection Using Physics-Informed Neural Network and Surface Temperature Data	Detect the presence of breast cancer	Data: Radiologic Imaging (infrared camera) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	This paper reviews works that used PINN to model the forward and inverse heat transfer problems. The forward problem consists of predicting the temperature distribution, and the inverse problem predicts the thermal conductivity from known temperature data. PINN solving the inverse problem can be used for breast cancer detection from surface temperature data collected by infrared cameras. The PINN-SDSD model leverages the Pennes bioheat equation and patient-specific infrared temperatures to obtain solutions for the bioheat equation in breast cancer thermal transport.
2024 - Nevirosky - Explicit Physics-Informed Deep Learning for Computer-Aided Diagnostic Tasks in Medical Imaging	Develop cell foundation models for cancer diagnosis and prognosis	Data: Molecular Representation: Auxiliary Datasets - public Integration: Data - Transfer learning	This paper propose GeneBag, a cell foundation model pretrained on 1.3 million human single-cell RNA sequencing data points from PanglaoDB on a masked gene prediction task. The model is a modified BERT that processses input sequences from single-cell or bulk tissue RNA-seq data through an encoder with multi-head attention. This model can be fine-tuned on tissue-level bulk RNA sequencing data to pave the way for downstream applications in diagnostic and prognostics. Results showed good performance for tissue and cancer classsification as well as stage and survival prediciton using bulk RNA-seq
2024 - Lari - A Holistic Physics-Informed Neural Network Solution for Precise Destruction of Breast Tumors using High-Intensity Focused Ultrasound on a Realistic Breast Model	Predict the outcomes of high-intensity focused ultrasound (HIFU) therapy for breast tumor destruction	Data: Radiologic Imaging (ultrasound) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	HIFU is a technique that uses high-frequency sound waves to precisely target and destroy cancer cells without surgey. However, simulating the effect of HIFU on tissues is challenging due to the complex multiphysics phenomena. This paper presents a method to predict the outcomes of HIFU therapy using PINN. First, a 3D anatomically realistic breast phantom (ARBP) is generated from T1-weighted MRI to reflect the breast's anatomy. Then, equations about the pressure acoustics physics and bioheat transfer physics were considered in a PINN to predict heat conduction and generate a temperature map within the tissue domain.
2024 - Jiang - IRnet- Immunotherapy response prediction using pathway knowledge-informed graph neural network	Predict cancer patient's response to immune checkpoint inhibitors (ICI) treament	Data: Molecular Representation: Auxiliary Datasets - public Integration: Data - Transfer Learning	One main challenge is the lack of patient samples, which is tackled by pretraining on patient transcriptomics data and survival information from TCGA dataset. This was feasible because the authors identified a correlation between ICI response and durvival time through a Bayesian network model, and based on this correlation, they pretrained the model using TCGA despite it had no ICI response labels, then fine-tuned the model on ICI data. Pathway interaction importance was measured based on the attention weights for the GAT layer, and gene importance was measured on the learned weights of the feed-forward network layer, which mimics the gene-to-pathway membership. A website is available: https://irnet.missouri.edu/
2024 - Hu - Radiotherapy toxicity prediction using knowledge-constrained generalized linear model	Predict radiotion toxicity or Normal Tissue Complication Probability (NTCP) to develop certain complication given the planning dose distirbution during radiotion therapy	Data: treatment Representation: Knowledge from Experts - Quantitative Integration: Knowledge-regularized objective	Existing univariate analysis have the limitations of not considering the strong correlation of dosimetric variables because of the physics of dose deposition in clinically realistic dose delivery methods. Among the multivariate methods, geleralized linear model (GLM) are easier to interpret. This work porposed a knowledge-constrained GLM to translate three pieces of domain knowelge into non-negativity of the coeffiicents (increasing the dosimetric variable is associated with an increase or no impact for the risk of toxicity but must not decrease the risk), monotonicity (the greater the dose, the higher risk of toxicity), and adjacent simmilarity constraints (dosimetric variables with adjacent dose levels are highly correlated). The model can be solved with standard optimizers. The model was demonstrated on lung cancer and prostate cancer datasets.
2024 - Hossain - Not all tickets are equal and we know it- Guiding pruning with domain-specific knowledge	Knowledge discovery of gene regulatory relationships in breast cancer	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Training - Knowledge-in-the-loop	This paper proposed an iterative pruning-based approach that accounts for prior knowledge by scoring parameters in terms of their alignment with the prior knowledge. The knowledge is represented as a matrix between inputs and outputs. For example, the matrix may represent prior experimental evidence regarding the binding of transcription factors (inputs) and target genes (outputs). During the pruning, a tuning parameter combines the data-informed weight matrices with prior-knowledge-informed effect matrices. This approach is shown to encourage sparse gene regulatory dynamics that allow for easier interpretation.
2024 - Hossain - Biologically informed NeuralODEs for genome-wide regulatory dynamics	Knowledge discovery of gene regulatory network (GRN) that can explain temporal gene expression patterns.	Data: Molecular Representation: Knowledge from Experts - Quantitative Integration: Knowledge-regularized objective	This paper proposed PHOENIX, a NeuralODE based on Hill-Langmuir kinetics (models dynamic transcription factor binding site occupancy) to promote sparse, biologically interpretable representations of GRN ordinary differential equations. The activation f unctions are designed to resemble Hill kinetics. A trained NN block encodes the ODEs governing the dynamics of gene expression and hence encodes the underlying GRN. To tackle the challenge that NeuralODEs have multiple solutions, of which many are inconsistency with our understanding of the biological phenomona, the authors introduced biologically motivated soft constraints to regularize the solution search to be consistent with a prior. In this case, the prior is formulated as a simple linear model of chemical reaction networks that encode likely network structure based on prior domain knowledge.
2024 - CancerLLM- A Large Language Model in Cancer Domain	Develop a medical Large Language Model (LLM) specifically designed for cancer domain.	Data: Clinical (textual) Representation: Auxiliary Datasets - public Integration: Data - Transfer Learning	The existing CancerBERT primarily focuses on breast cancer leaving other cancer types relatively unaddressed. This work introduces CancerLLM, a state-of-the-art LLM with 7B parameters specialized for the cancer domain with capabilities in cancer phenotype extraction and cancer diagnosis generation. The Mistral 7B was pre-trained using 2M cancer clinical notes and 500k pathology reports, then fine-tuned on three specialized datasets focusing on phenotype extraction and diagnosis generation. The CancerLLM was further fine-tuned using instruction learning for these tasks.
2024 - Alharbi - LASSO–MOGAT- a multi-omics graph attention framework for cancer classification	Classify 31 cancer types from multiomics data	Data: Molecular Representation: Knowledge graph Integration: Model design - Graph models	This paper proposed a graph-based deep learning framework that integrates messenger RNA, microRNA, and DNA methylation data to classify 31 cancer types and normal samples. mRNA or RNA-Seq data were processed using differential gene expression analysis. DNA methylation data were processed by fitting a Linear Models for Microarray (LIMMA) models. LASSO regression was used to select the most relevant features from multi-omics data (from the order of ten thousands to hundreds). The graph attention networks (GAT) was leveraged to incroporate protein-protein interaction networks. Each node represent a protein, node features include multi-omics data; edge represent interactions between proteins. So all patient's network have the same topology but with patient-specific node features. The proposed model achieved SOTA performance among multi-omics graph methods based on PPI networks.
2023 - Zhang - Accurate TCR-pMHC interaction prediction using a BERT-based transfer learning method	Prediction of TCR-pMHC binding for development of cancer immunotherapy	Data: Clinical Representation: Auxiliary Datasets - public Integration: Transfer learning	The authors proposes a TCR-BERT embedding model, a pre-trained architecture based on BERT. Training was done on over 113,000,000 CDR3beta sequeances from TCRdb database, coverting over 29 diseases and 10 tissues. The model was trained using masked language modeling tasks. The authors also propose a pMHC-BERT embedding model that share similar architecture as the TCR embedding model. Finally, the TABR-BERT model is trained to predct TCR-pMHC interaction. This approach mitigates the challenge of limited training data in TCR-pMHC binding prediction models.
2023 - Yin - SWENet- a physics-informed deep neural network (PINN) for shear wave elastography	Image reconstruction in Shear Wave Elastography (SWE)	Data: Radiologic Imaging (shear wave elastography) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	SWE is a new imaging technology that allows measuring mechnical parameters of soft tisssues in vivo and non-invasively. However, conventional SWE methods have the limitation of large errors in heterogeneous materials and when the cross-section dimension of a phantom tumor is smaller than ~1cm. DL-based SWE has the potential to provide a more accurate evaluation of the spatial distirbution of tissue mechanical properties. This paper proposed a framework to reconstruct image from SWE based on two PINN networks. The first NN takes temporal inputs and learns parameters that are in function of time. The second NN takes just spatial coordinates and learns the time-independent parameters. Both PINNS incorporates governing PDE of wave motion in elastic soft materials and the constraints. The model was tested on simulation data and ex vivo experiments on real tumors.
2023 - Xu - Knowledge-Infused Prompting Improves Clinical Text Generation with Large Language Models	Develop a Large Language Model (LLM) for clinical NLP tasks	Data: Clinical (textual) Representation: Auxiliary Datasets - public Integration: Data - Transfer Learning	This paper proposes ClinGen, a LLM for clinical NLP tasks. Main innovations involve clincal knowledge extraction and context-informed LLM prompting. For clinical knowledge extraction, the authors (i) sampled topics (nodes) from healthcare knowledge graphs such as iBKH KG, (ii) queried topics from ChatGPT3.5, and (iii) queried writing styles from LLM. This information is used to augment the prompts by sampling a topic and a writing style from the candidate set to add to the LLM instruction. The model is then fine-tuned on the downstream tasks. The model is evaluated on synthetic clinical data generation across 7 clinical NLP tasks and 16 datasets.
2023 - Wysocki - Transformers and the Representation of Biomedical Background Knowledge	Natural Language Inference in the context of a cancer precision medicine inference task to support Clinical s in the evaluation of the Clinical significance of findings compared to existing evidence	Data: Clinical Representation: Auxiliary Datasets - public Integration: Transfer learning	This paper compared two transfromers specialized for the biomedical domain (BioBERT and BioMegatron), both pre-trained on large biomedical text corpora (PubMed). The comparison evaluates their ability to capture entities, complex relationships, biomeidcal facts, etc.
2023 - Weiskittel - Network Biology-Inspired Machine Learning Features Predict Cancer Gene Targets and Reveal Target Coordinating Mechanisms	Predict dependency of genes to cancer lineage	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation	For each cancer type, the ARACNE algorithm was used to create regulatory networks for each cancer lineage in DepMap using Cancer Cell Line Encyclopedia RNAseq data. For each gene in each lineage-specific regulatory network, the following were extracted: (i) network features, (ii) number of cancer halllmark neighbors, (iii), sum of weights, and (iv) length of path to cancer-associated genes from the Cancer Gene Census.
2023 - Wang - Quantifying intra-tumoral genetic heterogeneity of glioblastoma toward precision medicine using MRI and a data-inclusive machine learning algorithm	Predict the regional genetic alteration of three GBM driver genes (EGFR, PDGFRA and PTEN)	Data: Radiologic Imaging (MRI) Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning	A novel Weakly Supervised Ordinal Support Vector Machine (WSO-SVM) was proposed to integrate unlabeled samples with imprecise/interval labels. The ordinal relationships between labels was considered in the formulation of WSO-SVM.
2023 - Wang - A Novel Hybrid Ordinal Learning Model with Health Care Application	Predict the regional genetic alteration of three GBM driver genes (EGFR, PDGFRA and PTEN)	Data: Radiologic Imaging (MRI) Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning	A Hybrid Ordinal Learner (HOL) was proposed to integrate samples with both precise and interval labels to train a robust Ordinal Learner (OL) model. The domain knowledge about imprecise labels of unlabeled samples, and ordinal relationships between labels is considered in the model. The model can be formulated as an optimization problem with ordinal constrains. A novel conversion method was developed that converts the HOL formulation into an equivalent formulation of learning a set of binary classifiers with coupled parameters.
2023 - Shoemaker - Bayesian feature selection for radiomics using reliability metrics	Selection of radiomic features	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Feature selection	The authors assign a prior distribution over gene coefficients, which allows more flexibilty. The inclusion of each feature is represented via a latent indicator. Prior knowledge about genes that are important or pairs of genes that are related can be used to specify this latent indicator.
2023 - Shi - MORGAT- A Model Based Knowledge-Informed Multi-omics Integration and Robust Graph Attention Network for Molecular Subtyping of Cancer	Classification of cancer types	Data: Molecular Representation: Knowledge graph Integration: Model design - Graph models	This paper prosed a framework for molecular subtyping of cancer based on the knowledge-informed multi-omics integration and robust graph attention network. The framework consists of four parts: (i) feature selection module for genomics, epigenomics, and trandscriptomics through differential analysis, (ii) removal of fake edges (far nodes) in the graph constructed with omics data based on the Gaussian distance between omics data, (iii) knowledge-informed multi-omics integration module, where omics data are sorted as a time series according to the "natural" sequence of genomics, epigenomics, transcriptomics, proteomics, and metabolomics, then integrated multi-omics data in two directions from the genome to the phenotype and from the phenotype to the genome using a bi-directional LSTM; (iv) compute the feature contribution of features on each sample.
2023 - Savchenko - Mathematical modeling of BCG-based bladder cancer treatment using socio-demographics	Predict tumor cell growth with personalized BCG models for improved treatment of breast cancer	Data: Clinical Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective	This study aims to improve the Bacillus Calmette-Gérin (BCG)-based treatment protocols by changing the “one-size-fits-all” approach with a more personalized one. This was done by integration of a machine learning component which is used to assess and adjust the model’s parameters over individuals and over time. Specifically, in the BCG model, the scalar parameters were replaced with functions that depend on time and the socio-demographics of the patient, and a term associated with the presence of immune cells based on the level of the patient’s activeness was added to the immune cell populations. The authors carefully designed a parameter fitting procedure to estimate the personalized model parameters overcoming the limited labeled samples (and explained right after):
2023 - Pratama - Physical restriction neural networks with restarting strategy for solving mathematical model of thermal heat equation for early diagnose breast cancer	Early breast cancer detection through surface temperature data	Data: Radiologic Imaging (infrared) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	This paper focuses on solving the two-dimensional heat equation's PDE for breast cancer detection (Penne's bio-heat transfer equation). A PINN was used to solve the PDE thermal analysis for normal and tumorous breasts to study the temperature behavior surrounding the cancer. A restarting strategy was designed where if the loss function does not decrease in several subsequent iterations, the process will stop and restart to the next cycle.
2023 - Perez - Thermal modeling of patient-specific breast cancer with physics-based artificial intelligence	Early breast cancer detection through surface temperature data	Data: Radiologic Imaging (infrared) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	This paper describes the application of PINNs to reconstruct digital models of breast cancer from infrared surface temperatures. The PINN uses random points to find the solution and solves the bioheat equation.
2023 - Peng - Boundary delineation in transrectal ultrasound images for region of interest of prostate	Segment prostate from transrectal ultrasound (TRUS) images for brachytherapy for prostate cancer	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Hyperparameter setting	Among the four common imaging modalities used for prostate diagnosis, TRUS has a higher dependence on prior information regarding shape features for prostate segmentation. This study proposes a semi-automatic contour detection that adopts radiologist-defined data seed points as a prior. The seed points are important because the segmentation algorithm (principal curve-based method) assumes that seed point is perfect without any abnormal points.
2023 - Park - Structured deep embedding model to generate composite clinical indices from electronic health records for early detection of pancreatic cancer	Early detection of pancreatic cancer from electronic health records	Data: Clinical (EHR) Representation: Scientific Knowledge - Probabilistic Relationships Integration: Model design - Customized model views; Feature engineering - Feature grouping	This paper proposes a deep embedding model that aims to reduce the dimensionality of the input variables in EHR by grouping related measurements as determined by domain experts. Results show that the grouped indices representing liver functions can be good indicators for early detection of pancreatic cancer. The deep embedding models are designed with three levels of hierarchy: (i) base model, which learns a reduced representation for each time-series varaible and concatenates them to project into a binary prediction space; (ii) combo model, which adds a grouping layer to group redundant variabless and geenrates 32 combo embeddings; (iii) and composite model that has another gouping layer to group relevant variables among 32 comboes. The second and third grouping strategies leverage a structure of variables determined by domain experts.
2023 - Mukhemetov - Physics-informed neural network for fast prediction of temperature distributions in cancerous breasts as a potential efficient portable AI-based diagnostic tool.pdf	Early breast cancer detection through surface temperature data	Data: Radiologic Imaging (infrared) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	This paper proposed a PINN that combines deep learning and physical principles to predict the temperature distributions in breast tissues and identify potential abnormal regions indicating the presence of tumors. The PINN model is normally trained by physics in terms of the residuals of the heat transfer equation, as well as boundary conditions with and without datasets of surface thermal imaging data concerning cancerous breast tissues, which can be used for future inverse thermal modeling to calculate tumor sizes and locations. The The PINN simulations are found to be 12 times faster than its FEM counterpart.
2023 - Meany - Predicting intratumoral fluid pressure and liposome accumulation using physics informed deep learning	Accurately measure spatial liposome accumulation and interstitial fluid pressure (IFP) distribution within a solid tumor for optimal treatment planning	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	Existing PINN works solve the forward problem. However, IFP is difficult to measure in vivo, so it is often more useful to consider the inverse problem: given the intratumoral accumulation of liposome post-administration, we want to derive the underyling IFP which resulted in that spatial distirbution. Such capability is important for forecasting tumor progression and desgining tumor treatment. This paper proposed a PINN for solving the inverse problem: predicting spatiotemporal accumulation of intratumoral liposome and spatial IFP in vivo from a single image of liposome accumulation post-administration and estimates fo the parameters present in the mathematical model. The PINN is based on a rely on a PDE to relate liposome transport and IFP within tumour tissue. The PINN learns the temporal process through the PDE and data provided at two time points, pre- and post-administration.
2023 - Meaney - Deep learning characterization of brain tumours with diffusion weighted imaging	Predict tumor progression curve and estaimte patient-specific tumor parameters of the PI model for patients with blioblastoma multiforme (GBM) for personalzed treatment	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	The PI model need two parameters to run. However, accurately estimating these two parameters is quite challenging. Existing appraoch use sumary measurements taken from imaging with several mathematical assumptions, which ignores granular measurements. The authors propose a deep learnning model to provide patient-specific estimates of the PI model parameters and simultaneously give a full forecast of the intermediate GBM progression curve. First, a cellularity map is obtained from the image using rules derived from known mapping between ADC nad tumor cellularity and the tissue class. Then, cellularity map is used as inputs of the deep learning model. PI parameters are incorporated into the customized loss, which is designed following concepts of PINN.
2023 - Mao - Weakly Supervised Transfer Learning with Application in Precision Medicine	Predict Tumor Cell Density for patients with GBM	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning	To build personalized models despite limited labeled samples are avaiable for each patient, this paper employed two strategies: (i) transfer learning from other patients with the same disease, and (ii) weakly supervised learning using weak labels derived from radiologist's expertise. Specifically, the enhancing ROI and non-enhancing ROI annotations were used to extract pairs of unlabeled samples that are expected to have ordering relationship in their labels. An additional term is added to the loss function to penalize the wrongly ordered weakly labeled pairs.
2023 - Kazemimogadam - A deep learning approach for automatic delineation of clinical target volume in stereotactic partial breast irradiation (S-PBI)	Accurately segment breast clinical target volume (CTV) in post-operative breast cancer radiotherapy	Data: Radiologic Imaging (CT) Representation: Knowledge from Experts - Qualitative Integration: Model design - Customized model views	CTV delineations are challenging because the extend of microscopic disease is not visualizable in radiological images. CTV is primarily identified by the clinicians to assure that tumor spread patterns are incorporated in the delineation process. Physician's contouring practice for CTV segmentations in stereotactic partial breast irradiation (S-PBI) involves deriving CTV from tumor bed volume (TBV) via margin expansion followed by correcting the extensions for anatomical barries for tumor invasion. The authors tried to mimic this practice in a 3D U-Net segmentation model. Specifically, the model takes in the CT image and the corresponding TBV mask as multi-channel inputs. U-Net is learnt to look for the location cues provided by the TBV to extract discriminative features. The expasion rules and anatomical barriers of the chest wall and skin boundary are also learnt to constrain TBV to CTV expansion.
2023 - Imperio - Pathologist Validation of a Machine Learning–Derived Feature for Colon Cancer Risk Stratification	evaluate whether pathologist scoring of a histopathologic feature previously identified by machine learning is associated with survival among patients with colon cancer.	Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Other	Tumor adipose feautre (TAF) was previously identified via a prognostic deep learning model developed with an independent colorectal cancer cohort. In study had 2 pathologists conduct semiquantitative scoring for TAF on colon cancer cases to verify the prognostic value of TAF for overall survival.
2023 - Cho - Interpretable Meta-learning of Multi-omics Data for Survival Analysis and Pathway Enrichmen	Survival analysis/prediction for patients with cancer	Data: Molecular Representation: Auxiliary Datasets - public Integration: Multitask learning	Proposed a meta-learning approach that uses various combinations of integrated omics datasets to train a prediction model of cancer survival. Results showed that this meta-learning design allowed finding robust functional and semantic relationships among genes.
2023 - Chen - TGM-Nets- A deep learning framework for enhanced forecasting of tumor growth by integrating imaging and modeling	Predict tumor growth	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	This paper proposed a PINN for prediction of tumor growth. The PINN is based on an equation of cellular microstructure and a diffusion-reaction equation for nutrient concentration. Fourier layers were added to extract features from the input images. Sequence learning was employed to re-train the same NN for solving the PDE over successive time segments while satisfying the already obtained solution for all previous time segments. Fine-tuning with physics was done for a reliable extrapolation. The model is trained using both synthetic and experimental data. This model is claimed to be able to make forecasts outside the training domain, for both early stage of the tumor and later development of the tumor.
2023 - Cao - X-LDA- An interpretable and knowledge-informed heterogeneous graph learning framework for LncRNA-disease association prediction	identify IncRNAs associated with prostate, colorectal, and breast cancer	Data: Molecular Representation: Knowledge graph Integration: Model design - Graph models	This paper proposed an interpretable and knowledge-informed graph learning model to predict lncRNA-disease associations (LDAs). First, lncRNA sequence similarity and disease semantic similarity are calculated based on corrrelations and domain knowledge. These similarity matrices are used to construct a graph representing the intricate topology among lncRNA, miRNA, and disease nodes. Then, graph patches, inspired by image patches in computer vision, are designed for each lncRNA-disease node pair to capture topological relationships and their features. The authors defined nine types of graph patches encompassing threee meta-paths and six diverse link modules. A CNN model takes graph patch features as input and convolution kernels are used to process the data for LDA prediction. A post-hoc attribution methods is used for interpretability.
2023 - Bogatu - Meta-analysis informed machine learning- Supporting cytokine storm detection during CAR-T cell Therapy	Predict adverse effects of chimeric antigen receptor therapies in cancer treatment	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature engineering - Feature transformation	Cytokine release syndrome (CRS) is one of the most consequential adverse effects of chimeric antigen receptor therapies in cancer treatment. This paper proposed a meta-review informed method for identification of CRS. Specifically, a set of biomarkers and their statistical information were extracted from the literature. This information is refered as a Knowledge Base (KB) of statistical evidence for CRS diagnosis. This knowledge is used to getting a matrix view of biomerker value similarities with respect to KB, so that ML algorithms with model the biomarker correlations not only within the psace defined by available data but relative to past discoveries as well. In addition, the KB was used to generate abductive evidence-based reasoning behind the predictions referring to the existing studies.
2022 - Xie - SRG-Vote- Predicting Mirna-Gene Relationships via Embedding and LSTM Ensemble	Predict potetial miRNA-gene interactions from large public datasets which might contribute to future diagnosis, prevention, and cancer therapy	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Graph models	The authors propose a voting system composing of three types of embeddings: (i) sequential embedding of miRNA and gene sequence via doc2vec, (ii) geometrical embedding of miRNA and gene via role2vec, and (iii) a graph convolutional networks embedding from miRNA-miRNA and gene-gene networks. Pairs of miRNA-gene were fed into LSTM and B-LSTM models to predict their relationships.
2022 - Wang - Biologically-informed deep neural networks provide quantitative assessment of intratumoral heterogeneity in post-treatment glioblastoma	Prediction of three gene modules: recurrent tumor tissue, treatment-induced changes, and normal brain tissue	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning	Proposed a two-stage model pipeline, including BioNet_Neu and BioNet_ProInf, to accommodate the hierarchical relationship of three gene modules relevant to GBM prognosis. Predictions are first generated for the gene module that has more supervision from biological knowledge. The outputs of this step then provided additional information for the interplay of the other two gene modules. The BioNet_ProInf is a multitask, semi-supervised learning model with a customknowledge attention loss.
2022 - Wang - A Novel Deep Learning Method to Predict Lung Cancer Long-Term Survival With Biological Knowledge Incorporated Gene Expression Images	Predict long-term survival for patients with lung cancer	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation	To overcome high-dimensionality, converted gene data into images incorporating pathway knowledge. Specifically, the knowledge provides hierarchical data where genes and proteins are root and gene expression are leaves. Used the hierarchical index sorting by mean across all samples to put gene expressions into a square matrix. Then, trained a CNN using four inputs: two gene expression images and two sets of Clinical data.
2022 - Shen - Identify Representative Samples by Conditional Random Field of Cancer Histology Images	1. Identify representative tissues within whole-slide images. 2. Classification of cancer prototype for colorectal cancer tissues. 3. Active learning -find representative patches to label to reduce annotation costs.	Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Knowledge-regularized objective	Recorded the spatial coordinates of each patch on the image grid. Then, used conditional random field (CRF) to model the probability of classes based on the classification output and spatial coordinates. The CRF model measures the spatial correlations to derive costs of different offsets and dynamically learns offsets that lead to locations of most representative patches. CNN was used as a feature extractor. The model can be used as feature extractor for any state-of-the-art classifiers. The model can also be used for active learning to reduce annotation costs.
2022 - Perez Raya - Thermal Modeling of Patient-Specific Breast Cancer with Physics-Based Artificial Intelligence	Learn parameters of a physics model (model the breast surface temperature from thermal properties of the tumor in order to detect tumor location)	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	Used physics-informed neural network to learn the parameters of a bioheat equation. Physics-informed neural network uses physical constraints, boundary conditions, and a PDE equations to define the loss function.
2022 - Peng - H-ProSeg- Hybrid ultrasound prostate segmentation based on explainability-guided mathematical model	Segment prostate from transrectal ultrasound (TRUS) images for brachytherapy for prostate cancer	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Hyperparameter setting	This study proposes a semi-automatic contour detection that adopts radiologist-defined data seed points as a prior. The model has three sub-networks. The first sub-network uses an improved principal curve-based methods to obtain seed points and their projection index. The second sub-network uses a differential eovlution-based neural network. The third sub-network uses the neurla network to explain the smooth mathematical description of the prostte contour.
2022 - Ma - DualGCN- a dual graph convolutional network model to predict cancer drug response	Predict anti-cancer drug response (sensitive vs resistant).	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation	Used chemical structure of drugs to represent each drug as a graph, where nodes are atoms and edges are connection between atoms. Used PPI network to represent each cancer sample as a graph, where nodes are proteins (genes) and edges are interaction between proteins. The two graphs as separately fed into GCN and their embeddings are concatenated and fed into MLP to predict drug response.
2022 - Liu - Multimodal Brain Tumor Segmentation Using Contrastive Learning Based Feature Comparison with Monomodal Normal Brain Images	Segment brain tumor from MRI.	Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Feature transformation	This paper used the MRI from healthy controls to help segment tumor for tumor brains. One challenge is that normal brain images are often monomodal and tumor brain images are multimodal. To tackle this, the authors used a "normal appearance network" to obtain a reconstructed "normal brain" image for the tumor brains that is comparable to the healthy controls. In detail, they trained IntroVAE using T1 image of healthy controls to obtain a lower dimensional representation of normal brain appearance, then projected the tumor brain image to this representation and obtain a reconstruction of "normal brain appearance". Then, in the segmentation network, they i) used a simple Siamese network to align the feature maps of the normal and tumor images in the normal regions, and ii) compared the feature maps of the normal and tumor images to identify low feature consistency regions as tumor regions.
2022 - Liu - MetBERT- a generalizable and pre-trained deep learning model for the prediction of metastatic cancer from Clinical notes	Predict cancer metastasis (all types of cancer)	Data: Clinical Representation: Auxiliary Datasets - public Integration: Transfer learning	Used four language models pre-trained on Clinical and scientic data (BioBERT, BlueBERT, Clinical BERT, PubmedBERT) and fine-tuned them using in-domain data (EHR Clinical notes).
2022 - Li - Coupling Deep Deformable Registration with Contextual Refinement for Semi-Supervised Medical Image Segmentation	Segmentation of anatomical structures from medical images (not targeted for tumor but can be applied to tumor and is related to cancer diagnosis)	Data: Radiologic Imaging Representation: Scientific Knowledge - Probabilistic Relationships Integration: Knowledge-in-the-loop	First, the authors applied an image registration model (deep diffeomorphic registration model) that allows regions predefined on templates to be overlaid on the image and produces a segmentation map of anatomical structures. Then, this segmentation map is refined by warping with the probabilistic atlas map and taking the nearest interpolation.
2022 - Lee - A Novel Knowledge Keeper Network for 7T-Free but 7T-Guided Brain Tissue Segmentation	Segment brain tissue from MRI (not explicitly for cancer but related)	Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Transfer learning	It is known that 7T images provide better insights for diagnosis. However, 7T MRI are not always available. This paper aims to build 7T-free but 7T-guided brain tiissue segmentation model. They used paired 3T-7T images to train a module that extracts 7T-like representations from 3T images using knowledge distillation and a discriminator. At inference time, the network is able to use only 3T images to do segmentation with 7T-image guided information.
2022 - Huang - MTL-ABS3Net- Atlas-Based Semi-Supervised Organ Segmentation Network With Multi-Task Learning for Medical Images	Segmentation of organs from medical images (not targeted for tumor but can be applied to tumor.)	Data: Radiologic Imaging Representation: Scientific Knowledge - Probabilistic Relationships Integration: Knowledge-in-the-loop	First, trained a segmentation network using labeled images. Applied the trained network on unlabeled data and computed a confidence map by comparing the segmentation results with the probability atlas. The high confidence pixels were selected and set as ground truth to compute loss for unlabeled samples.
2022 - Bhattacharya - Integrating zonal priors and pathomic MRI biomarkers for improved aggressive prostate cancer detection on MRI	Pixel-level classification/diagnosis of prostate cancer (normal, indolent, aggressive classes) from MRI	Data: Radiologic Imaging Representation: Scientific Knowledge - Probabilistic Relationships Integration: Hyperparameter setting	Zonal priors contain Clinical domain knowledge that aggressive cancer is more likely to occur in some zones. The zonal priors were formulated as mean probabilities of each class, conditional probabilities of each class given the zone and conditional probabilities of each zone given the class. The authors proposed two ways to integrate the zonal prior probabililites. One way is to to use it as a prior in the bayes decision model.
