POPPSlide: Patient Outcome Prediction Pipeline using Whole Slide Images

This repository contains the complete pipeline to predict patient outcome (categorical, time to event, or continuous) using whole slide images (WSIs). The method is described in detail in https://www.medrxiv.org/content/10.1101/2021.01.21.21250241v1 .

Requirements

• python ≥ 3.8
• Pytorch ≥ 1.7
• Pandas, Numpy
• sklearn
• tqdm
• large_image (https://github.com/girder/large_image)
• lifelines (https://lifelines.readthedocs.io/en/latest/)

Usage

0. Obtaining the datasets from TCGA

• Download WSIs from the TCGA repository, store in data/WSI_TCGA
• Download the meta file (with clinical and demographical information) from the TCGA repository in json, store in data/meta_files.

1. Create patches from the whole slide images

Example:

python utils/patch_extraction.py --cancer=LGG --num-cpus=8 --magnification=10 --patch-size=224 

2. Extract meta information; setup train/val/test splits

python utils/create_meta_info.py --cancer=LGG --ffpe-only --magnification=10 --stratify=status

If the variable is not available in the meta json file, you will need to manually add this variable to the generated data/meta_clinical_[].csv file. If this variable is a genetic mutation, such as IDH, downloading this variable from TCGA and merge it to the .csv file can be done by running

Rscript utils/obtain_gene_mutations.R

3. Training the deep learning model

Example:

• Classification

python train.py -b=8 --repeats-per-epoch=10 --num-patches=8 --num-val=100 --sample-id --save-interval=10 --outcome-type=classification --outcome=idh --pretrain --lr-backbone=1e-5 --lr-head=1e-5

• Survival

python train.py --stratify=status --sampling-ratio=1,1 -b=8 --repeats-per-epoch=10 --num-patches=8 --num-val=100 --sample-id --save-interval=10 --outcome-type=survival --pretrain --lr-backbone=1e-5 --lr-head=1e-5