Gigapixel Whole-Slide Images Classification using Locally Supervised Learning

Pytorch implementation for the locally supervised learning framework described in the paper Gigapixel Whole-Slide Images Classification using Locally Supervised Learning, arxiv and video (MICCAI 2022, accepted for oral presentation).

Installation

Install Anaconda/miniconda.
Required packages:

  $ conda create --name locallearning anaconda=2022.05=py38_0
  $ conda activate locallearning
  $ conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
  $ pip install opencv-contrib-python pytorch-lightning==1.5.10 torchmetrics albumentations tqdm pandas wandb

You can also use docker or singularity. We provide the Dockerfile we used in our experiments. If you do not want to use wandb, you can comment the logger in main.py (line 162 and 167).

Data organization

Our dataset is organized as csv indicated datasets. All the images should be stored in a single directory and the path of this directory (dataset_root) should be passed to main.py. The labels and train/validation/test separation should be listed in a csv file. This csv file should contain 3 columns: slide_id, label and type. slide_id is the filename of input image, including file extension. label is the integer label of a input WSI. The values in the type column should be "train", "valid" or "test". A sample csv file is provided.

Training

We used train.py to train and evaluate our framework.

usage: main.py [-h] [--data-mean DATA_MEAN] [--data-std DATA_STD]
               [--num-workers NUM_WORKERS] [--num-classes NUM_CLASSES]
               [--epochs EPOCHS] [--batch-size BATCH_SIZE]
               [--accumulate-grad-batches ACCUMULATE_GRAD_BATCHES] [--K K]
               [--load-weights LOAD_WEIGHTS] [--precision PRECISION] [--lr LR]
               [--lr-factor LR_FACTOR] [--loss-weight LOSS_WEIGHT]
               [--alpha ALPHA] [--decay-multi-epochs DECAY_MULTI_EPOCHS]
               [--weight-decay WEIGHT_DECAY] --output-dir OUTPUT_DIR
               [--project-name PROJECT_NAME] [--gpu-id GPU_ID]
               [--run-name RUN_NAME] [--progressive]
               dataset_root dataset_csv

Useful arguments:

--output-dir           # The path of directory to store outputs
[--num-classes]        # Number of classes of the input dataset
[--data-mean, --data-std DATA_STD] # the mean and stand devariation of the dataset
[--K]                  # K, the number of divided local modules
[--num_epochs]         # Number of training epochs
[--lr]                 # Initial learning rate
[--lr-factor]          # The multiplication factor on lr of the pretrained networks (eg. 0.5 and --lr 2e-5 means the lr of the pretrained networks is only 1e-5 )
[--loss-weight]        # The weight of each class, used in unbalanced datasets.
[--alpha]              # Hyperparameter alpha
[--decay-multi-epochs] # Epochs to decay lr (e.g. "10,20" means the lr will decay at epch 10 and 20 by a factor of 0.1)

Note that --batch-size should always be 1 as the input images usually have different sizes. We accumulate gradient of several batches by setting --accumulate-grad-batches, it is equivalent to a larger batch size.

The entire training can be done by, for example:

$ python main.py path_to_the_image_directory  path_to_the_csv_file --num-workers 4 --output-dir path_to_output_directory --data-mean 0.6909,0.4654,0.6119 --data-std 0.1786,0.2102,0.1795 --precision 32 --batch-size 1 --epochs 55 --lr 2e-5 --lr-factor 0.5 --loss-weight 0.964218456,1. --accumulate-grad-batches 8 --decay-multi-epochs 25,35,45 --weight-decay 1e-2 --K 4 --alpha 1. --project-name test --run-name test_run --gpu-id 0

Contact

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

Acknowledgments

Part of our code was borrowed from InfoPro-Pytorch. Thanks for their outstanding paper. Our framework used Pytorch Lightning. Thanks for this simple and efficient framework which facilitated our development.

Citation

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

@inproceedings{zhang2022gigapixel,
  title={Gigapixel Whole-Slide Images Classification Using Locally Supervised Learning},
  author={Zhang, Jingwei and Zhang, Xin and Ma, Ke and Gupta, Rajarsi and Saltz, Joel and Vakalopoulou, Maria and Samaras, Dimitris},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={192--201},
  year={2022},
  organization={Springer}
}