2021 - Zhou - Multi-task learning for segmentation and classification of tumors in 3D automated breast ultrasound images	Classify and segment tumor in 3D automated breast ultrasoudn images for breast disease diagnosis and surgical planning	Data: Radiologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Multitask learning	Since the benigh and malignant tumors usually have different shape characteristics, the authors assume that the segmentation and the classification task can help each other. They designed a multi-task model that simultaneously does segmentation and classification. Additionally, since ABUS volumes have severe noise and tumors in ABUS have ambiguous boundaries, they proposed a iterative feature-refining strategy during training to refine the feature maps using probability maps obtained from previous iterations.
2021 - Zhou - 3D multi-view tumor detection in automated whole breast ultrasound using deep convolutional neural network☆	Detect tumor from 3D images for patients with breast cancer	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Hyperparameter setting	(1) Defined the anchor size/aspect ratios for each of the three views based on the expected tumor size from each view. For example, tumors have closer to square shape in saggital view and elongated rectangular shape in axial and coronal views. (2) Also, tumors can be as small as occupying 2% of the full image which makes them hard to detect. So redesigned the VGG feature extraction network to better detect small tumors: i) added deconvolution layers to enlarge the last feature map, ii) concatenated feature maps that represented local and global textures.
2021 - Zhao - DeepOmix- A scalable and interpretable multi-omics deep learning framework and application in cancer survival analysis	Prognosis of lower grade glioma using multi-omics data	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	This work proposes DeepOmix, a deep learning model for cancer survivla analysis using multi-omics datasets. This network can incorporate prior biological knowledge of gene functional module networks as the function module layer (the second layer). The number of nodes and edges between the first and second layer are defined based on knowledge from tissue networks, gene co-expression networks, or prior biological signaling pathways.
2021 - Zhang - Predicting Anticancer Drug Response With Deep Learning Constrained by Signaling Pathways	Predict drug response for multiple cancer cell lines.	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	This work proposes consDeepSignaling, a deep learning model for drug response prediction. To reduce number of features and make the model more interpretable, 46 signaling pathways (from KEGG) and 929 genes were used to build the deep learning model. The layers corresponded to input layer, gene layer, pathway layer, and other hidden layers. Layer's connection also followed known biological knowledge.
2021 - Zangooei - Multiscale computational modeling of cancer growth using features derived from microCT images	Predict tumor and microvascular network growth patterns for breat cancer.	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-in-the-loop	To simulate complicated spatiotemporal and multiscale biological behavior of cancer growth, this paper used three multiscale image-based PDE models of tumor growth. The simulation performed at microscope, mesoscopic, and macroscopic scale. The model environment was formulated on a regular grid of lattice sites called tumor microenvironments. Knowledge was used to setup this environment with regulatory substances, diffusible factors, cancer cells, and pathways. A deep reinforcement learning framework was setup, with discrete policies that are functional map states. At microscope scale, an agent that seek to have a desired phenotype at each simulation time step. At mesoscopi sale, cells and vessels are represented as agents with their own phenotype. A neural netowrk predicts tumor cell growth.
2021 - Weiss - PILOT: Physics-Informed Learned Optimized Trajectories for Accelerated MRI	Find optimized trajectories for accelerated MRI	Data: Radiologic Imaging Representation: Scientific Knowledge - other Integration: Knowledge-regularized objective	This work aims to find optimized trajectories for accelerated MRI. Prior knowledge about the peak-current and the maximum slew-rate produced by the gradient coils were translated into geometric constraints in the optimization problem to ensure physically viable solutions.
2021 - Wang - Knowledge-Infused Global-Local Data Fusion for Spatial Predictive Modeling in Precision Medicine	Predict brain tumor cell density (TCD) by a Gaussian Process based model.	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective	Proliferation-invasion (PI) model is used to predict TCD. The difference between the predicted TCD from the proposed model and that from PI is constrained to be smaller than slack variables. (Soft constraints)
2021 - Sasse - Integration of multiomics data with graph convolutional networks to identify new cancer genes and their associated molecular mechanisms	Identify cancer-related genes from high-throughput molecular data	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation;	This paper proposed combining graph convolutional networks with protein-protein interaction (PPI) networks to predict cancer genes from multiomics pan-cancer data. The methods consists of (i) collecting average mutation rates, CNAa, DNA methylation, and gene expression changes from all genes across 16 tumor types annd concatenating into a single feature matrix, (ii) A network is constructed where nodes are high-confidence cancer/non-cancer genes and edges are known interactions between them from PPI. (iii) The GCN is trained to classify genes into predicted cancer and non-cancer genes. (iv) The most important features for the classification are extracted. (v) Genes are clustered according to their feature contributions to detect modules with important gene-gene connections in cancer.
2021 - Sanyal - Weakly supervised temporal model for prediction of breast cancer distant recurrence	Predict cancer recurrence for patients with breast cancer	Data: Clinical Representation: Auxiliary Datasets - private Integration: Weakly supervised learning	(Same model as Barnejee 2019). This work used a large set of Clinical notes to generate weak labels for training data, and trained a NLP model using the small, expert-labeled data and large, weakly labeled data. Used medical dictionary to clean the text data (combine related terms together, reduce sparsity).
2021 - Ramirez - Classification of Cancer Types Using Graph Convolutional Neural Networks	Classification of cancer types (33 cancers and normal)	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Graph models	Two different input graphs were generated, a gene co-expression graph (based on correlation between gene expressions) and a protein-protein interaction graph from the STRING database. These inputs are passed into a graph convolutional layer followed by hidden layers and a classifier.
2021 - Nave - Artificial immune system features added to breast cancer Clinical data for machine learning (ML) applications	Predict tumor size at different time points for breast cancer	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Feature transformation	This study aims to use ML to predict the tumor size for each time point based on the imaging data. A mathematical model was used to extract immunological features, which are then used as inputs into ML model together with Clinical features
2021 - Liu - TranSynergy- Mechanism-driven interpretable deep neural network for the synergistic prediction and pathway deconvolution of drug combinations	Predict efficiency of drug combination and serve as screening for downstream experimental evaluation	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation	In the drug response prediction, the input are two drug molecules in the drug combination and a cell line (gene expressions). The authors argue that using the chemical structure of drugs as features it is not straightforward to connect with the cellular mechanisms of drug action. Biological representation based on drug-target interaction profile is a better strategy to infer the drug representation vector. They used protein-protein-interaction network features. Random walk with restart algorithm was applied on a protein-protein interaction network to infer drug-target profile as the drug features. From gene expressions, the gene-gene dependency profiles were extracted.
2021 - Liu - Real time volumetric MRI for 3D motion tracking via geometry-informed deep learning	Reconstruction of accelerated MRI	Data: Radiologic Imaging Representation: Scientific Knowledge - other Integration: Customized model pipelines	In this neural network, the authors added an explicit geometry-module that encodes the k-space sampling patterns and known transforms between the k-space and image domains.
2021 - Li - Informed Attentive Predictors- A Generalisable Architecture for Prior Knowledge-Based Assisted Diagnosis of Cancers	Predict (classification) whether a sample is tumoral or normal for different types of cancers.	Data: Molecular Representation: Scientific Knowledge - Probabilistic Relationships Integration: Feature transformation	This model outputs prediction as a function of the data-driven prediction and knowledge-based prediction. They generated knowledge-based predictions based on the intuition that samples whose difference with normal samples is more similar to the mutation patterns known from knowledge are more likely to be tumoral. In detail: 1) Extracted a knowledge vector representing gene mutation rate for each gene from an exterior data source. 2) Generated a reduced version of original feature vectors to match the dimensions of the knowledge vector. 3) Computed a general representation for normal samples by projecting original features into 1-d vector. 4) For each sample, computed their difference with this normal sample vector. The similarity between this difference and the knowledge vector determines the likelihood of a sample between tumor, thus, generates a knowledge-based “prediction”. 5) The final prediction is a function of the prediction by original data and from this knowledge-based prediction.
2021 - Kaandorp - Improved unsupervised physics- informed deep learning for intravoxel incoherent motion modeling and evaluation in pancreatic cancer patients	Estimate three parameters of the Intravoxel incoherent motion (IVIM) model for diffusion-weighted imaging (DWI)	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks	Designed an unsupervised NN to predict the three IVIM parameters of interest. The model minimizes a physics-based loss that measures the difference between the original signal and “reconstructed” signal based on the IVIM model and NN estimated parameters.
2021 - Joshi - OncoNet- Weakly Supervised Siamese Network to automate cancer treatment response assessment between longitudinal FDG PET CT examinat	Classify the treatment outcome (preogression, resolution, stable) of patients with lung cancer	Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Weakly supervised learning	Since expert annotation is costly, this work used rule based heuristics to extract Standard Uptake Values (SUVmax) of metabolically active tumors from radiology reports. Then used differences in the SUVmax values between scans to derive the weak labels (progression, resolution, stable). This heuristic is based on the Lugano 2014 criteria for tumor evaluation, assessment, and prediciton.
2021 - Johnson - Integrating transcriptomics and bulk time course data into a mathematical framework to describe and predict therapeutic resistance in cancer	Predict therapeutic resistance in breast cancer by improved model calibration	Data: Molecular Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective	Mathematical models of tumor progression often include distinct subpopulations. Despite these models are calibrated to experimental data, single biomarkers do not capture longitudinal patterns. The study proposed an integrated model calibration scheme that integrates information from two data sources: (i) from longitudinal population data sampled at high temporal resolution for a number of doses, and (ii) machine learning outputs that estimate the phenotypic composition at three time points before and after treatment. The ML model was trained to classify phenotype of cells based on its transcriptome.
2021 - He - Cross-modality brain tumor segmentation via bidirectional global-to-local unsupervised domain adaptation	Segment brain tumors from multi-modal MRI for tumor diagnosis and treatment	Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Transfer learning	This paper proposed a framework that can segmenta brain tumors in one MRI modality (target domain) when trained on another MRI modality (source domain). To account for domain shift, they used a bidirectional module that learns how to generate synthetic image from one domain to another, and a global-to-local adaptor that aligns the real and synthetic images. Also, spatial-wise and channel-wise attention maps were used to transfer the most discriminate part of features through domains, assuming that silences of the images are consistent inter-domain.
2021 - Feng - Investigating the relevance of major signaling pathways in cancer survival using a biologically meaningful deep learning model	Predict drug response for cancer patients (breast cancer, lung, glioblastoma, skin)	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	The 46 signaling pathways from four types of cancer and 1967 genes were selected to design a customized deep learning model architecture. The input layer consisted two input features (gene expression and copy number variation) for each gene. Then genes were connected to the pathways based on known connections. Model was evaluated on four datasets of four types of cancer.
2021 - Erion - Improving performance of deep learning models with axiomatic attribution priors and expected gradients	Predict (regression) drug response for patients with acute myeloid leukemia (AML).	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective	The knowledge graph is used to regularizes the loss function so that relative feature contributions are consistent with the known probability of integration of gene (called graph prior).
2021 - Elmarakeby - Biologically informed deep neural network for prostate cancer discovery	Predict the disease sate (classification) of prostate cancer and better understanding of the relationship between Molecular features and the disease.	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	A customized neural network architecture is designed based on biological knowledge, where each layer and node have a specific biological interpretation, and only connections that follow known pathways are activated and others are zero-out.
2021 - Chuang - Convolutional neural network for human cancer types prediction by integrating protein interaction networks and omics data	Predict normal vs tumor samples for 11 cancer types	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation	A large number of PPIs of human proteins were collected from five public databases. Laplacian clustering approach was used to map PPI networks into 2D space, then combined with gene expression data, to generate normal and tumor sample images for CNN model training.
2021 - Chen - Modality-Specific Information Disentanglement from Multi-parametric MRI for Breast Tumor Segmentation and Computer-aided Diagnosis	Multi-modality image segmentation from MRI	Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Model design - Customized model pipelines	This paper proposed a method for multimodal image segmentation. The primary modality (DCE-MRI) is complemented by using supplementary modalities (T2w and ADC) with addititional modality-specific information to improve the segmentation. The framework comprises of three main steps: (i) two synthesis networks map the supplementary modalities to aligne with the feature spaces of the primary modality, (ii) information extractors that take in real and synthesized images to explore different cases of signal coherence across the modalities, (iii) modality specific attention module and modality trusty gating modues designed to improve the modality fusion and explore intra- and inter- modality relationships.
2021 - Azher - Development of Biologically Interpretable Multimodal Deep Learning Model for Cancer Prognosis Prediction	Predict patient survival risk scores for four types of cancers (colon, bladder, skin, liver)	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	First, pre-trained an encoder (VAE) that has biologically informed architecture. The nodes of the first two layers represent gene entities. The connections are initialized based on known biological relationships (CpG to gene, gene to pathway). After VAE is pre-trained, the encoder weights are transferred as initialization for an ANN to learn prediction model.
2021 - Anastopoulos - Patient Informed Domain Adaptation Improves Clinical Drug Response Prediction	Predict patient drug response using cross-domain data (cell line and primary tumor)	Data: Molecular Representation: Auxiliary Datasets - private Integration: Knowledge-regularized objective	Used Mean Maximum Discrepancy (MMD) as a loss term to encourage alignment of the latent space of cell line and tumor samples. Additionally, uses weight sharing for both cell line and tumor sample data neural networks. Represented each drug molecule as a graph (nodes are atoms and edges are bonds between atoms) then extracted latent representation via GCN.
2021 - A multiscale model of pulmonary micrometastasis and immune surveillance- towards cancer patient digital twins	Develop a framework to simulate the interactions between the immune system and te progression of micrometastases in the lung.	Data: Molecular Representation: Scientific Knowledge - Mathematical Models; Knowledge from Experts - Quantitative Integration: Model design - Customized model pipelines	This paper proposed a multiscale hybrid model that combines mechanistic models and machine learning to create virtual patient models, as a concept for Cancer Patient Digital Twin (CPDT). Several mechanistic models were used to model tumor growth, motility and infiltration of immune cells, and other biological processes. Machine learning methods were used in the parameter exploration for generating virtual patients. A patient classifier was trained to distinguish difeferent immune system responses based on the cancer cell population. This method is a preliminary concetp for building cancer digital twins.
2020 - Vu - Deep convolutional neural networks for automatic segmentation of thoracic	Segment thoracic organs‐at‐risk in radiation oncology for treatment planning of lung cancer, esophagus cancer, or breast cancer.	Data: Radiologic Imaging Representation: Auxiliary Datasets - public Integration: Transfer learning	Transfer learning is applied by using locking the weights from a VGG16 image classiﬁcation model trained on ImageNet. When adapting to the target dataset, activation function is modified, a batch normalization layer and a dropout layer are added.
2020 - Seninge - Biological network-inspired interpretable variational autoencoder	Knowledge discovery and better interpretation of biological mechanisms and the underlying gene regulatory networks (GRN) of single-cell GBM	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	Used a data-driven encoder to capture the complex patterns and a sparse knowledge-based decoder to allow explainability. In detail, the decoder is a single linear layer that zeros-out the connection between the latent layer and input layer if it belongs to a known biological abstraction. Also, the decoder is constraint to positive weights to maintain explainability of direction of biological activity.
2020 - Ozdemir - A 3D Probabilistic Deep Learning System for Detection and Diagnosis of Lung Cancer Using Low-Dose CT Scans	Screening and risk diagnosis of lung cancer	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning	Used expert-assigned scores as noisy labels to train a malignancy ranking network. This network ranks segmented nodules within each CT scan by their malignancy risk. The top-k nodules are then used to train a multitask classification network. (The intuition is that each CT scan may have a large number of nodules, the ranking helps reduce noise and focus on those that are more likely to be malignant.)
2020 - Oh - PathCNN- interpretable convolutional neural networks for survival prediction and pathway analysis applied to glioblastoma	Predict long-term survival for patients with glioblastoma multiforme (GBM).	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation	For each of the three omics types, organized data into a matrix where each column represent a pathway and each row represent a gene. Then, performed PCA on each matrix. Finally, combined PCs of all omics types into a pathway image ordering the pathways based on their correlation.
2020 - Liu - A Natural Language Processing Pipeline of Chinese Free-Text Radiology Reports for Liver Cancer Diagnosis	Classification of cancer vs normal for patients with liver cancer	Data: Clinical Representation: Knowledge from Experts - Qualitative Integration: Feature selection	Used expert knowledge to extract Clinical ly meaningful features (also called named entity recognition), which are then used for classification. First, used expert knowledge to derive a specialized lexicon (a list of Clinical ly relevant words and synonyms). Then, trained a language model using both original text and lexicon coded text to detect relevant entities.
2020 - Leung - A physics-guided modular deep-learning based automated framework for tumor segmentation in PET	Segmentation of tumor from PET	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Simulation	This method tackles the challenges of limited spatial resolution of PET and limited labeled data. Physic-sbased simulation is used generate synthetic images. Specifically, first- and second-order statistics for the intensity, size, shape, intra-tumor heterogeneity, and tumor-to-background intensity ratio are extracted. The kernel distribution of tumor descriptors are fed into a PET system modeling the major image-degrading processes in PET such as detector blurring with a 5 mm full-width-at-half-maximum (FWHM). Intra-tumor heterogeneity was simulated by incorporating unimodal variability in intensity values within the tumor and, for some tumors, by modeling the intensity distribution as a mixture model. Data from multiple patients were used to simulate patient heterogeneity. Tumor seed locations were manually selected such that generated tumors would appear at visually realistic locations. The simulated data is used to pre-train a U-Net. Real medical data is used to fine-tune the U-Net.
2020 - Lee - Cancer subtype classification and modeling by pathway attention and propagation	Classification of sample type and knowledge discovery/interpretation of important pathways	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Graph models	Pathways from KEGG are obtained as prior knowledge. Each pathway is represented as a graph. The gene expression profiles is amapped to the nodes of this pathway graph, where the node feature vector represent the numbe of genes in the pathway. The input graph is first passed to a GCN to capture localized patterns. One model is built for each pathway. Then, the pathway-level attention and ensemble-level attention are used to ensemble the features which are then used to produce final cancer subtype predictions.
2020 - Kuenzi - Predicting Drug Response and Synergy Using a Deep Learning Model of Human Cancer Cells	Predict drug sensitivity of tumor cells	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	Learnt genotype embedding using a network whose layers and connections are based on known biological processes. Learnt drug structure embedding using a CNN with the same number of output nodes as the genotype embedding so they can be concatenated.
2020 - Kozłowska - Mathematical model predicts response to chemotherapy in advanced non-resectable non-small cell lung cancer patients treated with platinum-based doublet	Predict anticancer treatment response to chemotherapy to find optimal treatment schedule for patients with	Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective	This study developed a method that combines mathematical mechanistic modeling with ML to estimate response to anticancer treatment for each patient. The authors replaced calibration of the mechanistic model using mixed-effect learning with calibration mainly based on a multivariate Gaussian-mixture model.
2020 - Ke - Identifying patch-level MSI from histological images of Colorectal Cancer by a Knowledge Distillation Model	Identify microsatellite instability (MSI) status in colorectal cancer for better prognosis.	Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Knowledge-in-the-loop	Most existing models just use the whole-slide label as groudn truth for each patch in the whole-slide image. However, this causes mislabeling of many local patches. To address the mislabeling problem, a self-distillation model was first used to separate data into high fidelity and low fidelity patches. Then, one-shot active learning was used where pathologists picked up the most representative samples from high-fidelity data and tagged them as clean data for noise-robust training. The self-distillation model reduced the efforts of pathologists.
2020 - He - Infusing Disease Knowledge into BERT for Health Question Answering, Medical Inference and Disease Name Recognition	Derive text data embeddings that can be used for downstream tasks including:1. consumer health question answering, 2. medical language inference, 3. disease name recognition. (Not just for cancer but I believe the diseases include some cancer)	Data: Clinical Representation: Auxiliary Datasets - public Integration: Weakly supervised learning	Used the structure of Wikipedia to extract “weakly-supervised passages” containing disease knowledge. Specifically, the title of the article is the disease name, and the title of the section is the topic aspect of the disease. Then, used these passages to create additional samples used in training.
2020 - Garcia - Transfer learning with convolutional neural networks for cancer survival prediction using gene-expression data	Predict progression of lung cancer	Data: Molecular Representation: Auxiliary Datasets - public Integration: Transfer learning	Pre-trained CNN on data of 31 cancers except lung cancer, then fine-tuned using lung cancer data.
2020 - Fortelny - Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data	Predict cell type from gene expression	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	Prior knowledge from signaling pathways and gene-regulartory networks are used to build knowledge-primed neural networks. Each node corresponds to a protein or a gene, and each edge corresponds to a regulatory relationship documented in biological databases.
2020 - Deng - Pathway-Guided Deep Neural Network toward Interpretable and Predictive Modeling of Drug Sensitivity	Predict drug sensitivities and undersand mechanisms of drug action	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	To reduce dimensionality, the authors selected 977 landmark genes. Inputs to the network include cell line feature nodes and drug target feature nodes. The architecture is customized by adding a layer of pathway nodes and their connections to input gene nodes. The rest of network are data-driven and fully connected.
2020 - Boso - Drug delivery- Experiments, mathematical modelling and machine learning	Simulate tumor growth and drug action for a proper modeling of drug delivery and efficacy in anticancer therapy	Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective	The drug-induced cell death rate is a parameter difficult to estimate experimentally. Machine learning was found helpful because it is able to catch the more complex physics and therefore to identify the necessary parameters. Neural networks was used to learn the relation between the involved quantities, by learning the mapping of input - output data.
2020 - Aguilar - A generalizable data-driven multicellular model of pancreatic ductal adenocarcinoma	Develop a data-driven computational framework that simulates the microenvironment of pancreatic tumor to understsand cell type-specific molecular interactions and cell-cell communications.	Data: Molecular Representation: Knowledge from Experts - Quantitative Integration: Model design - Customized model pipelines	This paper proposed an approach to model a block of cancerous tissue with a mixture of cancer, stromal, and immune cells randomly located inside a 3D rectangular space. Each cell has a Boolean Network (BN) that determines its cellular phenotype, the possible secretion of cytokines, and the state of membrane receptors. Within the cell, the BN contain n nodes, each with a perturbation probability. The cell-cell communication is modeled via diffusion of cytokines. The tissue is represented as a 3D point process of cells with a fixed density. Previous knowledge can be used to set constraints to this system. Model parameters were calibrated individually. This computational framework can be used to build patient-specific models given molecular data of their samples.
2019 - Zheng - Semi-supervised segmentation of liver using adversarial learning with deep atlas prior	Segmentation of liver in CT images (not targeted for tumor but can be applied to tumor.)	Data: Radiologic Imaging Representation: Scientific Knowledge - Probabilistic Relationships Integration: Knowledge-regularized objective	The probability atlas provides the prior statistical probabilty that the liver appears at each pixel location. The authors assumed that hard samples are those with atlas values close to 0.5, then the hard samples are given maximal weight in the loss function whereas the easy samples are given small weight.
2019 - Yazdjerdi - Reinforcement learning-based control of tumor growth under anti-angiogenic therapy	Control drug dosing in cancer antiangiogenic therapy	Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-in-the-loop	This work designed a RL-based controller for cancer antiangiogenic therapy by implementing the Q- learning algorithm to control drug dosing. The mathematical model of tumor growth under angiogenic inhibitor is simplified and used to represent a simulated patient and provide feedback to the controller.
2019 - Wang - Weakly supervised deep learning for whole slide lung cancer image analysis	Classification of cancer type into tumor vs non-tumor (lung0	Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning	The model is trained using both image-level class labels and a small number of coarse annotations of abnormal regions by pathologists. The cross-entropy loss is modified such that higher penalty is given when the CNN misclassifies patches that fall into annotated regions.
2019 - Wang - A dual-mode deep transfer learning (D2TL) system for breast cancer detection using contrast enhanced digital mammograms	Classify benign vs malignant tumors for breast cancer using the contrast-enhanced digital mammography (CEDM) data.	Data: Radiologic Imaging Representation: Auxiliary Datasets - public Integration: Transfer learning	Transfer learning is used in training the dual-mode DL model, in which a pre-trained Inception V3 on ImageNet is used to initialize the parameter estimation, and then the CEDM data are used to fine-tune the parameters.
2019 - Tomita - Attention-Based Deep Neural Networks for Detection of Cancerous and Precancerous Esophagus Tissue on Histopathological Slides	Classification of cancer prototype (skin cancer)	Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Customized model views	The model dynamically identifies ROIs in the whole-slide image and makes classification based on the selected ROIs. The model is composed of two modules, the first module extracts grid-based features and the second is attention module that generates an attention map used to weight the features, both trained end-to-end using image-level labels.
2019 - Shao - Weakly Supervised Deep Ordinal Cox Model for Survival Prediction From Whole-Slide Pathological Images	Predict survival of patients with multiple types of cancer (censored vs non-censored)	Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Feature transformation	Before classification, the samples (patches) were clustered into k distinct phenotype groups using Kmeans. Then generated a bag of patches by stratified sampling from the clusters. This intents to match the bag heterogeneity with WSI-level heterogeneity. The original label is binary. The authors compared the survival time (continuous) between pairs of patients to create ordinal labels of wether a sample survived long than another sample. They added a ranking loss to the objective function so the model can also learn from this ordinal information.
2019 - Sha - Multi‑Field‑of‑View Deep Learning Model Predicts Nonsmall Cell Lung Cancer Programmed Death‑Ligand 1 Status from Whole‑Slide Hematoxylin and Eosin Images	Predict the programmed death-ligand 1 (PD-L1) status from whole-slide images for nonsmall cell lung cancer (NSCLS) tumor samples.	Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Customized model views	Designed a NN with multi-field of view classification to mimic the pathologist using various zoom levels. The NN consists of one network that processes a large image and two additional branches that process small cuts of the central region of the image. This design ensures that the central region contributes more to classification.
2019 - Sevakula - Transfer Learning for Molecular Cancer Classification Using Deep Neural Networks	Classify any two tumor types with stacked autoencoder by (high dimensional) gene expression data.	Data: Molecular Representation: Auxiliary Datasets - public Integration: Transfer learning	A stacked sparse autoencoder is pre-trained by unlabeled data for feature selection and normalization for cancer classification on gene expression data. While classifying any two tumor types, data of other tumor types were used in unsupervised manner to improve the feature representation. Tested on 36 two-class benchmark datasets from the GEMLeR repository.
2019 - Nguyen - Incorporating human and learned domain knowledge into training deepneural networks- A differentiable dose-volume histogram and adversarialinspired framework for generating Pareto optimal dose distributions inradiation therapy	Predict Pareto optimal dose distributions in radiation therapy for prostate cancer patients with seven beam IMRT.	Data: Pathologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective	This work proposed a domain-specific loss, which was a differentiable loss function based on the dose-volume histogram (DVH), and combined it with an adversarial loss for the training of deep neural networks to generat Pareto optimal dose distributions in radiation therapy (for prostate cancer patients).
2019 - Nguyen - A feasibility study for predicting optimal radiation therapy dose distributions of prostate cancer patients from patient anatomy using deep learning	Predict optimal radiation therapy dose distributions (3D) of prostate cancer patients using U-net.	Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Other	U-net is used to predict fully 3D dose distribution prediction for prostate IMRT plans according to the treatment plans from physician.
2019 - Nguyen - 3D radiotherapy dose prediction on head and neck cancer patients with a hierarchically densely connected U-net deep learning architecture	Predict 3D radiotherapy dose on head and neck cancer patients by a hierarchically densely connected U-net.	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other	A deep learning-based dose prediction model, Hierarchically Densely Connected U-net, based on U-net and DenseNet is used to predict the 3D radiotherapy dose distribution of head and neck cancer patients.
2019 - Mercan - Assessment of Machine Learning of Breast Pathology Structures for Automated Differentiation of Breast Cancer and High-Risk Proliferative Lesions	Diagnosis (multi-class classification) of breast biopsy samples from whole slice images	Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Customized model pipelines	Used a sequential classification scheme in which a decision is made for a single diagnosis at a time to mimic the pathologists. The reasoning is features that describe one type of diagnosis do not apply to other diagnosis.
2019 - Ma - Incorporating Biological Knowledge with Factor Graph Neural Network for Interpretable Deep Learning	Predict Progression-Free Interval event	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective	(Very similar model as Ma, 2018) Computed a similarity matrix of learned feature representations (pairwise mutual information) and forces it to be consistent with feature relationships based on domain knowledge.
2019 - Ma - Incorporating Biological Knowledge with Factor Graph Neural Network for Interpretable Deep Learning	Predictor tumor stage for lung cancer (2 stages) and kidney cancer samples (3 stages); Classify tumor sample tyle (primary tumor vs normal)	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	Customized the network architecture such that i) each node corresponds to some biological entity, ii) only connect nodes that can reach each other in the knowledge graph within a number of steps. The later can be seen as an attention mechanism to capture hierarchical interactions among biological entities.
2019 - Liu - Automated detection and classification of thyroid nodules in ultrasound images using Clinical -knowledge-guided convolutional neural ne	Segment and classify nodule in ultrasoud images	Data: Radiologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Hyperparameter setting	Designs an expert-knowledge-guided multi-branch networkfor nodule detection. Similar to Faster R-CNN, the proposed network predicts region proposals for object detection. The authors computed the real distribution of nodule size and shape in training data to constrain the anchor boxes (lower feature maps to have smaller sizes and larger range of aspect ratios than those at higher feature maps). Based on how radiologists classify the challenging nodules (e.g. difference with surrounding tissies may reflect cell proliferation, malignant nodules tend ot have blur and irregular margin due to rapid growth), they proposed a triple-branch design to uses sonographic features known to be correlated with malignant nodules. The first branch is basic low-level features (takes original patches as input), second branch is context features (takes larger patches as input), and third branch is margin features (takes contour-masked margin regions as input).
