Exploring Low-Rank Property in Multiple Instance Learning for Whole Slide Image Classification (ICLR 2023) pdf

Authors: Jinxi Xiang, Xiyue Wang, Jun Zhang, Sen Yang, Xiao Han and Wei Yang.

In this study, we investigate the low-rank property of whole slide images to establish a novel multiple-instance learning paradigm. Specifically, we enhance performance through a two-stage process:

1. We introduce a low-rank contrastive learning approach designed to generate pathology-specific visual representations.

2. We augment the standard transformer model by incorporating a learnable low-rank matrix, which serves as a surrogate to facilitate higher-order interactions.

Requirements

use the environment.yaml with conda.

Structure

The folder structure is as follows:

mil/
    └──configs/    # create yaml file that contains dataloader, model, etc.
        ├── config_abmil_camelyon16_imagenet.yaml
        ├── config_clam_camelyon16_imagenet.yaml
        └── ...
    └──models/  # definition of MIL models
        ├── abmil.py
        ├── clam.py
        ├── ilra.py   # our proposed model
        └── ...
    └──splits/  # training ans test data split
        ├── camelyon16_test.csv
        ├── camelyon16_train_10fold.csv
        └── ...
    └──topk/ # dependency of CLAM
        └── ...
    └──train_mil.py/   # main function of training and test
    └──wsi_dataset.py/  # WSI dataset

Step 1

Prepare your WSI feature with CLAM and change the 'Data.feat_dir' in yaml file. To run this example code, you can download the CAMELYON16 features using ImageNet pre-trained Resnet50 from this link.

Step 2

change line 26 in train_mil.py to run different experiments.