Code for RoFormer-MIL

MICCAI - MLMI 2023: "RoFormer for Position Aware Multiple Instance Learning in Whole Slide Image Classification"

Arxiv

Setup

source setup.sh

To install miniconda, torch==2.0 and other packages.

Pre-processing

Data should be stored according to their classes

└- data
    |
    |- Class 1
    |  |- Image1.svs
    |  |- Image2.svs
    |  └ ...
    |
    └- Class 2
       |- Image1.svs
       |- Image2.svs
       └- ...

Pre-processing can be run with

python scripts/preprocessing_pipeline.py
python scripts/new_create_splits.py

• scripts/preprocessing_pipeline.py will run the CLAM preprocessing to tile the slides and extract resnet50 features. Data folder and some parameters can be set in conf/preprocessing.yaml.
• scripts/new_create_splits.py will run stratified train/test splitting. Parameters can be set in conf/create_splits.yaml.

Training

Code relies on

PyTorch Lightning - for boilerplate deep learning code/metrics

Hydra - for configuration files management

xFormers - memory efficient attention

Modeling code

• Model parameters can be set in conf/model_dict.yaml
• Training hyperparameters can be set in conf/training.yaml
• Modeling code is found in romil/models

python scripts/train.py

Will trigger a training run on the K folds, leveraging pytorch-lightning for boilerplate code, and mlflow for experiment tracking (easily configurable in conf/training.yaml:training_args.trainer.logger)

Reference

If you find our work useful in your research please consider citing our paper:

Pochet, E., Maroun, R., Trullo, R. RoFormer for Position Aware Multiple Instance Learning in Whole Slide Image Classification. Machine Learning in Medical Imaging. MLMI 2023.

@InProceedings{pochetroformer23,
author="Pochet, Etienne
and Maroun, Rami
and Trullo, Roger",
title="RoFormer for Position Aware Multiple Instance Learning in Whole Slide Image Classification",
booktitle="Machine Learning in Medical Imaging",
year="2024",
}