2019 - Hu - Accurate Patient-Specific Machine Learning Models of Glioblastoma Invasion Using Transfer Learning	Predict tumor Cell Density for patients with primary GBM	Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Transfer learning	This work implemented both univariate and multivariate individualized transfer learning predictive models, built in a Baysian framework, which harness the available population-level data but allow individual variability in their predictions (same methodology as 79).
2019 - Gaw - Integration of machine learning and mechanistic models accurately predicts variation in cell density of glioblastoma using multiparametric MRI	Predict tumor Cell Density for patients with primary GBM	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective	This work presented a novel hybrid model, based on multiparametric intensities, which combines machine learning (ML) with a mechanistic model of tumor growth (PI) to provide spatially resolved tumor cell density predictions. The ML component is an imaging data-driven graph-based semi-supervised learning model .
2019 - Gao - DeepCC- a novel deep learning-based framework for cancer Molecular subtype classification	Classify colorectal and breast cancer subtypes, i.e. cancer Molecular subtyping	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation	Transformed original gene expression profiles to a vector of enrichment scores by doing a gene set enrichment analysis (GSEA ) for each sample’s genes on thousands of gene sets from public knowledge database. This new feature vector quantify activities of biological pathways and is used as input to classifier.
2019 - Gao - A feature transfer enabled multi-task deep learning model on medical imaging	Mass classification, mass detection, and segmentation for breast cancer patients	Data: Radiologic Imaging Representation: Auxiliary Datasets - public Integration: Multitask learning	This work proposed an multitask deep learning model enabled by feature transfer. The proposed model utilized the different tasks from the same domain.
2019 - Deist - Simulation-assisted machine learning	Predict anti-cancer drug sensitivity (unknown cancer type)	Data: Molecular Representation: Scientific Knowledge - Mathematical Models Integration: Feature transformation	Generated simulations of each sample using the feature vector, then computed the similarity between the simulation output of each pair of samples. This similarity, rather than original high-dimensional features, is used to build a kernel for use in kernelized ML models. The simulation was done multiple times controlling for the uncertainty of simulation parameters. Shown effective when number of samples < number of features.
2019 - Cheng - Network-based prediction of drug combinations	Identify effective drug combinations for hypertension and cancer	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Graph models	This paper built a network of nearly two thousand Clinical ly approved drugs. Drug-target binding profiles were pooled from multiple data sources and the proximity between two drug’s targets were measured based on the interactome between the targets of each drug. Once the network was built, the topological relationship between two drug-target modules reflect whether drugs are pharmacologically distinct, complementary, indirectly similar, similar, or independent.
2019 - Campanella - Clinical -grade computational pathology using weakly supervised deep learning on whole slide images	Classification of cancer type into tumor vs non-tumor (breast, prostate)	Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning	The slide-level diagnosis are weak labels for the tiles within a slide. If a slide is positive, then at least one tile must contain tumor. If a slide is negative, then all tiles must not contain tumor. This can be formulated as multiple-instance learning (MIL) problem. This work trained a CNN with weak labels at tile-level. Ranks the tiles by their probability of being positive. Then learns a RNN by sequentially passing top-ranked tiles from each slide to integrate tile-level classification into slide-level.
2019 - Banerjee - Weakly supervised natural language processing for assessing patient-centered outcome following prostate cancer treatment	Annotate Clinical notes with patient outcomes (three classes: affirmed, negated, discussed risk) for patients with prostate cancer	Data: Clinical Representation: Auxiliary Datasets - public Integration: Weakly supervised learning	This work proposed a method to annotate Clinical notes without requiring ground-truth annotations. They used three domain-specific dictionaries to derive example artificial expressions for each class, which were used create weak labels for training data. A supervised model is then used to learn from weakly labeled training data. The proposed knowledge-guided model is compared with data-driven generative models. Used medical dictionary to clean the text data (combine related terms together, reduce sparsity).
2018 - Turki - A transfer learning approach via procrustes analysis and mean shiftfor cancer drug sensitivity prediction	Predict drug sensitivity for patients with myeloma and lung cancer.	Data: Molecular Representation: Auxiliary Datasets - private Integration: Transfer learning	Transferred half of the auxiliary data from a related task to be trained with the original dataset. In detail, performed CUR decomposition on original data to extract a low-rank low dimensional representation. Then, adjusted each sample from auxiliary data towards target domain using 1) kernel average of their neighbors from original dataset, and 2) SVD of the two datasets. Finally, selected a subset of samples from auxiliary data to be trained with the original dataset.
2018 - Rathore - Radiomic signature of infiltration in peritumoral edema predicts subsequent recurrence in glioblastoma: implications for personalized radiotherapy planning	Estimate peritumoral edema infiltration regions using radiomic signatures determined via ML	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Other	A classification model was trained to classify voxels into belonging to near-tumor ROI (N-ROI) or far-from-tumor ROI (F-ROI). This classification proability is used as the infiltration score. The assumption is that N-ROI and F-ROI are expected to have relatively higher and lower infiltration. The estimated infiltration maps are compared to recurring ROI (R-ROI) and nonrecurring ROI (NR-ROI) regions drawn by the experts.
2018 - Ma - OmicsMapNet- Transforming omics data to take advantage of Deep Convolutional Neural Network for discovery	Predict the grade of LGG and GBM tumor samples.	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation	Transformed gene data into images based on their functional hierarchy. Specifically, used the hierarchical gene structure from KEGG database to build a five-layer tree, mapping the gene expressions according to the child nodes of the tree (a gene can exists at multiple nodes). Then, converted the tree into an image using the pivot method based on their adjacency in the tree and sorting the genes according to the median values across all samples.
2018 - Ma - Multi-view Factorization AutoEncoder with Network Constraints for Multi-omic Integrative Analysis	Predict Progression-Free Interval event (a binary outcome that indicates the patient had a new tumor event in a fixed period, implying the treatment outcome is unfavorable)	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective	Extracted a Molecular interaction network from external knowledge database. Used Laplacian regularizer to force consistency between the learned feature representation and the feature interaction network from domain knowledge. Additionally, used patient similarity networks to regularize the same set of patients are consistent with each other.
2018 - Lyu - Deep Learning Based Tumor Type Classification Using Gene Expression Data	Classify tumor type (tumor vs normal) for multiple types of cancer	Data: Molecular Representation: Scientific Knowledge - Other Integration: Feature transformation	Transformed gene data into images based on their relative position according to their chromosome numbers. Each rectangle represents one gene and the value is normalized gene expression level. The intuition is that adjacent genes are more likely to interact with each other.
2018 - Liu - Integration of biological and statistical models toward personalized radiation therapy of cancer	Predict treatment outcome (rectal complication) for patients with prostate cancer	Data: Clinical Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective	Proposed a general framework to integrate biological equations with patient-specific data. First, identified parameters in the biological model that are meaningful to be personalized, i.e. learnt from patient-specific data. Then, derived the optimization problem that predicts the response variable based on the biological model but uses personalized parameters estimated from patient-specific data. Knowledge was also used to define nonnegativity constraints of certain variables.
2018 - Hu - Accurate Patient-Specific Machine Learning Models of Glioblastoma Invasion Using Transfer Learning	Predict Tumor Cell Density for patients with GBM	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Customized model pipelines	Used Transfer Learning to build one model for each patient then biasing the model towards the population pattern using a Bayesian framework. Also, constrained knowledge transfer from patients with a certain threshold for correlation between one imaging measure and the response.
2018 - Hao - PASNet- pathway-associated sparse deep neural network for prognosis prediction from high-throughput data	Predict survival for patients with GBM	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks	The authors extracted 574 pathways and 4359 genes from public databases. A deep learning model called PASNet is designed using these pathways. The architecture consists of a gene layer, a pathway layer, a hidden layer, and an output layer.
2018 - Dincer - DeepProfile- Deep learning of cancer Molecular profiles for precision medicine	Predict drug response for patients with acute myeloid leukemia (AML)	Data: Molecular Representation: Auxiliary Datasets - public Integration: Transfer learning	Trained a VAE using large amount of public unlabeled data from other patients with the same cancer type to learn a robust latent feature representation. Then used these feature vector to train drug response model.
2018 - Babier - Knowledge‐based automated planning for oropharyngeal cancer	Predict 3D dose treatment plans (applied to oropharyngeal cancer patients)	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other	First, a knowledge-based planning (KBP) model (bagging query (BQ) method or the generalized principal component analysis-based (gPCA) method) was trained to predict the dose distribution. These predictions from KPB model were then passed through an optimization pipeline to generate the ﬁnal ﬂuence-based treatment plans.
2018 - Babier - Knowledge-based automated planning with three-dimensional generative adversarial networks	Predict 3D dose treatment plans (applied to oropharyngeal cancer patients)	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other	First, GAN was trained to predict the dose distribution given a contoured CT image. These predictions from generators of GAN were then passed through an optimization pipeline to generate the ﬁnal ﬂuence-based treatment plans.
2018 - Abid - Exploring patterns enriched in a dataset with contrastive principal component analysis	Find trends and variationsin gene expression levels of cancer patients that are useful t identify Molecular subtypes of cancer.	Data: Molecular Representation: Auxiliary Datasets - private Integration: Transfer learning	This paper proposed contrastive PCA (cPCA), which finds low-dimensional sturcutres that are enriched in the dataset relative to comparison data. The method is applied to data of cancer patients to find low-dimensional representations that separate them from the data of healthy controls from a similar demographic background. cPCA can be used as the feature extraction method and for denoising, where the backgroudn dataset serves to remove the universal but uninsteresting variation in the target dataset.
2017 - Yang - Literature-based discovery of new candidates for drug repurposing	Discover new candidates for drug repurposing	Data: Treatment/Clinical Representation: Knowledge from Experts - Quantitative Integration: Feature transformation	Assuming that extensive knowledge is hidden in the large-scale literature, the authors propose a relationship extraction method that allows the detection of indirect relationships for drug repurposing. Articles that discussed disease-gene, gene-drug, and disease-drug relationships were pulled from the MEDLINE database (5.4 million documents). A list of relevant entity names was generated from the Therapeutic Target Database (2k diseases, 3k targets, 20k drugs). The Comparative Toxicogenomics Database provided >1 million chemical-gene, gene-disesase, chemical-disease relationships, formatted as triplets. NLP parser tool were used to convert sentences of target documents into the smallest common subtrees in dependency format. Then, drugs are represented in a numerical vector space based on connections ot the linked genes. Then, promising repurposing drug candidates were identified by computing the similarity of repurposed drug to approved drugs using Jacard Index computed on their binary representations. In their experiments, the identified candidates were compared with the literature and ClinicalTrials.gov database.
2017 - Padmanabhan - Reinforcement learning-based control of drug dosing for cancer chemotherapy treatment	Control cancer chemotherapy drug dosing.	Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-in-the-loop	This work developed a RL-based control strategy for cancer chemotherapy treatment doses. A mathematical model of cancer drug dynamics was used to provide feedback to the RL-based controller.
2017 - Min - Predicting activities of daily living for cancer patients using an ontology-guided machine learning methodology	Predict activities of daily living disabilities for the first year after a patient’s cancer diagnosis	Data: Clinical Representation: Auxiliary Datasets - public Integration: Feature transformation	Mapped data into concept unique identifiers (CUI) of UMLS, which helps translate medical terms among diverse data sources. Identified base concepts, and from them, extracted hierarchies of related concepts. Applied AQ21 algorithm to learn generalizable rules that can predict disabilities.
2017 - Grady - Experience Based Quality Control in IMRT Treatment Planning of High Risk Post-Prostatectomy Prostate Cancer with RapidPlan	Develop a knowledge based planning (KBP) model for the treatment plan of high risk post-prostatectomy prostate cancer	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature transformation	An initial dosimetric analysis was carried out to identify high quality plans from Clinical trial. Then, Principal Component Analysis (PCA) is used to characterize the salient features of the patient anatomy (CT) and dose distribution (treatment plans) and Support Vector Regression (SVR) is used to model their correlation.
2017 - Geeleher - Discovering novel pharmacogenomic biomarkers by imputing drug response in cancer patients from large genomics studies	Predict drug response and discover novel biomarkers in Clinical cancer datasets; Classification of tumor samples for breast, blood, lung, skin, and central nervous system cancers.	Data: Molecular Representation: Auxiliary Datasets - private Integration: Transfer learning	PreClinical datasets are small but more straightforward to collect drug response. The goal is to use large Clinical datasets for pharmacogenomics discovery without having to collect drug response data. Used a preClinical dataset to train a model for drug response and gene expression, then applied the estimated model to impute drug response for the Clinical dataset of interest. This strategy assumes that if it is possible to predict drug response using cell line-derived gene expression, it should be possible for similar models.
2017 - Ferranti - The value of prior knowledge in machine learning of complex network systems	Predict patient drug response and phenotype (both as classification)	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature grouping	Unlike other papers that uses real data, this paper used synthetic data to evaluate whether using knowledge can improve performance of ML models. They used synthetic biological networks to generate data and compared models without knowledge (using data from all nodes to predict the super-nodes) vs. with knowledge (using nodes that are known to be directly connected to the super-nodes).
2017 - Esteva - Dermatologist-level classification of skin cancer with deep neural networks	Classify disease subtype for patients with skin lesion (3-way and 9-way classication)	Data: Radiologic Imaging Representation: Auxiliary Datasets - public Integration: Customized model pipelines; Data - Transfer learning	The CNN model is first pre-trained on 1.51 natural images and then fine-tuned on medical images. The authors used the Clinical taxonomy to partition disease subtypes into fine-grained training classes (as many as +700 classes) because training on finer classes showed improved classification. A deep learning model is trained predict the finer classes, then, aggregated the prediction of finer classes to recover the prediction of coarser classes. However, this method requires labels for all subclasses to be available.
2017 - Corredor - Training a cell-level classifier for detecting basal-cell carcinoma by combining human visual attention maps with low-level handcrafted features	Classification of cancer type into tumor vs non-tumor (skin)	Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Feature selection	This model extracts a visual attention map from actions performed by pathologists while navigating WSIs to inform the low-level feature algorithm where in the images to initially “look”. First, segmented all nuclei from the WSI and computed their low-level features. When pathologists look at the images, they can only see a window of reference instead of the entire image. Each nucleus is assigned a likelihood of cancerous based on the number of times the nucleus is visited by pathologists. Then, computed the importance of each feature by performing linear regression of features with the likelihood. The classifier SVM was trained using weighted-features.
2017 - Azizi - Transfer learning from RF to B-mode temporal enhanced ultrasound features for prostate cancer detection	detect prostate cancer using tempoeral enhanced ultrasound data	Data: Radiologic Imaging (ultrasound) Representation: Auxiliary Datasets - private Integration: Data - Transfer learning	The authors trained a tissue classificaiton model on tempoeral enhanced ultrasound radiofrequency data, and deployed for classification using B-mode data collected from the same ultrasound scanner. Unsupervised domain adaptation and transfer learning techniques were used. It starts with a layer of PCA to whiten the spectral features and align the source and target PCA subspaces. Then, a shared deep belief networks (DBN) was trained using both source and target data. The loss function minimizes the reconstruction error for both source and target domain data, and minimizes the divergence between the source and target distributions. Such capability is meaningful because the radiofrequency data is not readily available on all commercial scanners in clinics.
2017 - Ammad-ud-din - Systematic identification of feature combinations for predicting drug response with Bayesian multi-view multi-task linear reg	Predict anti-cancer drug response and identify which feature combinations are the most predictive	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature grouping	Extracted functional-linked-networks from prior biological knowledge and used these to select features and split the data into “views”. Then used a multi-view regression, where each feature has a feature-specific weight and a view-specific weight, to predict drug response.
2016 - Shiraishi - Knowledge‐based prediction of three‐dimensional dose distributions for external beam	Predict three-dimensional dose distributions for external beam radiotherapy	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature selection	Using previously treated plans as training data, this work proposed an artificial neural network (ANN) to predict a dose matrix based on patient-specific geometric and planning parameters, such as the closest distance (r) to planning target volume (PTV) and organ-at-risks (OARs).
2016 - Kim - Integrating Domain Specific Knowledge and Network Analysis to Predict Drug Sensitivity of Cancer Cell Lines	Predict anticancer drug response (resistance vs sensitive) of cell lines	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature selection	Selected features (genes) based on domain knowledge about which genes are apoptotic. Conducted literature search of prescribed Clinical dose of the drugs. Then used this dose to derive two cutoff values to obtain binary labels for drug sensitivity and drug resistance.
2016 - Ammad-ud-din - Drug response prediction by inferring pathway-response associations with kernelized Bayesian matrix factorization	Predict anti-cancer drug response and generate hypothesis between drug and pathway activation	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature grouping	Used prior knowledge to split data into components such that each group represented one pathway. Then used component-wise multiple kernel learning to predict drug response. The model learns component-specific kernel weights, which can be used to infer hypothesis of pathways (components) associated with drug responses.
2016 - Aben - TANDEM a two-stage approach to maximize interpretability of drug response models based on multiple Molecular data types	Predict anti-cancer drug response	Data: Molecular Representation: Knowledge from Experts - Quantitative Integration: Customized model pipelines	Used a two-stage approach to improve the interpretability of prediction models and have more balanced contribution of data types. Separated features into upstream data (somatic mutations, CAN, etc., significantly associated with response) and downstream data (gene expression). The downstream is trained on the remaining variability of the upstream model.
2015 - Zou - A Transfer Learning Approach for Predictive Modeling of Degenerate Biological Systems	Predict relationship between TF binding sites and gene expression (cancer related cell lines) by other related cell lines	Data: Molecular Representation: Auxiliary Datasets - private Integration: Transfer learning	This work proposed a unique prior for the model coefﬁcients of the old domains and new domain. This prior has two hyperparameters characterizing the degeneracy and the correlation structure of the domains, respectively. This work proposed to use a graph to represent the qualitative knowledge about degeneracy (this is knowledge about the features, used to generate a graph and regularize the model) and set the corresponding hyperparameter to be the Laplacian matrix of the graph.
2015 - Shiraishi - Knowledge‐based prediction of plan quality metrics in intracranial stereotactic	Predict achievable dose–volume histograms (DVHs) and highly precise DVH-based quality metrics (QMs) in stereotactic radiosurgery/radiotherapy (SRS/SRT) plans	Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective	This work proposed mathematical models to predict achievable DVHs based on the geometric relationship of organs-at-risk (OARs) to planning target volumes (PTVs) for plans with/without OAR involvement respectively.
2015 - Morino - Predicting disease progression from short biomarker series using expert advice algorithm	Predict progression of patients with prostate cancer	Data: Clinical Representation: Auxiliary Datasets - private Integration: Transfer learning	This model predicts for a target patinet with short, unstable time-series data of biomarkers, while transferring knowledge learnt from other patients with longer time series data who suffered from the same disease. Two datasets were used to construct "experts" (models): a dataset of Canadian patients and another of japanese patients. The experts are weighted based on the similarity of their observations with the target patient.
2015 - Krayenbuehl - Evaluation of an automated knowledge based treatment planning system for head and neck	Predict treatment plans for patients with head and neck cancer	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other	This work re-planned the treatments for patients with head and neck cancer with Pinnacle Auto-Planning (AP) version 9.10. Only one single cycle of plan optimization using one single template was allowed for AP. The dose plans were evaluated. Additionally, two experienced radiation oncologists blind-reviewed and selected preferred plans.
2015 - Cohen - Mathematical Modelling of Molecular Pathways Enabling Tumour Cell Invasion and Migration	Predict the probability of metastasis and understand how genetic interactions between gene mutations can affect the probability of metastasis	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Customized model pipelines	Constructed an “influence network” using prior knowledge. The network is composed of nodes (biochemical species or phenotypes) and edges (activations or inhibitory influence). This network is modeled with Markov decision process with pre-assigned state transition probabilities. Simulated random walks on the network to obtain probabilities to reach a phenotype. Also analyzed the stable states of the network for interpretation of meaningful pathways.
2015 - Chen - Enhanced Performance of Brain Tumor Classification via Tumor Region Augmentation and Partition	Classification of tumor category for patients with brain tumor	Data: Radiologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Feature transformation	Based on biological knowledge, tumors of the same category are often located in similar places, such as gliomas usually involve white matter; meningiomas are typically adjacent to skull, gray matter, and cerebrospinal fluid; and pituitary tumors are often adjacent to sphenoidal sinus, internal carotid arteries, and optic chiasma. The authors proposed "tumor region augmentation and partition", which (i) enlarges the tumor ROI to capture possibly informative context, and (ii) partitions it into ring-form subregions, features from each subregion are weighted and concatenated to get final features.
2014 - Xu - Weakly supervised histopathology cancer image segmentation and classification	Classification of tumor type (cancer vs non-cancer image), image segmentation (cancer vs non-cancer tissue), and patch-level clusteirng (different classes). (Colo cancer	Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning	This paper used Multiple Instance Learning (MIL) to learn patch-level labels when only image-level labels are available. Adds an additional term of the loss that encourage nearby image patches within each bag to be in the same cluster. Assumes that multiple cancer subtypes with different morphological characteristics might co-exists in a WSI image. This model integrates the clustering concept into the MIL setting by assuming that exists a hidden variable that denotes the cluster of each positive instance. One classifier is trained for each cluster and each cluster's instances are weighted differently.
2014 - Wu - Improved robotic stereotactic body radiation therapy plan quality and planning efficacy for organ-confined prostate cancer utilizing overlap-volume histogram-driven planning methodology	Improve robotic stereotactic body radiation therapy plan quality and planning efﬁcacy for organ-conﬁned prostate cancer	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other	A database containing Clinical ly-delivered, robotic SBRT plans of patients with prostate cancer was used as a cohort to establish an organ’s distance-to-dose model. The overlap-volume histogram (OVH)-driven planning methodology was reﬁned by adding the PTV volume factor to counter the target’s dose fall-off effect and incorporated into Multiplan to automate SBRT planning.
2014 - Tol - Evaluation of a Knowledge-Based Planning Solution for Head and Neck Cancer	Predict treatment plans for patients with head and neck cancer	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other	Knowledge-based plans were created by RapidPlan. RapidPlan results were compared with Clinical plans (CP).
2014 - Kwon - Combining generative models for multifocal glioma segmentation and registration	Segmentation of gliobastoma tumor from MRI	Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Simulation	This is a generative model for tumor segmentation. This model grows tumors on an initialized atlas, and estimates parameters to grow the tumor via the diffusion-reaction-advection model, and registers the scans to this atlas to infer segmentation. Tumor shape prior is estimated via random walk with restart and incoporated into the model.
2014 - Kim - Knowledge boosting- a graph-based integration approach with multi-omics data and genomic knowledge for cancer Clinical outcome prediction	Classification of ovarian cancer grade (low vs high)	Data: Molecular Representation: Auxiliary Datasets - public Integration: Graph models	This work proposed a graph-based framework for integrating multi-omics data and genomic knowledge to improve power in predicting Clinical outcomes and elucidate interplay between different levels.
2013 - Wu - Using overlap volume histogram and IMRT plan data to guide and automate VMAT planning A head-and-neck case study	Predict treatment plans for patients with head and neck cancer	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other	Based on comparable head-and-neck dosimetric results between planner-generated VMAT and IMRT plans, an inhouse developed, overlap volume histogram (OVH)-driven automated planning application containing a database of prior Clinical head-and-neck IMRT plans was built into Pinnacle 3 SmartArc for volumetricmodulated arc therapy (VMAT) planning.
2013 - Good - A Knowledge-Based Approach to Improving and Homogenizing Intensity Modulated Radiation Therapy Planning Quality Among Treatment Centers- An Example Application to Prostate Cancer Planning	Predict optimal radiation therapy dose distributions (3D) of prostate cancer patients by matching and deforming the cases in the (assembled) knowledge base.	Data: Pathologic Imaging Representation: Auxiliary Datasets - private Integration: Feature transformation	A knowledge database was created from treatment plans for prostate cancer at their institute. For each “query” case from outside database, a similar “match” case was identiﬁed in the knowledge database, and the match case’s plan parameters were then adapted and optimized to the query case by using ﬁnal dose-volume planning constraints.
2012 - Yuan - Quantitative analysis of the factors which affect the interpatient organ‐at‐risk dose	Predict intensity modulated radiation therapy (IMRT) plans for patients with prostate or head-and-neck (HN) cancer	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature transformation	This work formulized the dependence of organs-at-risk (OAR) dose volume histograms (DVHs) on patient anatomical factors into feature models which were learned from prior plans by a stepwise multiple regression method. Features were extracted by PCA.
2012 - Cun - Prognostic gene signatures for patient stratification in breast cancer- accuracy, stability and interpretability of gene selection app	Predict metastasis/relapse for patients with breast cancer and interpret which genes are most important for the prediction	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Other	Compared 14 published methods (8 using network knowledge) on 6 public breast cancer datasets. Found that the incorporation of prior knowledge could generally not significantly improve prediction accuracy but sometimes improve gene selection stability and biological interpretability.
2011 - Zhu - A planning quality evaluation tool for prostate adaptive IMRT based on machine learning	Generate dose volume histograms (DVH) of organs-at-risk based on prior plans	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature transformation	Principal component analysis was applied to DVHs and distance-to-target histogram (DTH) to quantify salient features. Then, support vector regression was implemented to establish the correlation between the features of the DVH and the anatomical information (organ volumes and DTH).
2011 - Wu - Data-driven approach to generating achievable dose-volume histogram objectives in intensity-modulated radiotherapy planning	Predict IMRT treatment plans for patients with head and neck cancer	Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other	The overlap volume histogram (OVH) was used to compare the spatial relationships between the organs at risk and targets of a new patient with those of previous patients in a database. Clinical plans were compared with OVH-assisted plans (OPs).
2011 - Chanyavanich - Knowledge‐based IMRT treatment planning for prostate cancer	Predict IMRT treatment planning for prostate cancer	Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Feature transformation	In this study, the quality of each of treatment plan was evaluated by comparing the dose–volume histograms of the new semiautomated plan to that of the original plan developed manually by an expert (human) planner for the query case. The ﬁnal step consisted of optimizing the query case using the constraints and priorities from the reference case. A database was assembled into an information theoretic system. An algorithm based on mutual information was implemented to identify similar patient cases. Once the best match is identiﬁed, that Clinical ly approved plan is used to generate the new plan.
2010 - Vaske - Inference of patient-specific pathway activities from multi-dimensional cancer genomics data using PARADIGM	Understand patient-specific genetic interactions and identify relevant pathways involved in cancer progression (GBM and breast cancer)	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation	First converted the biologically known pathways into a directed graph and used it to construct factor graphs composed of observed entities and hidden entities. For each pathway, trained the factor graph for each patient using EM.
2010 - Johannes - Integration of pathway knowledge into a reweighted recursive feature elimination approach for risk stratification of cancer patie	Predict relapse for patients with breast cancer (risk stratification) and identify prognostic gene signatures	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature selection	Assumes that a gene with a low fold-change should have more influence if connected to differentially expressed genes. This method uses GeneRank, adapted from Google’s PageRank, to rank features based on prior knowledge (PPI) and fold-change information from data. This ranking is then combined with the standard SVM-RFE to perform feature selection.
2009 - Zhu - Network-based support vector machine for classification of microarray samples	Predict metastasis (classification) for patients with breast cancer.	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective	Assumes that neighboring genes in a network tend to function together in biological processes. Thus, given a network of genes, treated two neighboring genes in a network as one group. Applied F_infinity norm to each group of two features encouraging 1) grouped feature selection, 2) genes in the same group to have similar coefficients, 3) genes with more direct neighbors to have larger coefficients .
2009 - Guan - Lung cancer gene expression database analysis incorporating prior knowledge with support vector machine-based classification method	Classify disease subtype (cancer versus malignant pleural mesothelioma) for patients with lung cancer	Data: Molecular Representation: Auxiliary Datasets - public Integration: Feature selection	Used prior knowledge about lung disease related genes to select feature candidates, then tested the candidates through multiple testing and combined with feature genes selected by another method. The final set of features is used to run SVM.
2009 - Binder - Incorporating pathway information into boosting estimation of high-dimensional risk prediction models	Predict survival of lymphoma and ovarian cancer patients	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective	Assumes that connected genes in a network should have similar penalties. This PathBoost method uses pathway knowledge in the likelihood-based boosting procedure to perform gene selection. In detail, adapts the penalty for features that correspond to genes that have a connection to another gene by increasing the penalty for a selected feature and decreasing the penalty for connected feature.
2007 - Chuang - Network-based classification of breast cancer metastasis	Discover prognostic biomarkers for breast cancer; predict metastatic vs non-metastatic tumor samples	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature grouping	The gene expression profiles of tissue samples were mapped into the large network. Identified subnetworks as single connected components in the network and computed an activity score for each subnetwork by averaging the gene expression levels. Additionally, measured the discriminative score for each subnetwork as the mutual information of the activity score and the class label. Used greedy search to identify subnetworks whose discrimination power is locally maximal.
2005 - Guo - Towards precise classification of cancers based on robust gene functional expression profiles	Classify tumor type of multiple types of cancer	Data: Molecular Representation: Auxiliary Datasets - public Integration: Feature grouping	First, mapped genes to their categories in Gene Ontology, which are called functional modules. For each module, constructed a representative functional feature by summary measures (mean and median). These features are then used to train a classifier (random forest).
2002 - Lee - Gene selection: a Bayesian variable selection approach	Discover significant genes on breast tumors from patients carrying or not carrying mutations in the predisposing genes patients not expected	Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature selection	This paper proposed a hierarchical Bayesian model for gene selection. It tackles the limitation that genetic is high-dimensional and previous method tend to select a large number of genes. Rather than constraining the dimension or number of important genes, the authors assign a prior distribution over gene coefficients, which allows more flexibilty. The inclusino of each feature is represented via a latent indicator. Prior knowledge about genes that are important or pairs of genes that are related can be used to specify this latent indicator.

