PathOmics: Pathology-and-genomics Multimodal Transformer for Survival Outcome Prediction

The official code of "Pathology-and-genomics Multimodal Transformer for Survival Outcome Prediction" (Accepted to MICCAI2023, top 9%).

Our Paper [Link]

[2025.12 New Update!!!] We updated the paper list of pathology-and-genomics multimodal analysis approaches in healthcare at the end of this repo.

Workflow overview of the PathOmics

Workflow overview of the pathology-and-genomics multimodal transformer (PathOmics) for survival prediction. In (a), we show the pipeline of extracting image and genomics feature embedding via an unsupervised pretraining towards multimodal data fusion. In (b) and (c), our supervised finetuning scheme could flexibly handle multiple types of data for prognostic prediction. With the multimodal pretrained model backbones, both multi- or single-modal data can be applicable for our model fine-tuning.

Citation

@inproceedings{ding2023pathology,
  title={Pathology-and-genomics multimodal transformer for survival outcome prediction},
  author={Ding, Kexin and Zhou, Mu and Metaxas, Dimitris N and Zhang, Shaoting},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={622--631},
  year={2023},
  organization={Springer}
}

Prerequisites

python 3.8.18
Pytorch 2.0.1
pytorch-cuda 11.8
Torchvision 0.15.2
Pillow 9.4.0
numpy 1.24.3
pandas 2.0.3
scikit-survival 0.21.0 
scikit-learn 1.2.0
h5py 2.8.0

Usage

Data prerpocessing

1. Download WSIs from TCGA-COAD and TCGA-READ.
2. Download genomics data from CbioPortal and move the downloaded folder into "PathOmics" folder.
* "coadread_tcga_pan_can_atlas_2018" in `bash_main.py` and `bash_main_read.py` is the downloaded folder, please download it before you run the code.
3. Split WSIs into patches and only keep the foreground patches.
4. Extract patch features via pretrained models (e.g., ImageNet-pretrained ResNet50, ResNet101, etc).
5. Save patch features as .npz files. (For each slide, we generate one .npz file to save patch features).

For more details about extracting feature, please check Issue 1 and the code in split_tiles_utils/helper.py

Run code on TCGA-COAD only

Model will be pretrained and finetuned on theTCGA-COAD training set (4-fold cross-validation). The finetuned model will be evaluated on the TCGA-COAD hold-out set.

python bash_main.py --pretrain_loss 'MSE' --save_model_folder_name 'reproduce_experiments' --experiment_folder_name 'COAD_reproduce' --omic_modal 'miRNA' --kfold_split_seed 42 --pretrain_epochs 25 --finetune_epochs 25 --model_type 'PathOmics' --model_fusion_type 'concat' --model_pretrain_fusion_type 'concat' --cuda_device '2' --experiment_id '1' --use_GAP_in_pretrain_flag --seperate_test

Run code on TCGA-COAD and TCGA-READ

Model will be pretrained on TCGA-COAD (5-fold cross-validation). Model will be finetuned, validated, and evaluated on the TCGA-READ dataset.

python bash_main_read.py --k_fold 5 --fusion_mode 'concat' --prev_fusion_mode 'concat' --pretrain_loss 'MSE' --save_model_folder_name 'reproduce_experiments' --experiment_folder_name 'READ_reproduce' --omic_modal 'miRNA' --kfold_split_seed 42 --pretrain_epochs 25 --finetune_epochs 25 --model_type 'PathOmics' --cuda_device '2' --experiment_id '1' --use_GAP_in_pretrain_flag

If you want to use TCGA-COAD pretrain weights and skip the pretraining stage, please add --load_model_finetune into your script. Please modify the code to ensure your pretrain weights saving directory is correct.

Use data-efficient mode in finetuning stage

Please add --less_data into your script and set --finetune_test_ratio as your preferred ratio for indicating the ratio of data used for model finetuning.

