CAMContrast

This is the official code repository for the CiBM 2024 paper "Coarse-to-fine Visual Representation Learning for Medical Images via Class Activation Maps".

Environment setup

python -m virtualenv -p 3.9 env
source env/bin/activate

pip install -r requirements.txt
python setup.py install

Datasets

Dataset	Task	Link
OIA-ODIR	Fundus pretraining	original preprocessed
ChestX-ray14	X-ray pretraining Thorax disease classification	original preprocessed
IDRiD	DR-DME classification Lesions segmentation	source
REFUGE	Glaucoma classification Disc/cup segmentation	source
Vessel-seg	Vessel segmentation	DRIVE STARE CHASE_DB1
SIIM-ACR	Pneumothorax segmentation	source

Pretrained models

• CAMContrast-Fundus
• CAMContrast-X-ray

Scripts

• preprocessing and pretraining: script/upstream
• transfer learning: script/downstream
• shell scripts with commands for reproducing the experimental results: example, e.g., execute example/upstream/chestx-ray14/0_preprocess.sh to run the preprocessing step

Acknowledgement

The implementation of contrastive learning loss was adapted from the SupContrast repository.

Citation

@article{YAP2024108203,
  title = {Coarse-to-fine visual representation learning for medical images via class activation maps},
  journal = {Computers in Biology and Medicine},
  volume = {171},
  pages = {108203},
  year = {2024},
  issn = {0010-4825},
  doi = {https://doi.org/10.1016/j.compbiomed.2024.108203},
  url = {https://www.sciencedirect.com/science/article/pii/S0010482524002877},
  author = {Boon Peng Yap and Beng Koon Ng}
}