2024 - Zhao - Application of Physics Informed Neural Network for Breast Cancer Detection

Simulation of temperature distribution in 2D breast tissues to identify breast turmos.

Data: Radiologic Imaging (thermography) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

This paper used PINN to model the temperature distribution patterns of breast cancer thermography. The loss function was designed based off the heat conduction according to the Fourier heat equation. Initial conditions were assumed to be temperature distributions before the tumor manifestation. The PINN results were compared with FEM simulations which showed good accuracy.

2024 - Zarei - A Physics-informed Deep Neural Network for Harmonization of CT Images

Reduce variability in chest CT images by harmonizing the images.

Data: Radiologic Imaging (CT) Representation: Scientific Knowledge - Mathematical Models Integration: Feature transformation; Knowledge-regularized objective

Computed Tomography (CT) quantification is affected by the variability in image acquisition and rendition. This paper aimed to reduce this variability by harmonizing the images utilizing physics-based deep neural networks (DNNs). While there exist DL algorithms to harmonize images aposst-acquisiiont, these models require to be trained on ample data covering diverse imaging conditions, which is often not available. The authors simulated vritual CT images acquired under various imaging conditions using a vritual imaging platform with 40 computional patient models featuring lungs with different levels of pulmonary diseases. Then, an adversarial generative network was trained with a single CT slice and the modulation transfer functions (MTF, which quantifies the spatial resolution characterisitics o fthe CT system) of the scanner as inputs, and the harmonized CT image as output. Results showed that the harmonized images significantly enhanced image quality and quantification accuracy in CT imaging.

2024 - Yuan - Cell-ontology guided transcriptome foundation model

Construct a Transcriptome Foundation Model with integration of existing ontology knowledge.

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective

Existing Transcriptome Foundation Models (TFMs) use pre-training methods analogous to NLP like masked token prediction, treating genes as "tokens" and cells as "sentences". However, existing TFMs treat cells as independent samples and ignore their cell-type lineages. This paper proposed a way to leverage the existing ontology knowledge. The cell ontology knowleddge is downloaded from Open BIological and Biomedical Ontology databases as an unweighted directed graph, where edges denote hierarchical lineage relationship of the form "is a subtype of" between cell types. Personalized PageRank is used to estimate pairwise node structural similarities from the oncology graph. This foundation model is pre-trained with three levels of tasks: (i) gene-level masked gene prediction, (ii) intra-cellular cell type coherence, which is contrastive learning based on cell type classes, and (iii) inter-cellular level ontology alignment, which uses the pairwise node structural similarities computed from the ontology graph to contrast pairs of cells. The model is evaluated on a wide range of downstram tasks including zero-shot cell clustering, novel cell type classification, marker gene prediction, and cancer drug response prediction.

2024 - Voigt - Biologically-Informed Shallow Classification Learning Integrating Pathway Knowledge.

Develop classification models to distinguish Castration Resistant Prostate Cncacers (CRPC) from primary cancers

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

This work proposes an approach to build biologically-informed neural networks that is slightly different from the commonly used masking node/edge approach. This work uses the Generalized Matrix Learning Vector Quantization (GMLVQ) model as a the backbone shallow classifier. LVQ models identify a prototype set that minimizes the overall classification error with respect to the prototypes. The GMLVQ variant provides additional model interpretation possibilities beyond the standard vector quantization by providing a classification correlation matrix that indicates strength of correlations between features that contribute to the class discrimination. To incorporate the knowlege, the authors adjusted the matrix structure in GMLVQ according to the pathway knowledge encoded in a series of knowlege matrices. This shallow biologically-informed network provides a (i) even easier interpretability compared to standard biologically-informed NN, and (ii) allows shortcut paths achieved by inserting additional vertices between connected layers.

2024 - Sainz - Exploring the potential of Physics-Informed Neural Networks to extract vascularization data from DCE-MRI in the presence of diffusion

Learn pharmacokinetic (PK) model parameters by fitting the voxel-wise contrast agent concentration time curves of DCE-MRI imaging data to provide diagnosis and monitoring of tumors

Data: Radiologic Imaging (DCE-MRI) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

The authors focus on an extension of a standard PK model called the diffusion extended Tofts model (D-ETM). This differential equation is based on the concept of effective diffusivity applied to biological tissues. The differential equation is incorporated as a soft constraint in the loss function to ensure that the PINN solution obtained by NN satisfies the physical laws described by D-ETM. To mitigate the issue that error in the PDE parameters were high, the authors used the residual based adaptive refinement method, in which the PDE residual is evaluated at a new set of collocation points randomluy sampled after a certain number of epochs. The PINN is tested in various experimental settings, including homogeneous vs. heterogeneous parameter distributions, noisy and incomplete data, showing that PINN overcome typical convergence issues when fitting the D-ETM to DCE-MRI data.