Literature reviews of pathology-and-genomics multimodal analysis approaches in healthcare.

|Publish Date|Title|Paper Link|Code| |---|---|---|---| |2025.12|Pathway-Aware Multimodal Transformer (PAMT): Integrating Pathological Image and Gene Expression for Interpretable Cancer Survival Analysis|TPAMI|Code| |2025.12|KPVG: Knowledge-prompted vision-genomics model for cancer survival prediction in whole slide images|Information Fusion|NA| |2025.12|Cancer Survival Analysis via Zero-shot Tumor Microenvironment Segmentation on Low-resolution Whole Slide Pathology Images|NeurIPS2025|NA| |2025.10|Multimodal Hypergraph Guide Learning for Non-Invasive ccRCC Survival Prediction|MICCAI2025|Code| |2025.10|Memory-Augmented Incomplete Multimodal Survival Prediction via Cross-Slide and Gene-Attentive Hypergraph Learning|MICCAI2025|Code| |2025.10|Disentangled and Interpretable Multimodal Attention Fusion for Cancer Survival Prediction|MICCAI2025|Code| |2025.10|MurreNet: Modeling Holistic Multimodal Interactions Between Histopathology and Genomic Profiles for Survival Prediction|MICCAI2025|NA| |2025.10|PS3: A Multimodal Transformer Integrating Pathology Reports with Histology Images and Biological Pathways for Cancer Survival Prediction|ICCV2025|NA| |2025.10|Comparison of Digital Histology AI Models with Low-Dimensional Genomic and Clinical Models in Survival Modeling for Prostate Cancer|ICCV2025|NA| |2025.09|HSFSurv: A hybrid supervision framework at individual and feature levels for multimodal cancer survival analysis|MIA|Code| |2025.08|POMP: Pathology-omics Multimodal Pre-training Framework for Cancer Survival Prediction|IJCAI2025|Code| |2025.06|Distilled Prompt Learning for Incomplete Multimodal Survival Prediction|CVPR2025|Code| |2025.06|Robust Multimodal Survival Prediction with Conditional Latent Differentiation Variational AutoEncoder|CVPR2025|Code| |2025.06|A machine learning approach for multimodal data fusion for survival prediction in cancer patients|NPJ Precision Oncology|Code| |2025.04|Clustering-Enhanced Multimodal Pre-Training for Histology-Gene Joint Representation Learning|ISBI2025|NA| |2025.04|Pathogen-X: A Cross-Modal Genomic Feature Trans-Align Network for Enhanced Survival Prediction from Histopathology Images|ISBI2025|NA| |2025.01|Decoding pan-cancer treatment outcomes using multimodal real-world data and explainable artificial intelligence|Nature Cancer|Code| |2025.01|Histo-Genomic Knowledge Association For Cancer Prognosis From Histopathology Whole Slide Images|TMI|Code| |2024.10|MoME: Mixture of Multimodal Experts for Cancer Survival Prediction|MICCAI 2024|Code| |2024.10|PG-MLIF: Multimodal Low-rank Interaction Fusion Framework Integrating Pathological Images and Genomic Data for Cancer Prognosis Prediction|MICCAI 2024|Code| |2024.10|Multimodal Cross-Task Interaction for Survival Analysis in Whole Slide Pathological Images|MICCAI 2024|Code| |2024.10|HistGen: Histopathology Report Generation via Local-Global Feature Encoding and Cross-modal Context Interaction|MICCAI 2024|NA| |2024.07|A Multimodal Knowledge-enhanced Whole-slide Pathology Foundation Model|ArXiv|NA| |2024.06|Transcriptomics-guided Slide Representation Learning in Computational Pathology|CVPR 2024|Code| |2024.06|Modeling Dense Multimodal Interactions Between Biological Pathways and Histology for Survival Prediction|[CVPR 2024](https://openaccess.thecvf.com/content/CVPR2024/papers/Jaume_Modeling_Dense_Multimodal_Interactions_Between_B

PathOmics

Install / Use

README