SwinCheX

This repository provides a maintained implementation aimed at reproducing the results of the paper "SwinCheX: Multi-label classification on chest X-ray images with transformer".

The codebase was adapted to run efficiently on the Heisenberg HPC cluster at Ilum – School of Science, as part of the activities developed during the Winter Internship at the Brazilian Synchrotron Light Laboratory (LNLS) within the Data Science Group (GCD).

Main Code Adaptations

The original implementation was modified primarily to ensure compatibility with the HPC environment and modern GPU software stacks. The key changes include:

• Adoption of native PyTorch Automatic Mixed Precision (AMP), replacing NVIDIA Apex, which is unavailable under CUDA 12.1
• Upgrade of core dependencies to the most recent versions mutually compatible with each other and with CUDA 12.1

Training on the NIH ChestX-ray14 Dataset

Initial Setup and Pretrained Models

Instructions for environment setup and pretrained model preparation are provided in get_started.md.

In this reproduction, we use the ImageNet-22K pretrained Swin-T model, configured with an input resolution of 224×224, as the backbone for training.

Requirements Installation

All required Python libraries are listed in requirements.txt. To install them, run:

pip install -r requirements.txt

Dataset Preparation

Download the NIH ChestX-ray14 Dataset from Kaggle.

After downloading, merge the images from the provided subfolders into a single directory. Optionally, you may organize separate folders for training, validation, and testing.

Single-GPU Training

To launch training on a single GPU, execute:

python -m torch.distributed.launch --nproc_per_node 1 --master_port 12345 main.py \
  --local_rank 0 \
  --cfg configs/swin_tiny_patch4_window7_224.yaml \
  --resume swin_tiny_patch4_window7_224.pth \
  --trainset path/to/trainset/images \
  --validset path/to/validset/images \
  --testset  path/to/testset/images \
  --train_csv_path configs/NIH/train.csv \
  --valid_csv_path configs/NIH/validation.csv \
  --test_csv_path  configs/NIH/test.csv \
  --batch-size 4 \
  --accumulation-steps 4

Results

The Swin-T-based model was trained for 300 epochs, demonstrating strong generalization performance across both validation and test sets.

Training Runtime

• Average training time per epoch: 23.66 min

Validation Results

The model achieved high discriminative capability during validation:

Metric	Value
Mean AUC	0.75440
Accuracy	95.57%
Loss	0.16972

Test Results

On the held-out test set, the model maintained competitive performance:

Metric	Value
Mean AUC	0.72008
Accuracy	92.45%
Loss	0.23426

Cross-Dataset Evaluation

To assess the generalization capability of the trained model, inference was performed on the validation split of the CheXpert dataset, which was not used during training. The model achieved the following performance metrics:

Metric	Value
Mean AUC	0.72530
Accuracy	81.26%

📌 Notes and Observations

• If all images are stored in a single directory, then the --trainset, --validset, and --testset arguments should all point to the same folder.

• To resume training from an intermediate checkpoint, you must comment out the line marked with "TODO" inside utils.py.