This project hosts the code for implementing the SGCL algorithm for self-supervised representation learning on whole-slide pathological imaegs.

SCGL is based on our preliminary work

SSLP: Spatial Guided Self-supervised Learning on Pathological Images,
Jiajun Li*, Tiancheng Lin*, Yi Xu
In International Conference on Medical Image Computing and Computer-Assisted Intervention. (MICCAI), 2021

• Novel method: SGCL is a novel self-supervised learning method for whole-slide pathological images to explore the intra-invariance, inter-invariance and inner-invariance beyond the self-invariance of instance discrimination.
• Exploring two intrinsic properties: Spatial-propagated invariance and multi-object prior behind the WSIs are leveraged for designing the pretext tasks.
• Loss function: A new loss function is proposed and validated via a theoretical analysis.
• SOTA performance: The proposed method achieves state-of-the-art performance on multiple pathological datasets and tasks.

Please refer to INSTALL.md for installation and dataset preparation.

We conduct several evalutions on Camelyon16, NCTCRC, Nucls, BCSS.

Note:

• The metrics for VOC det and seg are AP (COCO-style) and mIoU. The results are averaged over 3 trials.
• The training time is measured on 4 3090 Ti GPUs.
• See our paper for more results on different benchmarks.

CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=29504 \
bash tools/dist_train.sh configs/selfsup/sgcl/npidsgclmcrop.py 4 \
--work_dir ./work_dirs/selfsup/npidsgclmcrop

WORK_DIR=work_dirs/selfsup/sgcl/npidsgclmcrop.py/
CHECKPOINT=${WORK_DIR}/epoch_100.pth
WEIGHT_FILE=${WORK_DIR}/extracted_sgcl_camelyon_100ep.pth

python tools/extract_backbone_weights.py ${CHECKPOINT} ${WEIGHT_FILE}

Please refer to README.md for transferring to object detection and semantic segmentation.

• After extracting the backbone weights, the model can be used to replace the original ImageNet pre-trained model as initialization for many prediction tasks.
• If your machine has a slow data loading issue, especially for ImageNet, your are suggested to convert ImageNet to lmdb format through folder2lmdb_imagenet.py or folder2lmdb_coco.py, and use this config_imagenet or config_coco for training.

We would like to thank the OpenSelfSup, LocalAgg and DenseCL for their open-source projects, PyContrast for its detection evaluation configs .

Please consider citing our paper in your publications if the project helps your research.

@article{lin2023sgcl,
  title={SGCL: Spatial guided contrastive learning on whole-slide pathological images},
  author={Lin, Tiancheng and Yu, Zhimiao and Xu, Zengchao and Hu, Hongyu and Xu, Yi and Chen, Chang-Wen},
  journal={Medical Image Analysis},
  pages={102845},
  year={2023},
  publisher={Elsevier}
}