monai-mil

Source code on GitHub.

README

Multiple Instance Learning (MIL) Examples

This tutorial contains a baseline method of Multiple Instance Learning (MIL) classification from Whole Slide Images (WSI). The dataset is from Prostate cANcer graDe Assessment (PANDA) Challenge - 2020 for cancer grade classification from prostate histology WSIs. The implementation is based on:

Andriy Myronenko, Ziyue Xu, Dong Yang, Holger Roth, Daguang Xu: "Accounting for Dependencies in Deep Learning Based Multiple Instance Learning for Whole Slide Imaging". In MICCAI (2021). arXiv

Requirements

The script is tested with:

• Ubuntu 18.04 | Python 3.6 | CUDA 11.0 | Pytorch 1.10

• the default pipeline requires about 16GB memory per gpu

• it is tested on 4x16gb multi-gpu machine

Dependencies and installation

MONAI

Please install the required dependencies

``bash pip install tifffile pip install imagecodecs `

For more information please check out the installation guide.

Data

Prostate biopsy WSI dataset can be downloaded from Prostate cANcer graDe Assessment (PANDA) Challenge on Kaggle. In this tutorial, we assume it is downloaded in the /PandaChallenge2020 folder

Examples

Check all possible options

`bash python ./panda_mil_train_evaluate_pytorch_gpu.py -h `

Train

Train in multi-gpu mode with AMP using all available gpus, assuming the training images in /PandaChallenge2020/train_images folder, it will use the pre-defined 80/20 data split in datalist_panda_0.json

`bash python -u panda_mil_train_evaluate_pytorch_gpu.py --data_root=/PandaChallenge2020/train_images --amp --distributed --mil_mode=att_trans --batch_size=4 --epochs=50 --logdir=./logs `

If you need to use only specific gpus, simply add the prefix CUDA_VISIBLE_DEVICES=...

`bash CUDA_VISIBLE_DEVICES=0,1,2,3 python -u panda_mil_train_evaluate_pytorch_gpu.py --data_root=/PandaChallenge2020/train_images --amp --distributed --mil_mode=att_trans --batch_size=4 --epochs=50 --logdir=./logs `

Validation

Run inference of the best checkpoint over the validation set

`bash

Validate checkpoint on a single gpu...