[NeurIPS 2024 Poster] Learning Low-Rank Feature for Thorax Disease Classification

Rajeev Goel*, Yancheng Wang*, Utkarsh Nath, Alvin C. Silva, Teresa Wu, Yingzhen Yang

*: Equal contribution. Corresponding author: Rajeev Goel

Paper PDF (OpenReview) The paper is now accepted to NeurIPS 2024 as poster.

Installing Requirements

Our codebase follows the MAE Official and uses some additional packages. You may use one of the following commands to build environments with Conda and Pip.

Conda:

conda create -n medical_mae -f medical_mae.yml 

Pip:

conda create -n medical_mae python=3.8
conda activate medical_mae
pip install -r requirements.txt 

Preparing Datasets:

The MIMIC-CXR, CheXpert, and ChestX-ray14 datasets are public available on their official sites. You can download or request the access to them under the agreements.

You may also download them through the following links for research only and follow the official agreements.

MIMIC-CXR (JPG): https://physionet.org/content/mimic-cxr-jpg/2.0.0/

CheXpert (v1.0-small): https://www.kaggle.com/datasets/ashery/chexpert

ChestX-ray14 : https://www.kaggle.com/datasets/nih-chest-xrays/data

Pre-training on ImageNet or Chest X-rays

The pre-training instruction is in PRETRAIN.md.