CODAMAL: CONTRASTIVE DOMAIN ADAPTATION FOR MALARIA FOR LOW-COST MICROSCOPES

Dataset

You can download the dataset at https://drive.google.com/drive/folders/1k2GuIu6obj3Nz--dOTLuwQnJ2qs1sXxE Additional information on the dataset is available at https://github.com/intelligentMachines-ITU/LowCostMalariaDetection_CVPR_2022

Set path of downloaded folder in data/m5_400x.yaml

Training

The model is trained on HCM images and validated on LCM images

The model is trained and validated on the same zoom resolution (100/400/1000x)

To train on 400x scale:

python train_contrastive.py --data m5_400x.yaml --epochs 60 --weights yolov5m.pt --img 640 --cfg yolov5m.yaml --cache disk  --batch-size 32 --hyp config.yaml

To train without using the contrastive loss:

python train.py --data m5_400x.yaml --epochs 60 --weights yolov5m.pt --img 640 --cfg yolov5m.yaml --cache disk  --batch-size 32 --hyp config.yaml

Change --data m5_400x.yaml to m5_100x.yaml/m5_1000x.yaml to try other resolutions

Results

The output of training will finish with: Results saved to runs\detect\train Navigate to folder to see prediction on both training and validation data

Citation

If you find the repo useful for your research, please consider citing our paper:

@article{dave2024codamal,
  title={CodaMal: Contrastive Domain Adaptation for Malaria Detection in Low-Cost Microscopes},
  author={Dave, Ishan Rajendrakumar and de Blegiers, Tristan and Chen, Chen and Shah, Mubarak},
  journal={arXiv preprint arXiv:2402.10478},
  year={2024}
}

For any questions, welcome to create an issue or contact Ishan Dave ([email protected]).