This is the code of Hybrid Supervision Learning for Pathology Whole Slide Image Classification, Sensetime Research, accepted in MICCAI2021. Jiahui Li, Wen Chen, Xiaodi Huang, Shuang Yang, Zhiqiang Hu, Qi Duan, Dimitris N. Metaxas, Hongsheng Li and Shaoting Zhang
Paper link is here(https://arxiv.org/abs/2107.00934).
Hybrid supervision learning in computational pathology is a difficult task for this reason: High resolution of whole slide images makes it difficult to do end-to-end classification model training. To handle this problem, we propose a hybrid supervision learning framework for this kind of high resolution images with sufficient image-level coarse annotations and a few pixel-level fine labels. This framework, when applied in training patch model, can carefully make use of coarse image-level labels to refine generated pixel-level pseudo labels. Complete strategy is proposed to suppress pixel-level false positives and false negatives.
This code need to run third times to ensemble pixel-level prediction, then run stage2 to
get 0.9243, 4th in Camelyon17 challenge, the submission.csv is my submission file.
(https://camelyon17.grand-challenge.org/evaluation/challenge/leaderboard/)
- I have a GPU cluster, using 'srun python' to submit task. Modify code in orders/ to fit your machine.
- Pytorch, torchvision, opencv, numpy, scikit-learn, scikit-image, PIL, openslide, albumentations. This code did not use strange function thus it could be process by any version of those packages.
- 'pip install' all those packages
- Openslide need to install with both 'apt-get install' and 'pip install'.
- 1, Replace paths of your whole slide images in 'data/wsi_and_anno/train_pos_1_wsi.txt' et al.
- 2, Firstly go to data/, 'python crop_segmentation.py' to generate finegrain pixel labels from xmls.
- 3, 'python orders/Stage1_M_step_init.py' to train initial models.
- 4, 'python orders/Stage1_E_step_init_generate_pseudo_label.py' to generate pseudo labels for all whole slide images.
- 5, 'python oswalk.py' to generate paths indexes, replace keys.
- 6, Repeat step 3~5 from 'init' to 'round3'
- 7, 'python orders/Stage1_E_step_round3_generate_pseudo_label_for_testdata.py' to generate pseudo labels for all whole slide images.
- 8, 'python Stage2_classification_generate_heatmap.py' to generate down sampled heatmaps.
- 9, Ensemble heatmaps from several trails.
- 10, 'python Stage2_classifier.py' to generate submission file from ensembled heatmaps.
- 11, Hyperparameters ( T, V, K, R ) are directly written in dataloader.py and Stage1_E_step_generate_pseudo_label.py.
Contributors names and contact info
Jiahui Li, lijiahui@sensetime.com or jarveelee@gmail.com
- 0.1
- Initial Release
- Multi instance learning, image-level baseline in the paper
- Generalized Framework for Whole-Slide Image Segmentation, pixel-level baseline
- Deep Layer Aggregation, my backbone model
- Streaming convolutional neural networks for end-to-end learning with multi-megapixel images, a better end to end image-level training pipeline with memory compression technique
- Cluster-to-Conquer: A Framework for End-to-End Multi-Instance Learning for Whole Slide Image Classification, a better end to end image-level training pipeline with patch clustering technique
- Data Efficient and Weakly Supervised Computational Pathology on Whole Slide Images, a better end to end image-level training pipeline with patch clustering technique