2024 - Ruan - Magnetic Resonance Electrical Properties Tomography Based on Modified Physics-Informed Neural Network and Multiconstraints

Develop deep learning models to retreive the distrbution of electrical properties of scanned tissues from the measured transmit radiofrequency in MRI

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

A fully connected NN is proposed to learn mapping from a measured transmit radiofrequency (RF) in MRI to the spatial distribution of electrical properties (EP) of scanned tissues. The PDE considered here is the Maxwell's equation to describe the relationship between the magnetic field and spatial distirbution of EP. In contrast to convention PINN that directly learn mapping fomr spatial coordinates to EP, the authors propose a PINN that learns the ampping from the residual of the central physical equation of convection-reaction and its spatial gradient and Laplacian to the spatial distribution of EPs. Multiple constraints were added to avoid the ill condition of the problem, including one soft similarity constriant between permittivity and conductivitiy, and another gradient constriant to mitigate reconstruction artifacts introduced by noisy measurement and numerical computation of spatial derivatives.

2024 - Rohani - Predicting transcriptional outcomes of novel multigene perturbations with GEARS

Predict transcriptional responses to single and multigene perturbations in single-cell RNA data from perturbational screens

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Model design - Graph models

The problem is formulated as: given unperturbed gene expresssion and perturbed gene expression, the DL model predicts the gene expression outcome. For each gene, the model generates initial embeddings for unperturbed and perturbed states. A gene perturbation similarity graph is constructed leveraging the pathway information contained in gene databases. A GNN is used to to process each separately into combined information. The output is combined across all genes in cross-gene MLP layers and fed into gen-specific layers, and finally prediction layers of perturbed state.

2024 - Rodriguez - Using Physics-Informed Neural Networks (PINNs) for Tumor Cell Growth Modeling

Predict the growth pattern of tumor cells

Data: Clinical Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

This paper focuses on two mathematical models used to describe population growth over time: Verhulst growth model and the Montroll growth model. Standard PINN was used to learn parameters of the two models on a small hamster dataset.

2024 - Raeisi - Mathematical modeling of interactions between colon cancer and immune system with a deep learning algorithm

Identify different immune patterns and cancer cell behaviors through estimating key parameters of the cancer models.

Data: Clinical Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

This paper used PINN to estimate parameters and numerical solutions for the fractional differential system that captures the dynamics of the interactions between colon cancer and the immune system. Then, the PINN was used to perform simulations to confirm the. effectiveness of dendritic cells in the control and treatment of colon cancer. The authors also looked into the equilibrium points of the model and their stability to drive further insights for cancer treatment.

2024 - Podina - Learning Chemotherapy Drug Action via Universal Physics-Informed Neural Networks

Predict drug toxicity or efficacy and individual response to the drug

Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

This paper focuses on quantitative systems pharmacology (QSP) models that are used to assess drug pharmacokinetics and pharmacodynamics before they go to clinical trial. First, the PINN is applied to learn the drug action of a cancer growth ODE. Then, PINN where used to obtain an approximation of the drug doses in three ODEs. Lastly, PINN is applied to learn the net proliferation rate of doxorubicin from in-vitro data after validating the method on similar synthetic data.

2024 - Perez - A Transformative Approach for Breast Cancer Detection Using Physics-Informed Neural Network and Surface Temperature Data

Detect the presence of breast cancer

Data: Radiologic Imaging (infrared camera) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

This paper reviews works that used PINN to model the forward and inverse heat transfer problems. The forward problem consists of predicting the temperature distribution, and the inverse problem predicts the thermal conductivity from known temperature data. PINN solving the inverse problem can be used for breast cancer detection from surface temperature data collected by infrared cameras. The PINN-SDSD model leverages the Pennes bioheat equation and patient-specific infrared temperatures to obtain solutions for the bioheat equation in breast cancer thermal transport.

2024 - Nevirosky - Explicit Physics-Informed Deep Learning for Computer-Aided Diagnostic Tasks in Medical Imaging

Develop cell foundation models for cancer diagnosis and prognosis

Data: Molecular Representation: Auxiliary Datasets - public Integration: Data - Transfer learning

This paper propose GeneBag, a cell foundation model pretrained on 1.3 million human single-cell RNA sequencing data points from PanglaoDB on a masked gene prediction task. The model is a modified BERT that processses input sequences from single-cell or bulk tissue RNA-seq data through an encoder with multi-head attention. This model can be fine-tuned on tissue-level bulk RNA sequencing data to pave the way for downstream applications in diagnostic and prognostics. Results showed good performance for tissue and cancer classsification as well as stage and survival prediciton using bulk RNA-seq

2024 - Lari - A Holistic Physics-Informed Neural Network Solution for Precise Destruction of Breast Tumors using High-Intensity Focused Ultrasound on a Realistic Breast Model

Predict the outcomes of high-intensity focused ultrasound (HIFU) therapy for breast tumor destruction

Data: Radiologic Imaging (ultrasound) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

HIFU is a technique that uses high-frequency sound waves to precisely target and destroy cancer cells without surgey. However, simulating the effect of HIFU on tissues is challenging due to the complex multiphysics phenomena. This paper presents a method to predict the outcomes of HIFU therapy using PINN. First, a 3D anatomically realistic breast phantom (ARBP) is generated from T1-weighted MRI to reflect the breast's anatomy. Then, equations about the pressure acoustics physics and bioheat transfer physics were considered in a PINN to predict heat conduction and generate a temperature map within the tissue domain.

2024 - Jiang - IRnet- Immunotherapy response prediction using pathway knowledge-informed graph neural network

Predict cancer patient's response to immune checkpoint inhibitors (ICI) treament

Data: Molecular Representation: Auxiliary Datasets - public Integration: Data - Transfer Learning

One main challenge is the lack of patient samples, which is tackled by pretraining on patient transcriptomics data and survival information from TCGA dataset. This was feasible because the authors identified a correlation between ICI response and durvival time through a Bayesian network model, and based on this correlation, they pretrained the model using TCGA despite it had no ICI response labels, then fine-tuned the model on ICI data. Pathway interaction importance was measured based on the attention weights for the GAT layer, and gene importance was measured on the learned weights of the feed-forward network layer, which mimics the gene-to-pathway membership. A website is available: https://irnet.missouri.edu/

2024 - Hu - Radiotherapy toxicity prediction using knowledge-constrained generalized linear model

Predict radiotion toxicity or Normal Tissue Complication Probability (NTCP) to develop certain complication given the planning dose distirbution during radiotion therapy

Data: treatment Representation: Knowledge from Experts - Quantitative Integration: Knowledge-regularized objective

Existing univariate analysis have the limitations of not considering the strong correlation of dosimetric variables because of the physics of dose deposition in clinically realistic dose delivery methods. Among the multivariate methods, geleralized linear model (GLM) are easier to interpret. This work porposed a knowledge-constrained GLM to translate three pieces of domain knowelge into non-negativity of the coeffiicents (increasing the dosimetric variable is associated with an increase or no impact for the risk of toxicity but must not decrease the risk), monotonicity (the greater the dose, the higher risk of toxicity), and adjacent simmilarity constraints (dosimetric variables with adjacent dose levels are highly correlated). The model can be solved with standard optimizers. The model was demonstrated on lung cancer and prostate cancer datasets.

2024 - Hossain - Not all tickets are equal and we know it- Guiding pruning with domain-specific knowledge

Knowledge discovery of gene regulatory relationships in breast cancer

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Training - Knowledge-in-the-loop

This paper proposed an iterative pruning-based approach that accounts for prior knowledge by scoring parameters in terms of their alignment with the prior knowledge. The knowledge is represented as a matrix between inputs and outputs. For example, the matrix may represent prior experimental evidence regarding the binding of transcription factors (inputs) and target genes (outputs). During the pruning, a tuning parameter combines the data-informed weight matrices with prior-knowledge-informed effect matrices. This approach is shown to encourage sparse gene regulatory dynamics that allow for easier interpretation.

2024 - Hossain - Biologically informed NeuralODEs for genome-wide regulatory dynamics

Knowledge discovery of gene regulatory network (GRN) that can explain temporal gene expression patterns.

Data: Molecular Representation: Knowledge from Experts - Quantitative Integration: Knowledge-regularized objective

This paper proposed PHOENIX, a NeuralODE based on Hill-Langmuir kinetics (models dynamic transcription factor binding site occupancy) to promote sparse, biologically interpretable representations of GRN ordinary differential equations. The activation f unctions are designed to resemble Hill kinetics. A trained NN block encodes the ODEs governing the dynamics of gene expression and hence encodes the underlying GRN. To tackle the challenge that NeuralODEs have multiple solutions, of which many are inconsistency with our understanding of the biological phenomona, the authors introduced biologically motivated soft constraints to regularize the solution search to be consistent with a prior. In this case, the prior is formulated as a simple linear model of chemical reaction networks that encode likely network structure based on prior domain knowledge.

2024 - CancerLLM- A Large Language Model in Cancer Domain

Develop a medical Large Language Model (LLM) specifically designed for cancer domain.

Data: Clinical (textual) Representation: Auxiliary Datasets - public Integration: Data - Transfer Learning

The existing CancerBERT primarily focuses on breast cancer leaving other cancer types relatively unaddressed. This work introduces CancerLLM, a state-of-the-art LLM with 7B parameters specialized for the cancer domain with capabilities in cancer phenotype extraction and cancer diagnosis generation. The Mistral 7B was pre-trained using 2M cancer clinical notes and 500k pathology reports, then fine-tuned on three specialized datasets focusing on phenotype extraction and diagnosis generation. The CancerLLM was further fine-tuned using instruction learning for these tasks.

2024 - Alharbi - LASSO–MOGAT- a multi-omics graph attention framework for cancer classification

Classify 31 cancer types from multiomics data

Data: Molecular Representation: Knowledge graph Integration: Model design - Graph models

This paper proposed a graph-based deep learning framework that integrates messenger RNA, microRNA, and DNA methylation data to classify 31 cancer types and normal samples. mRNA or RNA-Seq data were processed using differential gene expression analysis. DNA methylation data were processed by fitting a Linear Models for Microarray (LIMMA) models. LASSO regression was used to select the most relevant features from multi-omics data (from the order of ten thousands to hundreds). The graph attention networks (GAT) was leveraged to incroporate protein-protein interaction networks. Each node represent a protein, node features include multi-omics data; edge represent interactions between proteins. So all patient's network have the same topology but with patient-specific node features. The proposed model achieved SOTA performance among multi-omics graph methods based on PPI networks.

2023 - Zhang - Accurate TCR-pMHC interaction prediction using a BERT-based transfer learning method

Prediction of TCR-pMHC binding for development of cancer immunotherapy

Data: Clinical Representation: Auxiliary Datasets - public Integration: Transfer learning

The authors proposes a TCR-BERT embedding model, a pre-trained architecture based on BERT. Training was done on over 113,000,000 CDR3beta sequeances from TCRdb database, coverting over 29 diseases and 10 tissues. The model was trained using masked language modeling tasks. The authors also propose a pMHC-BERT embedding model that share similar architecture as the TCR embedding model. Finally, the TABR-BERT model is trained to predct TCR-pMHC interaction. This approach mitigates the challenge of limited training data in TCR-pMHC binding prediction models.

2023 - Yin - SWENet- a physics-informed deep neural network (PINN) for shear wave elastography

Image reconstruction in Shear Wave Elastography (SWE)

Data: Radiologic Imaging (shear wave elastography) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

SWE is a new imaging technology that allows measuring mechnical parameters of soft tisssues in vivo and non-invasively. However, conventional SWE methods have the limitation of large errors in heterogeneous materials and when the cross-section dimension of a phantom tumor is smaller than ~1cm. DL-based SWE has the potential to provide a more accurate evaluation of the spatial distirbution of tissue mechanical properties. This paper proposed a framework to reconstruct image from SWE based on two PINN networks. The first NN takes temporal inputs and learns parameters that are in function of time. The second NN takes just spatial coordinates and learns the time-independent parameters. Both PINNS incorporates governing PDE of wave motion in elastic soft materials and the constraints. The model was tested on simulation data and ex vivo experiments on real tumors.

2023 - Xu - Knowledge-Infused Prompting Improves Clinical Text Generation with Large Language Models

Develop a Large Language Model (LLM) for clinical NLP tasks

Data: Clinical (textual) Representation: Auxiliary Datasets - public Integration: Data - Transfer Learning

This paper proposes ClinGen, a LLM for clinical NLP tasks. Main innovations involve clincal knowledge extraction and context-informed LLM prompting. For clinical knowledge extraction, the authors (i) sampled topics (nodes) from healthcare knowledge graphs such as iBKH KG, (ii) queried topics from ChatGPT3.5, and (iii) queried writing styles from LLM. This information is used to augment the prompts by sampling a topic and a writing style from the candidate set to add to the LLM instruction. The model is then fine-tuned on the downstream tasks. The model is evaluated on synthetic clinical data generation across 7 clinical NLP tasks and 16 datasets.

2023 - Wysocki - Transformers and the Representation of Biomedical Background Knowledge

Natural Language Inference in the context of a cancer precision medicine inference task to support Clinical s in the evaluation of the Clinical significance of findings compared to existing evidence

Data: Clinical Representation: Auxiliary Datasets - public Integration: Transfer learning

This paper compared two transfromers specialized for the biomedical domain (BioBERT and BioMegatron), both pre-trained on large biomedical text corpora (PubMed). The comparison evaluates their ability to capture entities, complex relationships, biomeidcal facts, etc.

2023 - Weiskittel - Network Biology-Inspired Machine Learning Features Predict Cancer Gene Targets and Reveal Target Coordinating Mechanisms

Predict dependency of genes to cancer lineage

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation

For each cancer type, the ARACNE algorithm was used to create regulatory networks for each cancer lineage in DepMap using Cancer Cell Line Encyclopedia RNAseq data. For each gene in each lineage-specific regulatory network, the following were extracted: (i) network features, (ii) number of cancer halllmark neighbors, (iii), sum of weights, and (iv) length of path to cancer-associated genes from the Cancer Gene Census.

2023 - Wang - Quantifying intra-tumoral genetic heterogeneity of glioblastoma toward precision medicine using MRI and a data-inclusive machine learning algorithm

Predict the regional genetic alteration of three GBM driver genes (EGFR, PDGFRA and PTEN)

Data: Radiologic Imaging (MRI) Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning

A novel Weakly Supervised Ordinal Support Vector Machine (WSO-SVM) was proposed to integrate unlabeled samples with imprecise/interval labels. The ordinal relationships between labels was considered in the formulation of WSO-SVM.

2023 - Wang - A Novel Hybrid Ordinal Learning Model with Health Care Application

Predict the regional genetic alteration of three GBM driver genes (EGFR, PDGFRA and PTEN)

Data: Radiologic Imaging (MRI) Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning

A Hybrid Ordinal Learner (HOL) was proposed to integrate samples with both precise and interval labels to train a robust Ordinal Learner (OL) model. The domain knowledge about imprecise labels of unlabeled samples, and ordinal relationships between labels is considered in the model. The model can be formulated as an optimization problem with ordinal constrains. A novel conversion method was developed that converts the HOL formulation into an equivalent formulation of learning a set of binary classifiers with coupled parameters.

2023 - Shoemaker - Bayesian feature selection for radiomics using reliability metrics

Selection of radiomic features

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Feature selection

The authors assign a prior distribution over gene coefficients, which allows more flexibilty. The inclusion of each feature is represented via a latent indicator. Prior knowledge about genes that are important or pairs of genes that are related can be used to specify this latent indicator.

2023 - Shi - MORGAT- A Model Based Knowledge-Informed Multi-omics Integration and Robust Graph Attention Network for Molecular Subtyping of Cancer

Classification of cancer types

Data: Molecular Representation: Knowledge graph Integration: Model design - Graph models

This paper prosed a framework for molecular subtyping of cancer based on the knowledge-informed multi-omics integration and robust graph attention network. The framework consists of four parts: (i) feature selection module for genomics, epigenomics, and trandscriptomics through differential analysis, (ii) removal of fake edges (far nodes) in the graph constructed with omics data based on the Gaussian distance between omics data, (iii) knowledge-informed multi-omics integration module, where omics data are sorted as a time series according to the "natural" sequence of genomics, epigenomics, transcriptomics, proteomics, and metabolomics, then integrated multi-omics data in two directions from the genome to the phenotype and from the phenotype to the genome using a bi-directional LSTM; (iv) compute the feature contribution of features on each sample.

2023 - Savchenko - Mathematical modeling of BCG-based bladder cancer treatment using socio-demographics

Predict tumor cell growth with personalized BCG models for improved treatment of breast cancer

Data: Clinical Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective

This study aims to improve the Bacillus Calmette-Gérin (BCG)-based treatment protocols by changing the “one-size-fits-all” approach with a more personalized one. This was done by integration of a machine learning component which is used to assess and adjust the model’s parameters over individuals and over time. Specifically, in the BCG model, the scalar parameters were replaced with functions that depend on time and the socio-demographics of the patient, and a term associated with the presence of immune cells based on the level of the patient’s activeness was added to the immune cell populations. The authors carefully designed a parameter fitting procedure to estimate the personalized model parameters overcoming the limited labeled samples (and explained right after):

2023 - Pratama - Physical restriction neural networks with restarting strategy for solving mathematical model of thermal heat equation for early diagnose breast cancer

Early breast cancer detection through surface temperature data

Data: Radiologic Imaging (infrared) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

This paper focuses on solving the two-dimensional heat equation's PDE for breast cancer detection (Penne's bio-heat transfer equation). A PINN was used to solve the PDE thermal analysis for normal and tumorous breasts to study the temperature behavior surrounding the cancer. A restarting strategy was designed where if the loss function does not decrease in several subsequent iterations, the process will stop and restart to the next cycle.

2023 - Perez - Thermal modeling of patient-specific breast cancer with physics-based artificial intelligence

Early breast cancer detection through surface temperature data

Data: Radiologic Imaging (infrared) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

This paper describes the application of PINNs to reconstruct digital models of breast cancer from infrared surface temperatures. The PINN uses random points to find the solution and solves the bioheat equation.

2023 - Peng - Boundary delineation in transrectal ultrasound images for region of interest of prostate

Segment prostate from transrectal ultrasound (TRUS) images for brachytherapy for prostate cancer

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Hyperparameter setting

Among the four common imaging modalities used for prostate diagnosis, TRUS has a higher dependence on prior information regarding shape features for prostate segmentation. This study proposes a semi-automatic contour detection that adopts radiologist-defined data seed points as a prior. The seed points are important because the segmentation algorithm (principal curve-based method) assumes that seed point is perfect without any abnormal points.

2023 - Park - Structured deep embedding model to generate composite clinical indices from electronic health records for early detection of pancreatic cancer

Early detection of pancreatic cancer from electronic health records

Data: Clinical (EHR) Representation: Scientific Knowledge - Probabilistic Relationships Integration: Model design - Customized model views; Feature engineering - Feature grouping

This paper proposes a deep embedding model that aims to reduce the dimensionality of the input variables in EHR by grouping related measurements as determined by domain experts. Results show that the grouped indices representing liver functions can be good indicators for early detection of pancreatic cancer. The deep embedding models are designed with three levels of hierarchy: (i) base model, which learns a reduced representation for each time-series varaible and concatenates them to project into a binary prediction space; (ii) combo model, which adds a grouping layer to group redundant variabless and geenrates 32 combo embeddings; (iii) and composite model that has another gouping layer to group relevant variables among 32 comboes. The second and third grouping strategies leverage a structure of variables determined by domain experts.

2023 - Mukhemetov - Physics-informed neural network for fast prediction of temperature distributions in cancerous breasts as a potential efficient portable AI-based diagnostic tool.pdf

Early breast cancer detection through surface temperature data

Data: Radiologic Imaging (infrared) Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

This paper proposed a PINN that combines deep learning and physical principles to predict the temperature distributions in breast tissues and identify potential abnormal regions indicating the presence of tumors. The PINN model is normally trained by physics in terms of the residuals of the heat transfer equation, as well as boundary conditions with and without datasets of surface thermal imaging data concerning cancerous breast tissues, which can be used for future inverse thermal modeling to calculate tumor sizes and locations. The The PINN simulations are found to be 12 times faster than its FEM counterpart.

2023 - Meany - Predicting intratumoral fluid pressure and liposome accumulation using physics informed deep learning

Accurately measure spatial liposome accumulation and interstitial fluid pressure (IFP) distribution within a solid tumor for optimal treatment planning

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

Existing PINN works solve the forward problem. However, IFP is difficult to measure in vivo, so it is often more useful to consider the inverse problem: given the intratumoral accumulation of liposome post-administration, we want to derive the underyling IFP which resulted in that spatial distirbution. Such capability is important for forecasting tumor progression and desgining tumor treatment. This paper proposed a PINN for solving the inverse problem: predicting spatiotemporal accumulation of intratumoral liposome and spatial IFP in vivo from a single image of liposome accumulation post-administration and estimates fo the parameters present in the mathematical model. The PINN is based on a rely on a PDE to relate liposome transport and IFP within tumour tissue. The PINN learns the temporal process through the PDE and data provided at two time points, pre- and post-administration.

2023 - Meaney - Deep learning characterization of brain tumours with diffusion weighted imaging

Predict tumor progression curve and estaimte patient-specific tumor parameters of the PI model for patients with blioblastoma multiforme (GBM) for personalzed treatment

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

The PI model need two parameters to run. However, accurately estimating these two parameters is quite challenging. Existing appraoch use sumary measurements taken from imaging with several mathematical assumptions, which ignores granular measurements. The authors propose a deep learnning model to provide patient-specific estimates of the PI model parameters and simultaneously give a full forecast of the intermediate GBM progression curve. First, a cellularity map is obtained from the image using rules derived from known mapping between ADC nad tumor cellularity and the tissue class. Then, cellularity map is used as inputs of the deep learning model. PI parameters are incorporated into the customized loss, which is designed following concepts of PINN.

2023 - Mao - Weakly Supervised Transfer Learning with Application in Precision Medicine

Predict Tumor Cell Density for patients with GBM

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning

To build personalized models despite limited labeled samples are avaiable for each patient, this paper employed two strategies: (i) transfer learning from other patients with the same disease, and (ii) weakly supervised learning using weak labels derived from radiologist's expertise. Specifically, the enhancing ROI and non-enhancing ROI annotations were used to extract pairs of unlabeled samples that are expected to have ordering relationship in their labels. An additional term is added to the loss function to penalize the wrongly ordered weakly labeled pairs.

2023 - Kazemimogadam - A deep learning approach for automatic delineation of clinical target volume in stereotactic partial breast irradiation (S-PBI)

Accurately segment breast clinical target volume (CTV) in post-operative breast cancer radiotherapy

Data: Radiologic Imaging (CT) Representation: Knowledge from Experts - Qualitative Integration: Model design - Customized model views

CTV delineations are challenging because the extend of microscopic disease is not visualizable in radiological images. CTV is primarily identified by the clinicians to assure that tumor spread patterns are incorporated in the delineation process. Physician's contouring practice for CTV segmentations in stereotactic partial breast irradiation (S-PBI) involves deriving CTV from tumor bed volume (TBV) via margin expansion followed by correcting the extensions for anatomical barries for tumor invasion. The authors tried to mimic this practice in a 3D U-Net segmentation model. Specifically, the model takes in the CT image and the corresponding TBV mask as multi-channel inputs. U-Net is learnt to look for the location cues provided by the TBV to extract discriminative features. The expasion rules and anatomical barriers of the chest wall and skin boundary are also learnt to constrain TBV to CTV expansion.

