(MICCAI 2025) MOC: Meta-Optimized Classifier for Few-Shot Whole Slide Image Classification

This repository contains the official PyTorch implementation for our paper:

MOC: Meta-Optimized Classifier for Few-Shot Whole Slide Image Classification (MICCAI 2025)

Preparation

1. Data Preparation

• WSI Preprocessing: This repository does not include scripts for Whole Slide Image (WSI) preprocessing. Please follow the official CLAM pipeline to extract WSI features.

  • Note: Subtypes of the same cancer should be merged into a single folder. For example, place preprocessed luad and lusc features into the same folder named nsclc/merge_features_conch.
  • We primarily use h5_files as they preserve the coordinate information of the patches.
  • The default data storage path can be modified in main_moc.py at lines 205 and 266.

• Example Data Directory Structure:

  data/
  ├── nsclc/
  │   └── merge_features_conch/
  │        ├── h5_files/
  │        └── pt_files/
  └── rcc/
      └── ...

• Dataset Information: Dataset definitions and split information have been placed in the dataset_csv/ and splits/ directories, respectively.

2. Model Preparation

• Download the CONCH model checkpoint from Hugging Face.
• The default path for the checkpoint is models/conch_checkpoint.bin. This path can be changed in main_moc.py at line 135.

Training & Evaluation

Train

Modify the CUDA device ID and dataset name in the scripts/moc_train.sh script, then run:

bash scripts/moc_train.sh

Evaluation

Modify the relevant configurations in the scripts/moc_eval.sh script, then run:

bash scripts/moc_eval.sh

Citation

If you use this code or our method in your research, please cite our paper:

@article{xiang2025moc,
  title={MOC: Meta-Optimized Classifier for Few-Shot Whole Slide Image Classification},
  author={Xiang, Tianqi and Li, Yi and Zhang, Qixiang and Li, Xiaomeng},
  journal={arXiv preprint arXiv:2508.09967},
  year={2025}
}