FEAL

This is the official Pytorch implementation of our CVPR 2024 paper "Think Twice Before Selection: Federated Evidential Active Learning for Medical Image Analysis with Domain Shifts".

Requirements

Please review the following requirements and install the packages listed in the requirements.txt

$ pip install --upgrade pip
$ pip install -r requirements.txt

Data Preparation

Datasets

Classification
- Fed-ISIC: Download the skin lesion classification dataset (images and labels) and metadata following Flamby.
- Fed-Camelyon: Download the histology breast cancer classification dataset and metadata following HarmoFL.
Segmentation
- Fed-Polyp: Download the endoscopic polyp segmentation dataset following PraNet.
- Fed-Prostate: Download the prostate MRI segmentation dataset following FedDG.
- Fed-Fundus: Download the retinal fundus segmentation dataset following FedDG.

Data Preprocessing

After downloading the datasets, please naviagte to the FEAL/data/ directory and execute prepare_dataset.py for data preprocessing. The folder structure within Dataset/ should be organized as follows.

├── Dataset
  ├── FedISIC_npy
    ├── ISIC_0012653_downsampled.npy, ISIC_0012654_downsampled.npy, ...
  ├── FedCamelyon
    ├── patches
      ├── patient_004_node_4, patient_009_node_1, ...

  ├── FedPolyp_npy
    ├── client1
      ├── sample1.npy, sample2.npy, ...
    ├── client2
    ├── ...
  ├── FedProstate_npy
    ├── client1
      ├── Case00
        ├── slice_012.npy, slice_013.npy, ...
      ├── ...
    ├── client2
    ├── ...
  ├── FedFundus_npy
    ├── client1
      ├── sample1.npy, sample2.npy, ...
    ├── client2
    ├── ...

Data Split

The data split of Fed-ISIC and Fed-Camelyon follows Flamby and HarmoFL, respectively. For Fed-Polyp, Fed-Prostate, and Fed-Fundus, please navigate to the FEAL/data directory and execute train_test_split.py for the data split process.

Usage

For skin lesion classification using the Fed-ISIC dataset, the command for execution is as follows:

CUDA_VISIBLE_DEVICES=1 python main_cls_al.py --dataset FedISIC --al_method FEAL --query_model both --query_ratio 0 --budget 500 --al_round 5 --max_round 100 --batch_size 32 --base_lr 5e-4 --kl_weight 1e-2 --display_freq 20

Citation

If you find this work helpful for your research, please consider citing:

@inproceedings{chen2024think,
  title={Think Twice Before Selection: Federated Evidential Active Learning for Medical Image Analysis with Domain Shifts},
  author={Chen, Jiayi and Ma, Benteng and Cui, Hengfei and Xia, Yong},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={11439--11449},
  year={2024}
}

Acknowledgment

The codebase is adapted from FedDG, FedLC, and EDL. We sincerely appreciate their insightful work and contributions.

JiayiChen815/FEAL