Dootmaan / DTFD-MIL.PyTorch

Unofficial implementation of CVPR2022 paper DTFD-MIL. Use the official CAMELYON16 dataset instead of the .pickle file used in the official DTFT-MIL repo.

DTFD-MIL.PyTorch

Update 2023/02/22

We have updated the pretrained weights of ResNet50 to 'https://download.pytorch.org/models/resnet50-11ad3fa6.pth', and it achieved an AUC of 93, which is very close to the result using the official repo of DTFD-MIL with its provided pickle dataset.

Update 2023/01/07

After consulting the authors, patch feature extration method in this repo has now been modified.

Prevoious Version

This repo is built to help readers better understand CVPR2022 paper DTFD-MIL: Double-tier feature distillation multiple Instance learning for histopathology whole slide image classification.

This is NOT an official implementation of the paper. The official one is here

What is the difference?

1. Dataset

   • The official one uses a preprocessed pickle file for CAMELYON16 dataset. This file can be downloaded through Google Drive but is pretty huge.
   • In this repo, we provide the preprocessing code for converting a WSI into patches and then patches into 1024-channel embedding. Therefore, you only need to have the original CAMELYON16 dataset downloaded on your device to train the entire DTFD-MIL framework. However, it should also be noted that we use a different dataset split.

2. Model

   • The official code include three methods for tier-2 distillation, which are MaxMinS, MaxS, and AFS. The experimental result shows that MaxMinS and AFS usually performs well.
   • In this repo, we only have AFS currently.

3. Code-style

   • Compared with the official one, code in this repo is much simpler, which is also easier to understand. We have tested the CAMELYON16 (level 1 patches) on our code but the result shows only an AUC of 86.66% (val) / 82.04% (test).

How to use?

1. Download CAMELYON16 dataset here

2. Clone this repo. Cd into this repo.

3. Run 'Patch-Generation/gen_patch_noLabel_stride_MultiProcessing_multiScales.py' to convert each WSI into a folder of patches. This file is from the official DTFD-MIL repo.

4. Run 'patches2feature.py' to convert each patch into a 1024-dimension vector. This file will run through all the patches in a folder and eventually generate a feats1024.npy for each WSI. In other words, by running this file, we can convert a gigapixel WSI into a Nx1024 matrix (N is the number of patches in a WSI).

5. Run 'train_DTFD-MIL.py' to train and validate the framework.

Some common problems

• loss1.backward() raises an Exception that gradient has changed.

  • This is because you use a newer version of PyTorch. Changing to use PyTorch 1.4.0 can solve this problem.

• Prerequests:

  • Pytorch 1.4.0
  • scikit-learn (or not)
  • numpy
  • opencv