2023 - Imperio - Pathologist Validation of a Machine Learning–Derived Feature for Colon Cancer Risk Stratification

evaluate whether pathologist scoring of a histopathologic feature previously identified by machine learning is associated with survival among patients with colon cancer.

Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Other

Tumor adipose feautre (TAF) was previously identified via a prognostic deep learning model developed with an independent colorectal cancer cohort. In study had 2 pathologists conduct semiquantitative scoring for TAF on colon cancer cases to verify the prognostic value of TAF for overall survival.

2023 - Cho - Interpretable Meta-learning of Multi-omics Data for Survival Analysis and Pathway Enrichmen

Survival analysis/prediction for patients with cancer

Data: Molecular Representation: Auxiliary Datasets - public Integration: Multitask learning

Proposed a meta-learning approach that uses various combinations of integrated omics datasets to train a prediction model of cancer survival. Results showed that this meta-learning design allowed finding robust functional and semantic relationships among genes.

2023 - Chen - TGM-Nets- A deep learning framework for enhanced forecasting of tumor growth by integrating imaging and modeling

Predict tumor growth

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

This paper proposed a PINN for prediction of tumor growth. The PINN is based on an equation of cellular microstructure and a diffusion-reaction equation for nutrient concentration. Fourier layers were added to extract features from the input images. Sequence learning was employed to re-train the same NN for solving the PDE over successive time segments while satisfying the already obtained solution for all previous time segments. Fine-tuning with physics was done for a reliable extrapolation. The model is trained using both synthetic and experimental data. This model is claimed to be able to make forecasts outside the training domain, for both early stage of the tumor and later development of the tumor.

2023 - Cao - X-LDA- An interpretable and knowledge-informed heterogeneous graph learning framework for LncRNA-disease association prediction

identify IncRNAs associated with prostate, colorectal, and breast cancer

Data: Molecular Representation: Knowledge graph Integration: Model design - Graph models

This paper proposed an interpretable and knowledge-informed graph learning model to predict lncRNA-disease associations (LDAs). First, lncRNA sequence similarity and disease semantic similarity are calculated based on corrrelations and domain knowledge. These similarity matrices are used to construct a graph representing the intricate topology among lncRNA, miRNA, and disease nodes. Then, graph patches, inspired by image patches in computer vision, are designed for each lncRNA-disease node pair to capture topological relationships and their features. The authors defined nine types of graph patches encompassing threee meta-paths and six diverse link modules. A CNN model takes graph patch features as input and convolution kernels are used to process the data for LDA prediction. A post-hoc attribution methods is used for interpretability.

2023 - Bogatu - Meta-analysis informed machine learning- Supporting cytokine storm detection during CAR-T cell Therapy

Predict adverse effects of chimeric antigen receptor therapies in cancer treatment

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature engineering - Feature transformation

Cytokine release syndrome (CRS) is one of the most consequential adverse effects of chimeric antigen receptor therapies in cancer treatment. This paper proposed a meta-review informed method for identification of CRS. Specifically, a set of biomarkers and their statistical information were extracted from the literature. This information is refered as a Knowledge Base (KB) of statistical evidence for CRS diagnosis. This knowledge is used to getting a matrix view of biomerker value similarities with respect to KB, so that ML algorithms with model the biomarker correlations not only within the psace defined by available data but relative to past discoveries as well. In addition, the KB was used to generate abductive evidence-based reasoning behind the predictions referring to the existing studies.

2022 - Xie - SRG-Vote- Predicting Mirna-Gene Relationships via Embedding and LSTM Ensemble

Predict potetial miRNA-gene interactions from large public datasets which might contribute to future diagnosis, prevention, and cancer therapy

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Graph models

The authors propose a voting system composing of three types of embeddings: (i) sequential embedding of miRNA and gene sequence via doc2vec, (ii) geometrical embedding of miRNA and gene via role2vec, and (iii) a graph convolutional networks embedding from miRNA-miRNA and gene-gene networks. Pairs of miRNA-gene were fed into LSTM and B-LSTM models to predict their relationships.

2022 - Wang - Biologically-informed deep neural networks provide quantitative assessment of intratumoral heterogeneity in post-treatment glioblastoma

Prediction of three gene modules: recurrent tumor tissue, treatment-induced changes, and normal brain tissue

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning

Proposed a two-stage model pipeline, including BioNet_Neu and BioNet_ProInf, to accommodate the hierarchical relationship of three gene modules relevant to GBM prognosis. Predictions are first generated for the gene module that has more supervision from biological knowledge. The outputs of this step then provided additional information for the interplay of the other two gene modules. The BioNet_ProInf is a multitask, semi-supervised learning model with a customknowledge attention loss.

2022 - Wang - A Novel Deep Learning Method to Predict Lung Cancer Long-Term Survival With Biological Knowledge Incorporated Gene Expression Images

Predict long-term survival for patients with lung cancer

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation

To overcome high-dimensionality, converted gene data into images incorporating pathway knowledge. Specifically, the knowledge provides hierarchical data where genes and proteins are root and gene expression are leaves. Used the hierarchical index sorting by mean across all samples to put gene expressions into a square matrix. Then, trained a CNN using four inputs: two gene expression images and two sets of Clinical data.

2022 - Shen - Identify Representative Samples by Conditional Random Field of Cancer Histology Images

1. Identify representative tissues within whole-slide images. 2. Classification of cancer prototype for colorectal cancer tissues. 3. Active learning -find representative patches to label to reduce annotation costs.

Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Knowledge-regularized objective

Recorded the spatial coordinates of each patch on the image grid. Then, used conditional random field (CRF) to model the probability of classes based on the classification output and spatial coordinates. The CRF model measures the spatial correlations to derive costs of different offsets and dynamically learns offsets that lead to locations of most representative patches. CNN was used as a feature extractor. The model can be used as feature extractor for any state-of-the-art classifiers. The model can also be used for active learning to reduce annotation costs.

2022 - Perez Raya - Thermal Modeling of Patient-Specific Breast Cancer with Physics-Based Artificial Intelligence

Learn parameters of a physics model (model the breast surface temperature from thermal properties of the tumor in order to detect tumor location)

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

Used physics-informed neural network to learn the parameters of a bioheat equation. Physics-informed neural network uses physical constraints, boundary conditions, and a PDE equations to define the loss function.

2022 - Peng - H-ProSeg- Hybrid ultrasound prostate segmentation based on explainability-guided mathematical model

Segment prostate from transrectal ultrasound (TRUS) images for brachytherapy for prostate cancer

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Hyperparameter setting

This study proposes a semi-automatic contour detection that adopts radiologist-defined data seed points as a prior. The model has three sub-networks. The first sub-network uses an improved principal curve-based methods to obtain seed points and their projection index. The second sub-network uses a differential eovlution-based neural network. The third sub-network uses the neurla network to explain the smooth mathematical description of the prostte contour.

2022 - Ma - DualGCN- a dual graph convolutional network model to predict cancer drug response

Predict anti-cancer drug response (sensitive vs resistant).

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation

Used chemical structure of drugs to represent each drug as a graph, where nodes are atoms and edges are connection between atoms. Used PPI network to represent each cancer sample as a graph, where nodes are proteins (genes) and edges are interaction between proteins. The two graphs as separately fed into GCN and their embeddings are concatenated and fed into MLP to predict drug response.

2022 - Liu - Multimodal Brain Tumor Segmentation Using Contrastive Learning Based Feature Comparison with Monomodal Normal Brain Images

Segment brain tumor from MRI.

Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Feature transformation

This paper used the MRI from healthy controls to help segment tumor for tumor brains. One challenge is that normal brain images are often monomodal and tumor brain images are multimodal. To tackle this, the authors used a "normal appearance network" to obtain a reconstructed "normal brain" image for the tumor brains that is comparable to the healthy controls. In detail, they trained IntroVAE using T1 image of healthy controls to obtain a lower dimensional representation of normal brain appearance, then projected the tumor brain image to this representation and obtain a reconstruction of "normal brain appearance". Then, in the segmentation network, they i) used a simple Siamese network to align the feature maps of the normal and tumor images in the normal regions, and ii) compared the feature maps of the normal and tumor images to identify low feature consistency regions as tumor regions.

2022 - Liu - MetBERT- a generalizable and pre-trained deep learning model for the prediction of metastatic cancer from Clinical notes

Predict cancer metastasis (all types of cancer)

Data: Clinical Representation: Auxiliary Datasets - public Integration: Transfer learning

Used four language models pre-trained on Clinical and scientic data (BioBERT, BlueBERT, Clinical BERT, PubmedBERT) and fine-tuned them using in-domain data (EHR Clinical notes).

2022 - Li - Coupling Deep Deformable Registration with Contextual Refinement for Semi-Supervised Medical Image Segmentation

Segmentation of anatomical structures from medical images (not targeted for tumor but can be applied to tumor and is related to cancer diagnosis)

Data: Radiologic Imaging Representation: Scientific Knowledge - Probabilistic Relationships Integration: Knowledge-in-the-loop

First, the authors applied an image registration model (deep diffeomorphic registration model) that allows regions predefined on templates to be overlaid on the image and produces a segmentation map of anatomical structures. Then, this segmentation map is refined by warping with the probabilistic atlas map and taking the nearest interpolation.

2022 - Lee - A Novel Knowledge Keeper Network for 7T-Free but 7T-Guided Brain Tissue Segmentation

Segment brain tissue from MRI (not explicitly for cancer but related)

Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Transfer learning

It is known that 7T images provide better insights for diagnosis. However, 7T MRI are not always available. This paper aims to build 7T-free but 7T-guided brain tiissue segmentation model. They used paired 3T-7T images to train a module that extracts 7T-like representations from 3T images using knowledge distillation and a discriminator. At inference time, the network is able to use only 3T images to do segmentation with 7T-image guided information.

2022 - Huang - MTL-ABS3Net- Atlas-Based Semi-Supervised Organ Segmentation Network With Multi-Task Learning for Medical Images

Segmentation of organs from medical images (not targeted for tumor but can be applied to tumor.)

Data: Radiologic Imaging Representation: Scientific Knowledge - Probabilistic Relationships Integration: Knowledge-in-the-loop

First, trained a segmentation network using labeled images. Applied the trained network on unlabeled data and computed a confidence map by comparing the segmentation results with the probability atlas. The high confidence pixels were selected and set as ground truth to compute loss for unlabeled samples.

2022 - Bhattacharya - Integrating zonal priors and pathomic MRI biomarkers for improved aggressive prostate cancer detection on MRI

Pixel-level classification/diagnosis of prostate cancer (normal, indolent, aggressive classes) from MRI

Data: Radiologic Imaging Representation: Scientific Knowledge - Probabilistic Relationships Integration: Hyperparameter setting

Zonal priors contain Clinical domain knowledge that aggressive cancer is more likely to occur in some zones. The zonal priors were formulated as mean probabilities of each class, conditional probabilities of each class given the zone and conditional probabilities of each zone given the class. The authors proposed two ways to integrate the zonal prior probabililites. One way is to to use it as a prior in the bayes decision model.

2021 - Zhou - Multi-task learning for segmentation and classification of tumors in 3D automated breast ultrasound images

Classify and segment tumor in 3D automated breast ultrasoudn images for breast disease diagnosis and surgical planning

Data: Radiologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Multitask learning

Since the benigh and malignant tumors usually have different shape characteristics, the authors assume that the segmentation and the classification task can help each other. They designed a multi-task model that simultaneously does segmentation and classification. Additionally, since ABUS volumes have severe noise and tumors in ABUS have ambiguous boundaries, they proposed a iterative feature-refining strategy during training to refine the feature maps using probability maps obtained from previous iterations.

2021 - Zhou - 3D multi-view tumor detection in automated whole breast ultrasound using deep convolutional neural network☆

Detect tumor from 3D images for patients with breast cancer

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Hyperparameter setting

(1) Defined the anchor size/aspect ratios for each of the three views based on the expected tumor size from each view. For example, tumors have closer to square shape in saggital view and elongated rectangular shape in axial and coronal views. (2) Also, tumors can be as small as occupying 2% of the full image which makes them hard to detect. So redesigned the VGG feature extraction network to better detect small tumors: i) added deconvolution layers to enlarge the last feature map, ii) concatenated feature maps that represented local and global textures.

2021 - Zhao - DeepOmix- A scalable and interpretable multi-omics deep learning framework and application in cancer survival analysis

Prognosis of lower grade glioma using multi-omics data

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

This work proposes DeepOmix, a deep learning model for cancer survivla analysis using multi-omics datasets. This network can incorporate prior biological knowledge of gene functional module networks as the function module layer (the second layer). The number of nodes and edges between the first and second layer are defined based on knowledge from tissue networks, gene co-expression networks, or prior biological signaling pathways.

2021 - Zhang - Predicting Anticancer Drug Response With Deep Learning Constrained by Signaling Pathways

Predict drug response for multiple cancer cell lines.

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

This work proposes consDeepSignaling, a deep learning model for drug response prediction. To reduce number of features and make the model more interpretable, 46 signaling pathways (from KEGG) and 929 genes were used to build the deep learning model. The layers corresponded to input layer, gene layer, pathway layer, and other hidden layers. Layer's connection also followed known biological knowledge.

2021 - Zangooei - Multiscale computational modeling of cancer growth using features derived from microCT images

Predict tumor and microvascular network growth patterns for breat cancer.

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-in-the-loop

To simulate complicated spatiotemporal and multiscale biological behavior of cancer growth, this paper used three multiscale image-based PDE models of tumor growth. The simulation performed at microscope, mesoscopic, and macroscopic scale. The model environment was formulated on a regular grid of lattice sites called tumor microenvironments. Knowledge was used to setup this environment with regulatory substances, diffusible factors, cancer cells, and pathways. A deep reinforcement learning framework was setup, with discrete policies that are functional map states. At microscope scale, an agent that seek to have a desired phenotype at each simulation time step. At mesoscopi sale, cells and vessels are represented as agents with their own phenotype. A neural netowrk predicts tumor cell growth.

2021 - Weiss - PILOT: Physics-Informed Learned Optimized Trajectories for Accelerated MRI

Find optimized trajectories for accelerated MRI

Data: Radiologic Imaging Representation: Scientific Knowledge - other Integration: Knowledge-regularized objective

This work aims to find optimized trajectories for accelerated MRI. Prior knowledge about the peak-current and the maximum slew-rate produced by the gradient coils were translated into geometric constraints in the optimization problem to ensure physically viable solutions.

2021 - Wang - Knowledge-Infused Global-Local Data Fusion for Spatial Predictive Modeling in Precision Medicine

Predict brain tumor cell density (TCD) by a Gaussian Process based model.

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective

Proliferation-invasion (PI) model is used to predict TCD. The difference between the predicted TCD from the proposed model and that from PI is constrained to be smaller than slack variables. (Soft constraints)

2021 - Sasse - Integration of multiomics data with graph convolutional networks to identify new cancer genes and their associated molecular mechanisms

Identify cancer-related genes from high-throughput molecular data

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation;

This paper proposed combining graph convolutional networks with protein-protein interaction (PPI) networks to predict cancer genes from multiomics pan-cancer data. The methods consists of (i) collecting average mutation rates, CNAa, DNA methylation, and gene expression changes from all genes across 16 tumor types annd concatenating into a single feature matrix, (ii) A network is constructed where nodes are high-confidence cancer/non-cancer genes and edges are known interactions between them from PPI. (iii) The GCN is trained to classify genes into predicted cancer and non-cancer genes. (iv) The most important features for the classification are extracted. (v) Genes are clustered according to their feature contributions to detect modules with important gene-gene connections in cancer.

2021 - Sanyal - Weakly supervised temporal model for prediction of breast cancer distant recurrence

Predict cancer recurrence for patients with breast cancer

Data: Clinical Representation: Auxiliary Datasets - private Integration: Weakly supervised learning

(Same model as Barnejee 2019). This work used a large set of Clinical notes to generate weak labels for training data, and trained a NLP model using the small, expert-labeled data and large, weakly labeled data. Used medical dictionary to clean the text data (combine related terms together, reduce sparsity).

2021 - Ramirez - Classification of Cancer Types Using Graph Convolutional Neural Networks

Classification of cancer types (33 cancers and normal)

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Graph models

Two different input graphs were generated, a gene co-expression graph (based on correlation between gene expressions) and a protein-protein interaction graph from the STRING database. These inputs are passed into a graph convolutional layer followed by hidden layers and a classifier.

2021 - Nave - Artificial immune system features added to breast cancer Clinical data for machine learning (ML) applications

Predict tumor size at different time points for breast cancer

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Feature transformation

This study aims to use ML to predict the tumor size for each time point based on the imaging data. A mathematical model was used to extract immunological features, which are then used as inputs into ML model together with Clinical features

2021 - Liu - TranSynergy- Mechanism-driven interpretable deep neural network for the synergistic prediction and pathway deconvolution of drug combinations

Predict efficiency of drug combination and serve as screening for downstream experimental evaluation

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation

In the drug response prediction, the input are two drug molecules in the drug combination and a cell line (gene expressions). The authors argue that using the chemical structure of drugs as features it is not straightforward to connect with the cellular mechanisms of drug action. Biological representation based on drug-target interaction profile is a better strategy to infer the drug representation vector. They used protein-protein-interaction network features. Random walk with restart algorithm was applied on a protein-protein interaction network to infer drug-target profile as the drug features. From gene expressions, the gene-gene dependency profiles were extracted.

2021 - Liu - Real time volumetric MRI for 3D motion tracking via geometry-informed deep learning

Reconstruction of accelerated MRI

Data: Radiologic Imaging Representation: Scientific Knowledge - other Integration: Customized model pipelines

In this neural network, the authors added an explicit geometry-module that encodes the k-space sampling patterns and known transforms between the k-space and image domains.

2021 - Li - Informed Attentive Predictors- A Generalisable Architecture for Prior Knowledge-Based Assisted Diagnosis of Cancers

Predict (classification) whether a sample is tumoral or normal for different types of cancers.

Data: Molecular Representation: Scientific Knowledge - Probabilistic Relationships Integration: Feature transformation

This model outputs prediction as a function of the data-driven prediction and knowledge-based prediction. They generated knowledge-based predictions based on the intuition that samples whose difference with normal samples is more similar to the mutation patterns known from knowledge are more likely to be tumoral. In detail: 1) Extracted a knowledge vector representing gene mutation rate for each gene from an exterior data source. 2) Generated a reduced version of original feature vectors to match the dimensions of the knowledge vector. 3) Computed a general representation for normal samples by projecting original features into 1-d vector. 4) For each sample, computed their difference with this normal sample vector. The similarity between this difference and the knowledge vector determines the likelihood of a sample between tumor, thus, generates a knowledge-based “prediction”. 5) The final prediction is a function of the prediction by original data and from this knowledge-based prediction.

2021 - Kaandorp - Improved unsupervised physics- informed deep learning for intravoxel incoherent motion modeling and evaluation in pancreatic cancer patients

Estimate three parameters of the Intravoxel incoherent motion (IVIM) model for diffusion-weighted imaging (DWI)

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Physics-informed Neural Networks

Designed an unsupervised NN to predict the three IVIM parameters of interest. The model minimizes a physics-based loss that measures the difference between the original signal and “reconstructed” signal based on the IVIM model and NN estimated parameters.

2021 - Joshi - OncoNet- Weakly Supervised Siamese Network to automate cancer treatment response assessment between longitudinal FDG PET CT examinat

Classify the treatment outcome (preogression, resolution, stable) of patients with lung cancer

Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Weakly supervised learning

Since expert annotation is costly, this work used rule based heuristics to extract Standard Uptake Values (SUVmax) of metabolically active tumors from radiology reports. Then used differences in the SUVmax values between scans to derive the weak labels (progression, resolution, stable). This heuristic is based on the Lugano 2014 criteria for tumor evaluation, assessment, and prediciton.

2021 - Johnson - Integrating transcriptomics and bulk time course data into a mathematical framework to describe and predict therapeutic resistance in cancer

Predict therapeutic resistance in breast cancer by improved model calibration

Data: Molecular Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective

Mathematical models of tumor progression often include distinct subpopulations. Despite these models are calibrated to experimental data, single biomarkers do not capture longitudinal patterns. The study proposed an integrated model calibration scheme that integrates information from two data sources: (i) from longitudinal population data sampled at high temporal resolution for a number of doses, and (ii) machine learning outputs that estimate the phenotypic composition at three time points before and after treatment. The ML model was trained to classify phenotype of cells based on its transcriptome.

2021 - He - Cross-modality brain tumor segmentation via bidirectional global-to-local unsupervised domain adaptation

Segment brain tumors from multi-modal MRI for tumor diagnosis and treatment

Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Transfer learning

This paper proposed a framework that can segmenta brain tumors in one MRI modality (target domain) when trained on another MRI modality (source domain). To account for domain shift, they used a bidirectional module that learns how to generate synthetic image from one domain to another, and a global-to-local adaptor that aligns the real and synthetic images. Also, spatial-wise and channel-wise attention maps were used to transfer the most discriminate part of features through domains, assuming that silences of the images are consistent inter-domain.

2021 - Feng - Investigating the relevance of major signaling pathways in cancer survival using a biologically meaningful deep learning model

Predict drug response for cancer patients (breast cancer, lung, glioblastoma, skin)

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

The 46 signaling pathways from four types of cancer and 1967 genes were selected to design a customized deep learning model architecture. The input layer consisted two input features (gene expression and copy number variation) for each gene. Then genes were connected to the pathways based on known connections. Model was evaluated on four datasets of four types of cancer.

2021 - Erion - Improving performance of deep learning models with axiomatic attribution priors and expected gradients

Predict (regression) drug response for patients with acute myeloid leukemia (AML).

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective

The knowledge graph is used to regularizes the loss function so that relative feature contributions are consistent with the known probability of integration of gene (called graph prior).

2021 - Elmarakeby - Biologically informed deep neural network for prostate cancer discovery

Predict the disease sate (classification) of prostate cancer and better understanding of the relationship between Molecular features and the disease.

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

A customized neural network architecture is designed based on biological knowledge, where each layer and node have a specific biological interpretation, and only connections that follow known pathways are activated and others are zero-out.

2021 - Chuang - Convolutional neural network for human cancer types prediction by integrating protein interaction networks and omics data

Predict normal vs tumor samples for 11 cancer types

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation

A large number of PPIs of human proteins were collected from five public databases. Laplacian clustering approach was used to map PPI networks into 2D space, then combined with gene expression data, to generate normal and tumor sample images for CNN model training.

2021 - Chen - Modality-Specific Information Disentanglement from Multi-parametric MRI for Breast Tumor Segmentation and Computer-aided Diagnosis

Multi-modality image segmentation from MRI

Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Model design - Customized model pipelines

This paper proposed a method for multimodal image segmentation. The primary modality (DCE-MRI) is complemented by using supplementary modalities (T2w and ADC) with addititional modality-specific information to improve the segmentation. The framework comprises of three main steps: (i) two synthesis networks map the supplementary modalities to aligne with the feature spaces of the primary modality, (ii) information extractors that take in real and synthesized images to explore different cases of signal coherence across the modalities, (iii) modality specific attention module and modality trusty gating modues designed to improve the modality fusion and explore intra- and inter- modality relationships.

2021 - Azher - Development of Biologically Interpretable Multimodal Deep Learning Model for Cancer Prognosis Prediction

Predict patient survival risk scores for four types of cancers (colon, bladder, skin, liver)

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

First, pre-trained an encoder (VAE) that has biologically informed architecture. The nodes of the first two layers represent gene entities. The connections are initialized based on known biological relationships (CpG to gene, gene to pathway). After VAE is pre-trained, the encoder weights are transferred as initialization for an ANN to learn prediction model.

