Pathomic Fusion: An Integrated Framework for Fusing Histopathology and Genomic Features for Cancer Diagnosis and Prognosis
- 05/26/2021: Updated Google Drive with all models and processed data for TCGA-GBMLGG and TCGA-KIRC. found using the following link
© Mahmood Lab - This code is made available under the GPLv3 License and is available for non-commercial academic purposes.
If you find our work useful in your research or if you use parts of this code please consider citing our paper:
@article{chen2020pathomic,
title={Pathomic fusion: an integrated framework for fusing histopathology and genomic features for cancer diagnosis and prognosis},
author={Chen, Richard J and Lu, Ming Y and Wang, Jingwen and Williamson, Drew FK and Rodig, Scott J and Lindeman, Neal I and Mahmood, Faisal},
journal={IEEE Transactions on Medical Imaging},
year={2020},
publisher={IEEE}
}
We propose a simple and scalable method for integrating histology images and -omic data using attention gating and tensor fusion. Histopathology images can be processed using CNNs or GCNs for parameter efficiency or a combination of the the two. The setup is adaptable for integrating multiple -omic modalities with histopathology and can be used for improved diagnostic, prognostic and theriputic response determinations.
- Linux (Tested on Ubuntu 18.04)
- NVIDIA GPU (Tested on Nvidia GeForce RTX 2080 Tis on local workstations, and Nvidia V100s using Google Cloud)
- CUDA + cuDNN (Tested on CUDA 10.1 and cuDNN 7.5. CPU mode and CUDA without CuDNN may work with minimal modification, but untested.)
- torch>=1.1.0
- torch_geometric=1.3.0
The code base structure is explained below:
- train_cv.py: Cross-validation script for training unimodal and multimodal networks. This script will save evaluation metrics and predictions on the train + test split for each epoch on every split in checkpoints.
- test_cv.py: Script for testing unimodal and unimodal networks on only the test split.
- train_test.py: Contains the definitions for "train" and "test".
- networks.py: Contains PyTorch model definitions for all unimodal and multimodal network.
- fusion.py: Contains PyTorch model definitions for fusion.
- data_loaders.py: Contains the PyTorch DatasetLoader definition for loading multimodal data.
- options.py: Contains all the options for the argparser.
- make_splits.py: Script for generating a pickle file that saves + aligns the path for multimodal data for cross-validation.
- run_cox_baselines.py: Script for running Cox baselines.
- utils.py: Contains definitions for collating, survival loss functions, data preprocessing, evaluation, figure plotting, etc...
The directory structure for your multimodal dataset should look similar to the following:
./
├── data
└── PROJECT
├── INPUT A (e.g. Image)
├── image_001.png
├── image_002.png
├── ...
├── INPUT B (e.g. Graph)
├── image_001.pkl
├── image_002.pkl
├── ...
└── INPUT C (e.g. Genomic)
└── genomic_data.csv
└── checkpoints
└── PROJECT
├── TASK X (e.g. Survival Analysis)
├── path
├── ...
├── ...
└── TASK Y (e.g. Grade Classification)
├── path
├── ...
├── ...Depending on which modalities you are interested in combining, you must: (1) write your own function for aligning multimodal data in make_splits.py, (2) create your DatasetLoader in data_loaders.py, (3) modify the options.py for your data and task. Models will be saved to the checkpoints directory, with each model for each task saved in its own directory. At the moment, the only supervised learning tasks implemented are survival outcome prediction and grade classification.
Here are example commands for training unimodal + multimodal networks.
Example shown below for training a survival model for mode A and saving the model checkpoints + predictions at the end of each split. In this example, we would create a folder called "CNN_A" in "./checkpoints/example/" for all the models in cross-validation. It assumes that "A" is defined as a mode in dataset_loaders.py for handling modality-specific data-preprocessing steps (random crop + flip + jittering for images), and that there is a network defined for input A in networks.py. "surv" is already defined as a task for training networks for survival analysis in options.py, networks.py, train_test.py, train_cv.py.
python train_cv.py --exp_name surv --dataroot ./data/example/ --checkpoints_dir ./checkpoints/example/ --task surv --mode A --model_name CNN_A --niter 0 --niter_decay 50 --batch_size 64 --reg_type none --init_type max --lr 0.002 --weight_decay 4e-4 --gpu_ids 0
To obtain test predictions on only the test splits in your cross-validation, you can replace "train_cv" with "test_cv".
python test_cv.py --exp_name surv --dataroot ./data/example/ --checkpoints_dir ./checkpoints/example/ --task surv --mode input_A --model input_A_CNN --niter 0 --niter_decay 50 --batch_size 64 --reg_type none --init_type max --lr 0.002 --weight_decay 4e-4 --gpu_ids 0
Example shown below for training a grade classification model for fusing modes A and B. Similar to the previous example, we would create a folder called "Fusion_AB" in "./checkpoints/example/" for all the models in cross-validation. It assumes that "AB" is defined as a mode in dataset_loaders.py for handling multiple inputs A and B at the same time. "grad" is already defined as a task for training networks for grade classification in options.py, networks.py, train_test.py, train_cv.py.
python train_cv.py --exp_name surv --dataroot ./data/example/ --checkpoints_dir ./checkpoints/example/ --task grad --mode AB --model_name Fusion_AB --niter 0 --niter_decay 50 --batch_size 64 --reg_type none --init_type max --lr 0.002 --weight_decay 4e-4 --gpu_ids 0
To reporduce the results in our paper and for exact data preprocessing, implementation, and experimental details please follow the instructions here: ./data/TCGA_GBMLGG/. Processed data and trained models can be downloaded here.
- Please report all issues on the public forum.
This project is licensed under the GNU GPLv3 License - see the LICENSE.md file for details
- A provisional patent on this work has been filed by the Brigham and Women's Hospital.
- This code is inspired by SALMON and SCNN, code base structure was inspired by pytorch-CycleGAN-and-pix2pix.
- Subsidized computing resources for this project were provided by Nvidia and Google Cloud.