Skip to content

Latest commit

 

History

History
50 lines (38 loc) · 3.32 KB

README.md

File metadata and controls

50 lines (38 loc) · 3.32 KB
Raw

Precise Location Matching Improves Dense Contrastive Learning in Digital Pathology

Pytorch implementation for the locally supervised learning framework described in the paper Precise Location Matching Improves Dense Contrastive Learning in Digital Pathology, (IPMI 2023, arxiv).

The overview of our framework.

Installation

Please refer to the OpenSelfSup for installation instructions. You can also check our core implementation in openselfsup/datasets/contrastive_bbox.py, openselfsup/models/locmat_cl.py and openselfsup/models/necks_matching.py to integrate it to your own frameworks.

Training

Training scripts can be found under tools folder, e.g. to conduct our PLM_SSL pretraining on 4 GPUs, you can use:

python -W ignore -mtorch.distributed.launch --nproc_per_node=4 --master_port=60123 tools/train.py configs/PLM_SSL/r18_bs256_ep200_lambda_1.0.py --work_dir [output_directory] --seed 0 --launcher pytorch

For downstream segmentation, please install detectron2. You can use segmentation/detection/train_net.py to train and evaluate on downstream dataset, e.g.

python tools/extract_backbone_weights.py  $weights_dir/epoch_200.pth  $weights_dir/segmentation/backbone.pth

cd segmetation/detection

python convert-pretrain-to-detectron2.py $weights_dir/segmentation/backbone.pth $weights_dir/segmentation/backbone.pkl

python train_net.py --config-file configs/GlaS/GlaS_R_18_FPN_1x_moco.yaml --num-gpus 4 --datasets-folder /gpfs/scratch/jingwezhang/data/segmentation MODEL.WEIGHTS $weights_dir/segmentation/backbone.pkl SOLVER.IMS_PER_BATCH 16 OUTPUT_DIR $weights_dir/segmentation/GlaS_R_18_FPN_1x_moco/1

python train_net.py --config-file configs/CRAG/CRAG_R_18_FPN_1x_moco.yaml --num-gpus 4 --datasets-folder /gpfs/scratch/jingwezhang/data/segmentation  MODEL.WEIGHTS $weights_dir/segmentation/backbone.pkl SOLVER.IMS_PER_BATCH 16 OUTPUT_DIR $weights_dir/segmentation/CRAG_R_18_FPN_1x_moco/1

You can also check this repository for more information.

Contact

If you have any questions or concerns, feel free to report issues or directly contact us (Jingwei Zhang [email protected] or Saarthak Kapse [email protected]).

Acknowledgments

Most of our code was borrowed from OpenSelfSup and ConCL. Thanks for their outstanding work.

Citation

If you use the code or results in your research, please use the following BibTeX entry.

@inproceedings{zhang2023precise,
  title={Precise location matching improves dense contrastive learning in digital pathology},
  author={Zhang, Jingwei and Kapse, Saarthak and Ma, Ke and Prasanna, Prateek and Vakalopoulou, Maria and Saltz, Joel and Samaras, Dimitris},
  booktitle={International Conference on Information Processing in Medical Imaging},
  pages={783--794},
  year={2023},
  organization={Springer}
}