2021 - Anastopoulos - Patient Informed Domain Adaptation Improves Clinical Drug Response Prediction

Predict patient drug response using cross-domain data (cell line and primary tumor)

Data: Molecular Representation: Auxiliary Datasets - private Integration: Knowledge-regularized objective

Used Mean Maximum Discrepancy (MMD) as a loss term to encourage alignment of the latent space of cell line and tumor samples. Additionally, uses weight sharing for both cell line and tumor sample data neural networks. Represented each drug molecule as a graph (nodes are atoms and edges are bonds between atoms) then extracted latent representation via GCN.

2021 - A multiscale model of pulmonary micrometastasis and immune surveillance- towards cancer patient digital twins

Develop a framework to simulate the interactions between the immune system and te progression of micrometastases in the lung.

Data: Molecular Representation: Scientific Knowledge - Mathematical Models; Knowledge from Experts - Quantitative Integration: Model design - Customized model pipelines

This paper proposed a multiscale hybrid model that combines mechanistic models and machine learning to create virtual patient models, as a concept for Cancer Patient Digital Twin (CPDT). Several mechanistic models were used to model tumor growth, motility and infiltration of immune cells, and other biological processes. Machine learning methods were used in the parameter exploration for generating virtual patients. A patient classifier was trained to distinguish difeferent immune system responses based on the cancer cell population. This method is a preliminary concetp for building cancer digital twins.

2020 - Vu - Deep convolutional neural networks for automatic segmentation of thoracic

Segment thoracic organs‐at‐risk in radiation oncology for treatment planning of lung cancer, esophagus cancer, or breast cancer.

Data: Radiologic Imaging Representation: Auxiliary Datasets - public Integration: Transfer learning

Transfer learning is applied by using locking the weights from a VGG16 image classiﬁcation model trained on ImageNet. When adapting to the target dataset, activation function is modified, a batch normalization layer and a dropout layer are added.

2020 - Seninge - Biological network-inspired interpretable variational autoencoder

Knowledge discovery and better interpretation of biological mechanisms and the underlying gene regulatory networks (GRN) of single-cell GBM

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

Used a data-driven encoder to capture the complex patterns and a sparse knowledge-based decoder to allow explainability. In detail, the decoder is a single linear layer that zeros-out the connection between the latent layer and input layer if it belongs to a known biological abstraction. Also, the decoder is constraint to positive weights to maintain explainability of direction of biological activity.

2020 - Ozdemir - A 3D Probabilistic Deep Learning System for Detection and Diagnosis of Lung Cancer Using Low-Dose CT Scans

Screening and risk diagnosis of lung cancer

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning

Used expert-assigned scores as noisy labels to train a malignancy ranking network. This network ranks segmented nodules within each CT scan by their malignancy risk. The top-k nodules are then used to train a multitask classification network. (The intuition is that each CT scan may have a large number of nodules, the ranking helps reduce noise and focus on those that are more likely to be malignant.)

2020 - Oh - PathCNN- interpretable convolutional neural networks for survival prediction and pathway analysis applied to glioblastoma

Predict long-term survival for patients with glioblastoma multiforme (GBM).

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation

For each of the three omics types, organized data into a matrix where each column represent a pathway and each row represent a gene. Then, performed PCA on each matrix. Finally, combined PCs of all omics types into a pathway image ordering the pathways based on their correlation.

2020 - Liu - A Natural Language Processing Pipeline of Chinese Free-Text Radiology Reports for Liver Cancer Diagnosis

Classification of cancer vs normal for patients with liver cancer

Data: Clinical Representation: Knowledge from Experts - Qualitative Integration: Feature selection

Used expert knowledge to extract Clinical ly meaningful features (also called named entity recognition), which are then used for classification. First, used expert knowledge to derive a specialized lexicon (a list of Clinical ly relevant words and synonyms). Then, trained a language model using both original text and lexicon coded text to detect relevant entities.

2020 - Leung - A physics-guided modular deep-learning based automated framework for tumor segmentation in PET

Segmentation of tumor from PET

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Simulation

This method tackles the challenges of limited spatial resolution of PET and limited labeled data. Physic-sbased simulation is used generate synthetic images. Specifically, first- and second-order statistics for the intensity, size, shape, intra-tumor heterogeneity, and tumor-to-background intensity ratio are extracted. The kernel distribution of tumor descriptors are fed into a PET system modeling the major image-degrading processes in PET such as detector blurring with a 5 mm full-width-at-half-maximum (FWHM). Intra-tumor heterogeneity was simulated by incorporating unimodal variability in intensity values within the tumor and, for some tumors, by modeling the intensity distribution as a mixture model. Data from multiple patients were used to simulate patient heterogeneity. Tumor seed locations were manually selected such that generated tumors would appear at visually realistic locations. The simulated data is used to pre-train a U-Net. Real medical data is used to fine-tune the U-Net.

2020 - Lee - Cancer subtype classification and modeling by pathway attention and propagation

Classification of sample type and knowledge discovery/interpretation of important pathways

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Graph models

Pathways from KEGG are obtained as prior knowledge. Each pathway is represented as a graph. The gene expression profiles is amapped to the nodes of this pathway graph, where the node feature vector represent the numbe of genes in the pathway. The input graph is first passed to a GCN to capture localized patterns. One model is built for each pathway. Then, the pathway-level attention and ensemble-level attention are used to ensemble the features which are then used to produce final cancer subtype predictions.

2020 - Kuenzi - Predicting Drug Response and Synergy Using a Deep Learning Model of Human Cancer Cells

Predict drug sensitivity of tumor cells

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

Learnt genotype embedding using a network whose layers and connections are based on known biological processes. Learnt drug structure embedding using a CNN with the same number of output nodes as the genotype embedding so they can be concatenated.

2020 - Kozłowska - Mathematical model predicts response to chemotherapy in advanced non-resectable non-small cell lung cancer patients treated with platinum-based doublet

Predict anticancer treatment response to chemotherapy to find optimal treatment schedule for patients with

Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective

This study developed a method that combines mathematical mechanistic modeling with ML to estimate response to anticancer treatment for each patient. The authors replaced calibration of the mechanistic model using mixed-effect learning with calibration mainly based on a multivariate Gaussian-mixture model.

2020 - Ke - Identifying patch-level MSI from histological images of Colorectal Cancer by a Knowledge Distillation Model

Identify microsatellite instability (MSI) status in colorectal cancer for better prognosis.

Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Knowledge-in-the-loop

Most existing models just use the whole-slide label as groudn truth for each patch in the whole-slide image. However, this causes mislabeling of many local patches. To address the mislabeling problem, a self-distillation model was first used to separate data into high fidelity and low fidelity patches. Then, one-shot active learning was used where pathologists picked up the most representative samples from high-fidelity data and tagged them as clean data for noise-robust training. The self-distillation model reduced the efforts of pathologists.

2020 - He - Infusing Disease Knowledge into BERT for Health Question Answering, Medical Inference and Disease Name Recognition

Derive text data embeddings that can be used for downstream tasks including:1. consumer health question answering, 2. medical language inference, 3. disease name recognition. (Not just for cancer but I believe the diseases include some cancer)

Data: Clinical Representation: Auxiliary Datasets - public Integration: Weakly supervised learning

Used the structure of Wikipedia to extract “weakly-supervised passages” containing disease knowledge. Specifically, the title of the article is the disease name, and the title of the section is the topic aspect of the disease. Then, used these passages to create additional samples used in training.

2020 - Garcia - Transfer learning with convolutional neural networks for cancer survival prediction using gene-expression data

Predict progression of lung cancer

Data: Molecular Representation: Auxiliary Datasets - public Integration: Transfer learning

Pre-trained CNN on data of 31 cancers except lung cancer, then fine-tuned using lung cancer data.

2020 - Fortelny - Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data

Predict cell type from gene expression

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

Prior knowledge from signaling pathways and gene-regulartory networks are used to build knowledge-primed neural networks. Each node corresponds to a protein or a gene, and each edge corresponds to a regulatory relationship documented in biological databases.

2020 - Deng - Pathway-Guided Deep Neural Network toward Interpretable and Predictive Modeling of Drug Sensitivity

Predict drug sensitivities and undersand mechanisms of drug action

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

To reduce dimensionality, the authors selected 977 landmark genes. Inputs to the network include cell line feature nodes and drug target feature nodes. The architecture is customized by adding a layer of pathway nodes and their connections to input gene nodes. The rest of network are data-driven and fully connected.

2020 - Boso - Drug delivery- Experiments, mathematical modelling and machine learning

Simulate tumor growth and drug action for a proper modeling of drug delivery and efficacy in anticancer therapy

Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective

The drug-induced cell death rate is a parameter difficult to estimate experimentally. Machine learning was found helpful because it is able to catch the more complex physics and therefore to identify the necessary parameters. Neural networks was used to learn the relation between the involved quantities, by learning the mapping of input - output data.

2020 - Aguilar - A generalizable data-driven multicellular model of pancreatic ductal adenocarcinoma

Develop a data-driven computational framework that simulates the microenvironment of pancreatic tumor to understsand cell type-specific molecular interactions and cell-cell communications.

Data: Molecular Representation: Knowledge from Experts - Quantitative Integration: Model design - Customized model pipelines

This paper proposed an approach to model a block of cancerous tissue with a mixture of cancer, stromal, and immune cells randomly located inside a 3D rectangular space. Each cell has a Boolean Network (BN) that determines its cellular phenotype, the possible secretion of cytokines, and the state of membrane receptors. Within the cell, the BN contain n nodes, each with a perturbation probability. The cell-cell communication is modeled via diffusion of cytokines. The tissue is represented as a 3D point process of cells with a fixed density. Previous knowledge can be used to set constraints to this system. Model parameters were calibrated individually. This computational framework can be used to build patient-specific models given molecular data of their samples.

2019 - Zheng - Semi-supervised segmentation of liver using adversarial learning with deep atlas prior

Segmentation of liver in CT images (not targeted for tumor but can be applied to tumor.)

Data: Radiologic Imaging Representation: Scientific Knowledge - Probabilistic Relationships Integration: Knowledge-regularized objective

The probability atlas provides the prior statistical probabilty that the liver appears at each pixel location. The authors assumed that hard samples are those with atlas values close to 0.5, then the hard samples are given maximal weight in the loss function whereas the easy samples are given small weight.

2019 - Yazdjerdi - Reinforcement learning-based control of tumor growth under anti-angiogenic therapy

Control drug dosing in cancer antiangiogenic therapy

Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-in-the-loop

This work designed a RL-based controller for cancer antiangiogenic therapy by implementing the Q- learning algorithm to control drug dosing. The mathematical model of tumor growth under angiogenic inhibitor is simplified and used to represent a simulated patient and provide feedback to the controller.

2019 - Wang - Weakly supervised deep learning for whole slide lung cancer image analysis

Classification of cancer type into tumor vs non-tumor (lung0

Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning

The model is trained using both image-level class labels and a small number of coarse annotations of abnormal regions by pathologists. The cross-entropy loss is modified such that higher penalty is given when the CNN misclassifies patches that fall into annotated regions.

2019 - Wang - A dual-mode deep transfer learning (D2TL) system for breast cancer detection using contrast enhanced digital mammograms

Classify benign vs malignant tumors for breast cancer using the contrast-enhanced digital mammography (CEDM) data.

Data: Radiologic Imaging Representation: Auxiliary Datasets - public Integration: Transfer learning

Transfer learning is used in training the dual-mode DL model, in which a pre-trained Inception V3 on ImageNet is used to initialize the parameter estimation, and then the CEDM data are used to fine-tune the parameters.

2019 - Tomita - Attention-Based Deep Neural Networks for Detection of Cancerous and Precancerous Esophagus Tissue on Histopathological Slides

Classification of cancer prototype (skin cancer)

Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Customized model views

The model dynamically identifies ROIs in the whole-slide image and makes classification based on the selected ROIs. The model is composed of two modules, the first module extracts grid-based features and the second is attention module that generates an attention map used to weight the features, both trained end-to-end using image-level labels.

2019 - Shao - Weakly Supervised Deep Ordinal Cox Model for Survival Prediction From Whole-Slide Pathological Images

Predict survival of patients with multiple types of cancer (censored vs non-censored)

Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Feature transformation

Before classification, the samples (patches) were clustered into k distinct phenotype groups using Kmeans. Then generated a bag of patches by stratified sampling from the clusters. This intents to match the bag heterogeneity with WSI-level heterogeneity. The original label is binary. The authors compared the survival time (continuous) between pairs of patients to create ordinal labels of wether a sample survived long than another sample. They added a ranking loss to the objective function so the model can also learn from this ordinal information.

2019 - Sha - Multi‑Field‑of‑View Deep Learning Model Predicts Nonsmall Cell Lung Cancer Programmed Death‑Ligand 1 Status from Whole‑Slide Hematoxylin and Eosin Images

Predict the programmed death-ligand 1 (PD-L1) status from whole-slide images for nonsmall cell lung cancer (NSCLS) tumor samples.

Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Customized model views

Designed a NN with multi-field of view classification to mimic the pathologist using various zoom levels. The NN consists of one network that processes a large image and two additional branches that process small cuts of the central region of the image. This design ensures that the central region contributes more to classification.

2019 - Sevakula - Transfer Learning for Molecular Cancer Classification Using Deep Neural Networks

Classify any two tumor types with stacked autoencoder by (high dimensional) gene expression data.

Data: Molecular Representation: Auxiliary Datasets - public Integration: Transfer learning

A stacked sparse autoencoder is pre-trained by unlabeled data for feature selection and normalization for cancer classification on gene expression data. While classifying any two tumor types, data of other tumor types were used in unsupervised manner to improve the feature representation. Tested on 36 two-class benchmark datasets from the GEMLeR repository.

2019 - Nguyen - Incorporating human and learned domain knowledge into training deepneural networks- A differentiable dose-volume histogram and adversarialinspired framework for generating Pareto optimal dose distributions inradiation therapy

Predict Pareto optimal dose distributions in radiation therapy for prostate cancer patients with seven beam IMRT.

Data: Pathologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective

This work proposed a domain-specific loss, which was a differentiable loss function based on the dose-volume histogram (DVH), and combined it with an adversarial loss for the training of deep neural networks to generat Pareto optimal dose distributions in radiation therapy (for prostate cancer patients).

2019 - Nguyen - A feasibility study for predicting optimal radiation therapy dose distributions of prostate cancer patients from patient anatomy using deep learning

Predict optimal radiation therapy dose distributions (3D) of prostate cancer patients using U-net.

Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Other

U-net is used to predict fully 3D dose distribution prediction for prostate IMRT plans according to the treatment plans from physician.

2019 - Nguyen - 3D radiotherapy dose prediction on head and neck cancer patients with a hierarchically densely connected U-net deep learning architecture

Predict 3D radiotherapy dose on head and neck cancer patients by a hierarchically densely connected U-net.

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other

A deep learning-based dose prediction model, Hierarchically Densely Connected U-net, based on U-net and DenseNet is used to predict the 3D radiotherapy dose distribution of head and neck cancer patients.

2019 - Mercan - Assessment of Machine Learning of Breast Pathology Structures for Automated Differentiation of Breast Cancer and High-Risk Proliferative Lesions

Diagnosis (multi-class classification) of breast biopsy samples from whole slice images

Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Customized model pipelines

Used a sequential classification scheme in which a decision is made for a single diagnosis at a time to mimic the pathologists. The reasoning is features that describe one type of diagnosis do not apply to other diagnosis.

2019 - Ma - Incorporating Biological Knowledge with Factor Graph Neural Network for Interpretable Deep Learning

Predict Progression-Free Interval event

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective

(Very similar model as Ma, 2018) Computed a similarity matrix of learned feature representations (pairwise mutual information) and forces it to be consistent with feature relationships based on domain knowledge.

2019 - Ma - Incorporating Biological Knowledge with Factor Graph Neural Network for Interpretable Deep Learning

Predictor tumor stage for lung cancer (2 stages) and kidney cancer samples (3 stages); Classify tumor sample tyle (primary tumor vs normal)

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

Customized the network architecture such that i) each node corresponds to some biological entity, ii) only connect nodes that can reach each other in the knowledge graph within a number of steps. The later can be seen as an attention mechanism to capture hierarchical interactions among biological entities.

2019 - Liu - Automated detection and classification of thyroid nodules in ultrasound images using Clinical -knowledge-guided convolutional neural ne

Segment and classify nodule in ultrasoud images

Data: Radiologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Hyperparameter setting

Designs an expert-knowledge-guided multi-branch networkfor nodule detection. Similar to Faster R-CNN, the proposed network predicts region proposals for object detection. The authors computed the real distribution of nodule size and shape in training data to constrain the anchor boxes (lower feature maps to have smaller sizes and larger range of aspect ratios than those at higher feature maps). Based on how radiologists classify the challenging nodules (e.g. difference with surrounding tissies may reflect cell proliferation, malignant nodules tend ot have blur and irregular margin due to rapid growth), they proposed a triple-branch design to uses sonographic features known to be correlated with malignant nodules. The first branch is basic low-level features (takes original patches as input), second branch is context features (takes larger patches as input), and third branch is margin features (takes contour-masked margin regions as input).

2019 - Hu - Accurate Patient-Specific Machine Learning Models of Glioblastoma Invasion Using Transfer Learning

Predict tumor Cell Density for patients with primary GBM

Data: Radiologic Imaging Representation: Auxiliary Datasets - private Integration: Transfer learning

This work implemented both univariate and multivariate individualized transfer learning predictive models, built in a Baysian framework, which harness the available population-level data but allow individual variability in their predictions (same methodology as 79).

2019 - Gaw - Integration of machine learning and mechanistic models accurately predicts variation in cell density of glioblastoma using multiparametric MRI

Predict tumor Cell Density for patients with primary GBM

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective

This work presented a novel hybrid model, based on multiparametric intensities, which combines machine learning (ML) with a mechanistic model of tumor growth (PI) to provide spatially resolved tumor cell density predictions. The ML component is an imaging data-driven graph-based semi-supervised learning model .

2019 - Gao - DeepCC- a novel deep learning-based framework for cancer Molecular subtype classification

Classify colorectal and breast cancer subtypes, i.e. cancer Molecular subtyping

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation

Transformed original gene expression profiles to a vector of enrichment scores by doing a gene set enrichment analysis (GSEA ) for each sample’s genes on thousands of gene sets from public knowledge database. This new feature vector quantify activities of biological pathways and is used as input to classifier.

2019 - Gao - A feature transfer enabled multi-task deep learning model on medical imaging

Mass classification, mass detection, and segmentation for breast cancer patients

Data: Radiologic Imaging Representation: Auxiliary Datasets - public Integration: Multitask learning

This work proposed an multitask deep learning model enabled by feature transfer. The proposed model utilized the different tasks from the same domain.

2019 - Deist - Simulation-assisted machine learning

Predict anti-cancer drug sensitivity (unknown cancer type)

Data: Molecular Representation: Scientific Knowledge - Mathematical Models Integration: Feature transformation

Generated simulations of each sample using the feature vector, then computed the similarity between the simulation output of each pair of samples. This similarity, rather than original high-dimensional features, is used to build a kernel for use in kernelized ML models. The simulation was done multiple times controlling for the uncertainty of simulation parameters. Shown effective when number of samples < number of features.

2019 - Cheng - Network-based prediction of drug combinations

Identify effective drug combinations for hypertension and cancer

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Graph models

This paper built a network of nearly two thousand Clinical ly approved drugs. Drug-target binding profiles were pooled from multiple data sources and the proximity between two drug’s targets were measured based on the interactome between the targets of each drug. Once the network was built, the topological relationship between two drug-target modules reflect whether drugs are pharmacologically distinct, complementary, indirectly similar, similar, or independent.

2019 - Campanella - Clinical -grade computational pathology using weakly supervised deep learning on whole slide images

Classification of cancer type into tumor vs non-tumor (breast, prostate)

Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning

The slide-level diagnosis are weak labels for the tiles within a slide. If a slide is positive, then at least one tile must contain tumor. If a slide is negative, then all tiles must not contain tumor. This can be formulated as multiple-instance learning (MIL) problem. This work trained a CNN with weak labels at tile-level. Ranks the tiles by their probability of being positive. Then learns a RNN by sequentially passing top-ranked tiles from each slide to integrate tile-level classification into slide-level.

2019 - Banerjee - Weakly supervised natural language processing for assessing patient-centered outcome following prostate cancer treatment

Annotate Clinical notes with patient outcomes (three classes: affirmed, negated, discussed risk) for patients with prostate cancer

Data: Clinical Representation: Auxiliary Datasets - public Integration: Weakly supervised learning

This work proposed a method to annotate Clinical notes without requiring ground-truth annotations. They used three domain-specific dictionaries to derive example artificial expressions for each class, which were used create weak labels for training data. A supervised model is then used to learn from weakly labeled training data. The proposed knowledge-guided model is compared with data-driven generative models. Used medical dictionary to clean the text data (combine related terms together, reduce sparsity).

2018 - Turki - A transfer learning approach via procrustes analysis and mean shiftfor cancer drug sensitivity prediction

Predict drug sensitivity for patients with myeloma and lung cancer.

Data: Molecular Representation: Auxiliary Datasets - private Integration: Transfer learning

Transferred half of the auxiliary data from a related task to be trained with the original dataset. In detail, performed CUR decomposition on original data to extract a low-rank low dimensional representation. Then, adjusted each sample from auxiliary data towards target domain using 1) kernel average of their neighbors from original dataset, and 2) SVD of the two datasets. Finally, selected a subset of samples from auxiliary data to be trained with the original dataset.

2018 - Rathore - Radiomic signature of infiltration in peritumoral edema predicts subsequent recurrence in glioblastoma: implications for personalized radiotherapy planning

Estimate peritumoral edema infiltration regions using radiomic signatures determined via ML

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Other

A classification model was trained to classify voxels into belonging to near-tumor ROI (N-ROI) or far-from-tumor ROI (F-ROI). This classification proability is used as the infiltration score. The assumption is that N-ROI and F-ROI are expected to have relatively higher and lower infiltration. The estimated infiltration maps are compared to recurring ROI (R-ROI) and nonrecurring ROI (NR-ROI) regions drawn by the experts.

2018 - Ma - OmicsMapNet- Transforming omics data to take advantage of Deep Convolutional Neural Network for discovery

Predict the grade of LGG and GBM tumor samples.

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation

Transformed gene data into images based on their functional hierarchy. Specifically, used the hierarchical gene structure from KEGG database to build a five-layer tree, mapping the gene expressions according to the child nodes of the tree (a gene can exists at multiple nodes). Then, converted the tree into an image using the pivot method based on their adjacency in the tree and sorting the genes according to the median values across all samples.

2018 - Ma - Multi-view Factorization AutoEncoder with Network Constraints for Multi-omic Integrative Analysis

Predict Progression-Free Interval event (a binary outcome that indicates the patient had a new tumor event in a fixed period, implying the treatment outcome is unfavorable)

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective

Extracted a Molecular interaction network from external knowledge database. Used Laplacian regularizer to force consistency between the learned feature representation and the feature interaction network from domain knowledge. Additionally, used patient similarity networks to regularize the same set of patients are consistent with each other.

2018 - Lyu - Deep Learning Based Tumor Type Classification Using Gene Expression Data

Classify tumor type (tumor vs normal) for multiple types of cancer

Data: Molecular Representation: Scientific Knowledge - Other Integration: Feature transformation

Transformed gene data into images based on their relative position according to their chromosome numbers. Each rectangle represents one gene and the value is normalized gene expression level. The intuition is that adjacent genes are more likely to interact with each other.

2018 - Liu - Integration of biological and statistical models toward personalized radiation therapy of cancer

Predict treatment outcome (rectal complication) for patients with prostate cancer

Data: Clinical Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective

Proposed a general framework to integrate biological equations with patient-specific data. First, identified parameters in the biological model that are meaningful to be personalized, i.e. learnt from patient-specific data. Then, derived the optimization problem that predicts the response variable based on the biological model but uses personalized parameters estimated from patient-specific data. Knowledge was also used to define nonnegativity constraints of certain variables.

