MIMMIC-CXR

This repository contains a pipeline for multi-label classification of thoracic pathologies using the MIMIC-CXR dataset, for URIS progress report.

---

Project structure

├── checkpoints/          # Saved model weights (.pth)
├── data/                 # Dataset directory 
├── config.py             # Hyperparameters
├── data_loader.py        # Data loader
├── model.py              # DenseNet
├── train.py              # Main training script 
├── evaluate.py           # Per-class performance eval
├── visualize.py          # Script to generate loss/AUROC plots
├── submit_job.sh         # PBS script for training
├── evaluate.sh           # PBS script for evaluation
├── utils.py              # Helpers
├── create_master.py      # Merges raw metadata into one
└── split_files.py        # Generates `mimic_train.csv` and `mimic_val.csv`

Getting started

1. Environment Setup

conda create -n icv python=3.7
conda activate icv
pip install torch torchvision tqdm pandas scikit-learn matplotlib

2. Training on the cluster

Prepare data:

python create_master.py
python split_files.py

To submit a training job:

qsub submit_job.sh

3. Evaluation

Once training is complete, run:

qsub evaluate.sh

This generates class_performance.csv, containing the AUROC score for each of the 14 pathologies.