2018 - Hu - Accurate Patient-Specific Machine Learning Models of Glioblastoma Invasion Using Transfer Learning

Predict Tumor Cell Density for patients with GBM

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Customized model pipelines

Used Transfer Learning to build one model for each patient then biasing the model towards the population pattern using a Bayesian framework. Also, constrained knowledge transfer from patients with a certain threshold for correlation between one imaging measure and the response.

2018 - Hao - PASNet- pathway-associated sparse deep neural network for prognosis prediction from high-throughput data

Predict survival for patients with GBM

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Biologically-informed neural networks

The authors extracted 574 pathways and 4359 genes from public databases. A deep learning model called PASNet is designed using these pathways. The architecture consists of a gene layer, a pathway layer, a hidden layer, and an output layer.

2018 - Dincer - DeepProfile- Deep learning of cancer Molecular profiles for precision medicine

Predict drug response for patients with acute myeloid leukemia (AML)

Data: Molecular Representation: Auxiliary Datasets - public Integration: Transfer learning

Trained a VAE using large amount of public unlabeled data from other patients with the same cancer type to learn a robust latent feature representation. Then used these feature vector to train drug response model.

2018 - Babier - Knowledge‐based automated planning for oropharyngeal cancer

Predict 3D dose treatment plans (applied to oropharyngeal cancer patients)

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other

First, a knowledge-based planning (KBP) model (bagging query (BQ) method or the generalized principal component analysis-based (gPCA) method) was trained to predict the dose distribution. These predictions from KPB model were then passed through an optimization pipeline to generate the ﬁnal ﬂuence-based treatment plans.

2018 - Babier - Knowledge-based automated planning with three-dimensional generative adversarial networks

Predict 3D dose treatment plans (applied to oropharyngeal cancer patients)

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other

First, GAN was trained to predict the dose distribution given a contoured CT image. These predictions from generators of GAN were then passed through an optimization pipeline to generate the ﬁnal ﬂuence-based treatment plans.

2018 - Abid - Exploring patterns enriched in a dataset with contrastive principal component analysis

Find trends and variationsin gene expression levels of cancer patients that are useful t identify Molecular subtypes of cancer.

Data: Molecular Representation: Auxiliary Datasets - private Integration: Transfer learning

This paper proposed contrastive PCA (cPCA), which finds low-dimensional sturcutres that are enriched in the dataset relative to comparison data. The method is applied to data of cancer patients to find low-dimensional representations that separate them from the data of healthy controls from a similar demographic background. cPCA can be used as the feature extraction method and for denoising, where the backgroudn dataset serves to remove the universal but uninsteresting variation in the target dataset.

2017 - Yang - Literature-based discovery of new candidates for drug repurposing

Discover new candidates for drug repurposing

Data: Treatment/Clinical Representation: Knowledge from Experts - Quantitative Integration: Feature transformation

Assuming that extensive knowledge is hidden in the large-scale literature, the authors propose a relationship extraction method that allows the detection of indirect relationships for drug repurposing. Articles that discussed disease-gene, gene-drug, and disease-drug relationships were pulled from the MEDLINE database (5.4 million documents). A list of relevant entity names was generated from the Therapeutic Target Database (2k diseases, 3k targets, 20k drugs). The Comparative Toxicogenomics Database provided >1 million chemical-gene, gene-disesase, chemical-disease relationships, formatted as triplets. NLP parser tool were used to convert sentences of target documents into the smallest common subtrees in dependency format. Then, drugs are represented in a numerical vector space based on connections ot the linked genes. Then, promising repurposing drug candidates were identified by computing the similarity of repurposed drug to approved drugs using Jacard Index computed on their binary representations. In their experiments, the identified candidates were compared with the literature and ClinicalTrials.gov database.

2017 - Padmanabhan - Reinforcement learning-based control of drug dosing for cancer chemotherapy treatment

Control cancer chemotherapy drug dosing.

Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-in-the-loop

This work developed a RL-based control strategy for cancer chemotherapy treatment doses. A mathematical model of cancer drug dynamics was used to provide feedback to the RL-based controller.

2017 - Min - Predicting activities of daily living for cancer patients using an ontology-guided machine learning methodology

Predict activities of daily living disabilities for the first year after a patient’s cancer diagnosis

Data: Clinical Representation: Auxiliary Datasets - public Integration: Feature transformation

Mapped data into concept unique identifiers (CUI) of UMLS, which helps translate medical terms among diverse data sources. Identified base concepts, and from them, extracted hierarchies of related concepts. Applied AQ21 algorithm to learn generalizable rules that can predict disabilities.

2017 - Grady - Experience Based Quality Control in IMRT Treatment Planning of High Risk Post-Prostatectomy Prostate Cancer with RapidPlan

Develop a knowledge based planning (KBP) model for the treatment plan of high risk post-prostatectomy prostate cancer

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature transformation

An initial dosimetric analysis was carried out to identify high quality plans from Clinical trial. Then, Principal Component Analysis (PCA) is used to characterize the salient features of the patient anatomy (CT) and dose distribution (treatment plans) and Support Vector Regression (SVR) is used to model their correlation.

2017 - Geeleher - Discovering novel pharmacogenomic biomarkers by imputing drug response in cancer patients from large genomics studies

Predict drug response and discover novel biomarkers in Clinical cancer datasets; Classification of tumor samples for breast, blood, lung, skin, and central nervous system cancers.

Data: Molecular Representation: Auxiliary Datasets - private Integration: Transfer learning

PreClinical datasets are small but more straightforward to collect drug response. The goal is to use large Clinical datasets for pharmacogenomics discovery without having to collect drug response data. Used a preClinical dataset to train a model for drug response and gene expression, then applied the estimated model to impute drug response for the Clinical dataset of interest. This strategy assumes that if it is possible to predict drug response using cell line-derived gene expression, it should be possible for similar models.

2017 - Ferranti - The value of prior knowledge in machine learning of complex network systems

Predict patient drug response and phenotype (both as classification)

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature grouping

Unlike other papers that uses real data, this paper used synthetic data to evaluate whether using knowledge can improve performance of ML models. They used synthetic biological networks to generate data and compared models without knowledge (using data from all nodes to predict the super-nodes) vs. with knowledge (using nodes that are known to be directly connected to the super-nodes).

2017 - Esteva - Dermatologist-level classification of skin cancer with deep neural networks

Classify disease subtype for patients with skin lesion (3-way and 9-way classication)

Data: Radiologic Imaging Representation: Auxiliary Datasets - public Integration: Customized model pipelines; Data - Transfer learning

The CNN model is first pre-trained on 1.51 natural images and then fine-tuned on medical images. The authors used the Clinical taxonomy to partition disease subtypes into fine-grained training classes (as many as +700 classes) because training on finer classes showed improved classification. A deep learning model is trained predict the finer classes, then, aggregated the prediction of finer classes to recover the prediction of coarser classes. However, this method requires labels for all subclasses to be available.

2017 - Corredor - Training a cell-level classifier for detecting basal-cell carcinoma by combining human visual attention maps with low-level handcrafted features

Classification of cancer type into tumor vs non-tumor (skin)

Data: Pathologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Feature selection

This model extracts a visual attention map from actions performed by pathologists while navigating WSIs to inform the low-level feature algorithm where in the images to initially “look”. First, segmented all nuclei from the WSI and computed their low-level features. When pathologists look at the images, they can only see a window of reference instead of the entire image. Each nucleus is assigned a likelihood of cancerous based on the number of times the nucleus is visited by pathologists. Then, computed the importance of each feature by performing linear regression of features with the likelihood. The classifier SVM was trained using weighted-features.

2017 - Azizi - Transfer learning from RF to B-mode temporal enhanced ultrasound features for prostate cancer detection

detect prostate cancer using tempoeral enhanced ultrasound data

Data: Radiologic Imaging (ultrasound) Representation: Auxiliary Datasets - private Integration: Data - Transfer learning

The authors trained a tissue classificaiton model on tempoeral enhanced ultrasound radiofrequency data, and deployed for classification using B-mode data collected from the same ultrasound scanner. Unsupervised domain adaptation and transfer learning techniques were used. It starts with a layer of PCA to whiten the spectral features and align the source and target PCA subspaces. Then, a shared deep belief networks (DBN) was trained using both source and target data. The loss function minimizes the reconstruction error for both source and target domain data, and minimizes the divergence between the source and target distributions. Such capability is meaningful because the radiofrequency data is not readily available on all commercial scanners in clinics.

2017 - Ammad-ud-din - Systematic identification of feature combinations for predicting drug response with Bayesian multi-view multi-task linear reg

Predict anti-cancer drug response and identify which feature combinations are the most predictive

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature grouping

Extracted functional-linked-networks from prior biological knowledge and used these to select features and split the data into “views”. Then used a multi-view regression, where each feature has a feature-specific weight and a view-specific weight, to predict drug response.

2016 - Shiraishi - Knowledge‐based prediction of three‐dimensional dose distributions for external beam

Predict three-dimensional dose distributions for external beam radiotherapy

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature selection

Using previously treated plans as training data, this work proposed an artificial neural network (ANN) to predict a dose matrix based on patient-specific geometric and planning parameters, such as the closest distance (r) to planning target volume (PTV) and organ-at-risks (OARs).

2016 - Kim - Integrating Domain Specific Knowledge and Network Analysis to Predict Drug Sensitivity of Cancer Cell Lines

Predict anticancer drug response (resistance vs sensitive) of cell lines

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature selection

Selected features (genes) based on domain knowledge about which genes are apoptotic. Conducted literature search of prescribed Clinical dose of the drugs. Then used this dose to derive two cutoff values to obtain binary labels for drug sensitivity and drug resistance.

2016 - Ammad-ud-din - Drug response prediction by inferring pathway-response associations with kernelized Bayesian matrix factorization

Predict anti-cancer drug response and generate hypothesis between drug and pathway activation

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature grouping

Used prior knowledge to split data into components such that each group represented one pathway. Then used component-wise multiple kernel learning to predict drug response. The model learns component-specific kernel weights, which can be used to infer hypothesis of pathways (components) associated with drug responses.

2016 - Aben - TANDEM a two-stage approach to maximize interpretability of drug response models based on multiple Molecular data types

Predict anti-cancer drug response

Data: Molecular Representation: Knowledge from Experts - Quantitative Integration: Customized model pipelines

Used a two-stage approach to improve the interpretability of prediction models and have more balanced contribution of data types. Separated features into upstream data (somatic mutations, CAN, etc., significantly associated with response) and downstream data (gene expression). The downstream is trained on the remaining variability of the upstream model.

2015 - Zou - A Transfer Learning Approach for Predictive Modeling of Degenerate Biological Systems

Predict relationship between TF binding sites and gene expression (cancer related cell lines) by other related cell lines

Data: Molecular Representation: Auxiliary Datasets - private Integration: Transfer learning

This work proposed a unique prior for the model coefﬁcients of the old domains and new domain. This prior has two hyperparameters characterizing the degeneracy and the correlation structure of the domains, respectively. This work proposed to use a graph to represent the qualitative knowledge about degeneracy (this is knowledge about the features, used to generate a graph and regularize the model) and set the corresponding hyperparameter to be the Laplacian matrix of the graph.

2015 - Shiraishi - Knowledge‐based prediction of plan quality metrics in intracranial stereotactic

Predict achievable dose–volume histograms (DVHs) and highly precise DVH-based quality metrics (QMs) in stereotactic radiosurgery/radiotherapy (SRS/SRT) plans

Data: Treatment Representation: Scientific Knowledge - Mathematical Models Integration: Knowledge-regularized objective

This work proposed mathematical models to predict achievable DVHs based on the geometric relationship of organs-at-risk (OARs) to planning target volumes (PTVs) for plans with/without OAR involvement respectively.

2015 - Morino - Predicting disease progression from short biomarker series using expert advice algorithm

Predict progression of patients with prostate cancer

Data: Clinical Representation: Auxiliary Datasets - private Integration: Transfer learning

This model predicts for a target patinet with short, unstable time-series data of biomarkers, while transferring knowledge learnt from other patients with longer time series data who suffered from the same disease. Two datasets were used to construct "experts" (models): a dataset of Canadian patients and another of japanese patients. The experts are weighted based on the similarity of their observations with the target patient.

2015 - Krayenbuehl - Evaluation of an automated knowledge based treatment planning system for head and neck

Predict treatment plans for patients with head and neck cancer

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other

This work re-planned the treatments for patients with head and neck cancer with Pinnacle Auto-Planning (AP) version 9.10. Only one single cycle of plan optimization using one single template was allowed for AP. The dose plans were evaluated. Additionally, two experienced radiation oncologists blind-reviewed and selected preferred plans.

2015 - Cohen - Mathematical Modelling of Molecular Pathways Enabling Tumour Cell Invasion and Migration

Predict the probability of metastasis and understand how genetic interactions between gene mutations can affect the probability of metastasis

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Customized model pipelines

Constructed an “influence network” using prior knowledge. The network is composed of nodes (biochemical species or phenotypes) and edges (activations or inhibitory influence). This network is modeled with Markov decision process with pre-assigned state transition probabilities. Simulated random walks on the network to obtain probabilities to reach a phenotype. Also analyzed the stable states of the network for interpretation of meaningful pathways.

2015 - Chen - Enhanced Performance of Brain Tumor Classification via Tumor Region Augmentation and Partition

Classification of tumor category for patients with brain tumor

Data: Radiologic Imaging Representation: Knowledge from Experts - Qualitative Integration: Feature transformation

Based on biological knowledge, tumors of the same category are often located in similar places, such as gliomas usually involve white matter; meningiomas are typically adjacent to skull, gray matter, and cerebrospinal fluid; and pituitary tumors are often adjacent to sphenoidal sinus, internal carotid arteries, and optic chiasma. The authors proposed "tumor region augmentation and partition", which (i) enlarges the tumor ROI to capture possibly informative context, and (ii) partitions it into ring-form subregions, features from each subregion are weighted and concatenated to get final features.

2014 - Xu - Weakly supervised histopathology cancer image segmentation and classification

Classification of tumor type (cancer vs non-cancer image), image segmentation (cancer vs non-cancer tissue), and patch-level clusteirng (different classes). (Colo cancer

Data: Pathologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Weakly supervised learning

This paper used Multiple Instance Learning (MIL) to learn patch-level labels when only image-level labels are available. Adds an additional term of the loss that encourage nearby image patches within each bag to be in the same cluster. Assumes that multiple cancer subtypes with different morphological characteristics might co-exists in a WSI image. This model integrates the clustering concept into the MIL setting by assuming that exists a hidden variable that denotes the cluster of each positive instance. One classifier is trained for each cluster and each cluster's instances are weighted differently.

2014 - Wu - Improved robotic stereotactic body radiation therapy plan quality and planning efficacy for organ-confined prostate cancer utilizing overlap-volume histogram-driven planning methodology

Improve robotic stereotactic body radiation therapy plan quality and planning efﬁcacy for organ-conﬁned prostate cancer

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other

A database containing Clinical ly-delivered, robotic SBRT plans of patients with prostate cancer was used as a cohort to establish an organ’s distance-to-dose model. The overlap-volume histogram (OVH)-driven planning methodology was reﬁned by adding the PTV volume factor to counter the target’s dose fall-off effect and incorporated into Multiplan to automate SBRT planning.

2014 - Tol - Evaluation of a Knowledge-Based Planning Solution for Head and Neck Cancer

Predict treatment plans for patients with head and neck cancer

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other

Knowledge-based plans were created by RapidPlan. RapidPlan results were compared with Clinical plans (CP).

2014 - Kwon - Combining generative models for multifocal glioma segmentation and registration

Segmentation of gliobastoma tumor from MRI

Data: Radiologic Imaging Representation: Scientific Knowledge - Mathematical Models Integration: Simulation

This is a generative model for tumor segmentation. This model grows tumors on an initialized atlas, and estimates parameters to grow the tumor via the diffusion-reaction-advection model, and registers the scans to this atlas to infer segmentation. Tumor shape prior is estimated via random walk with restart and incoporated into the model.

2014 - Kim - Knowledge boosting- a graph-based integration approach with multi-omics data and genomic knowledge for cancer Clinical outcome prediction

Classification of ovarian cancer grade (low vs high)

Data: Molecular Representation: Auxiliary Datasets - public Integration: Graph models

This work proposed a graph-based framework for integrating multi-omics data and genomic knowledge to improve power in predicting Clinical outcomes and elucidate interplay between different levels.

2013 - Wu - Using overlap volume histogram and IMRT plan data to guide and automate VMAT planning A head-and-neck case study

Predict treatment plans for patients with head and neck cancer

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other

Based on comparable head-and-neck dosimetric results between planner-generated VMAT and IMRT plans, an inhouse developed, overlap volume histogram (OVH)-driven automated planning application containing a database of prior Clinical head-and-neck IMRT plans was built into Pinnacle 3 SmartArc for volumetricmodulated arc therapy (VMAT) planning.

2013 - Good - A Knowledge-Based Approach to Improving and Homogenizing Intensity Modulated Radiation Therapy Planning Quality Among Treatment Centers- An Example Application to Prostate Cancer Planning

Predict optimal radiation therapy dose distributions (3D) of prostate cancer patients by matching and deforming the cases in the (assembled) knowledge base.

Data: Pathologic Imaging Representation: Auxiliary Datasets - private Integration: Feature transformation

A knowledge database was created from treatment plans for prostate cancer at their institute. For each “query” case from outside database, a similar “match” case was identiﬁed in the knowledge database, and the match case’s plan parameters were then adapted and optimized to the query case by using ﬁnal dose-volume planning constraints.

2012 - Yuan - Quantitative analysis of the factors which affect the interpatient organ‐at‐risk dose

Predict intensity modulated radiation therapy (IMRT) plans for patients with prostate or head-and-neck (HN) cancer

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature transformation

This work formulized the dependence of organs-at-risk (OAR) dose volume histograms (DVHs) on patient anatomical factors into feature models which were learned from prior plans by a stepwise multiple regression method. Features were extracted by PCA.

2012 - Cun - Prognostic gene signatures for patient stratification in breast cancer- accuracy, stability and interpretability of gene selection app

Predict metastasis/relapse for patients with breast cancer and interpret which genes are most important for the prediction

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Other

Compared 14 published methods (8 using network knowledge) on 6 public breast cancer datasets. Found that the incorporation of prior knowledge could generally not significantly improve prediction accuracy but sometimes improve gene selection stability and biological interpretability.

2011 - Zhu - A planning quality evaluation tool for prostate adaptive IMRT based on machine learning

Generate dose volume histograms (DVH) of organs-at-risk based on prior plans

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Feature transformation

Principal component analysis was applied to DVHs and distance-to-target histogram (DTH) to quantify salient features. Then, support vector regression was implemented to establish the correlation between the features of the DVH and the anatomical information (organ volumes and DTH).

2011 - Wu - Data-driven approach to generating achievable dose-volume histogram objectives in intensity-modulated radiotherapy planning

Predict IMRT treatment plans for patients with head and neck cancer

Data: Treatment Representation: Knowledge from Experts - Quantitative Integration: Other

The overlap volume histogram (OVH) was used to compare the spatial relationships between the organs at risk and targets of a new patient with those of previous patients in a database. Clinical plans were compared with OVH-assisted plans (OPs).

2011 - Chanyavanich - Knowledge‐based IMRT treatment planning for prostate cancer

Predict IMRT treatment planning for prostate cancer

Data: Radiologic Imaging Representation: Knowledge from Experts - Quantitative Integration: Feature transformation

In this study, the quality of each of treatment plan was evaluated by comparing the dose–volume histograms of the new semiautomated plan to that of the original plan developed manually by an expert (human) planner for the query case. The ﬁnal step consisted of optimizing the query case using the constraints and priorities from the reference case. A database was assembled into an information theoretic system. An algorithm based on mutual information was implemented to identify similar patient cases. Once the best match is identiﬁed, that Clinical ly approved plan is used to generate the new plan.

2010 - Vaske - Inference of patient-specific pathway activities from multi-dimensional cancer genomics data using PARADIGM

Understand patient-specific genetic interactions and identify relevant pathways involved in cancer progression (GBM and breast cancer)

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature transformation

First converted the biologically known pathways into a directed graph and used it to construct factor graphs composed of observed entities and hidden entities. For each pathway, trained the factor graph for each patient using EM.

2010 - Johannes - Integration of pathway knowledge into a reweighted recursive feature elimination approach for risk stratification of cancer patie

Predict relapse for patients with breast cancer (risk stratification) and identify prognostic gene signatures

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature selection

Assumes that a gene with a low fold-change should have more influence if connected to differentially expressed genes. This method uses GeneRank, adapted from Google’s PageRank, to rank features based on prior knowledge (PPI) and fold-change information from data. This ranking is then combined with the standard SVM-RFE to perform feature selection.

2009 - Zhu - Network-based support vector machine for classification of microarray samples

Predict metastasis (classification) for patients with breast cancer.

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective

Assumes that neighboring genes in a network tend to function together in biological processes. Thus, given a network of genes, treated two neighboring genes in a network as one group. Applied F_infinity norm to each group of two features encouraging 1) grouped feature selection, 2) genes in the same group to have similar coefficients, 3) genes with more direct neighbors to have larger coefficients .

2009 - Guan - Lung cancer gene expression database analysis incorporating prior knowledge with support vector machine-based classification method

Classify disease subtype (cancer versus malignant pleural mesothelioma) for patients with lung cancer

Data: Molecular Representation: Auxiliary Datasets - public Integration: Feature selection

Used prior knowledge about lung disease related genes to select feature candidates, then tested the candidates through multiple testing and combined with feature genes selected by another method. The final set of features is used to run SVM.

2009 - Binder - Incorporating pathway information into boosting estimation of high-dimensional risk prediction models

Predict survival of lymphoma and ovarian cancer patients

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Knowledge-regularized objective

Assumes that connected genes in a network should have similar penalties. This PathBoost method uses pathway knowledge in the likelihood-based boosting procedure to perform gene selection. In detail, adapts the penalty for features that correspond to genes that have a connection to another gene by increasing the penalty for a selected feature and decreasing the penalty for connected feature.

2007 - Chuang - Network-based classification of breast cancer metastasis

Discover prognostic biomarkers for breast cancer; predict metastatic vs non-metastatic tumor samples

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature grouping

The gene expression profiles of tissue samples were mapped into the large network. Identified subnetworks as single connected components in the network and computed an activity score for each subnetwork by averaging the gene expression levels. Additionally, measured the discriminative score for each subnetwork as the mutual information of the activity score and the class label. Used greedy search to identify subnetworks whose discrimination power is locally maximal.

2005 - Guo - Towards precise classification of cancers based on robust gene functional expression profiles

Classify tumor type of multiple types of cancer

Data: Molecular Representation: Auxiliary Datasets - public Integration: Feature grouping

First, mapped genes to their categories in Gene Ontology, which are called functional modules. For each module, constructed a representative functional feature by summary measures (mean and median). These features are then used to train a classifier (random forest).

2002 - Lee - Gene selection: a Bayesian variable selection approach

Discover significant genes on breast tumors from patients carrying or not carrying mutations in the predisposing genes patients not expected

Data: Molecular Representation: Scientific Knowledge - Knowledge Graph Integration: Feature selection

This paper proposed a hierarchical Bayesian model for gene selection. It tackles the limitation that genetic is high-dimensional and previous method tend to select a large number of genes. Rather than constraining the dimension or number of important genes, the authors assign a prior distribution over gene coefficients, which allows more flexibilty. The inclusino of each feature is represented via a latent indicator. Prior knowledge about genes that are important or pairs of genes that are related can be used to specify this latent indicator.

Knowledge-Informed Machine Learning for Cancer Applications

Machine Learning for Cancer

A Review of KIML for Cancer

Live Summary Table