Home

Awesome

DoFE: Domain-oriented Feature Embeddingfor Generalizable Fundus Image Segmentationon Unseen Datasets

by Shujun Wang, Lequan Yu, Kang Li, Yang Xin, Chi-Wing Fu, and Pheng-Ann Heng.

Introduction

This repository is for our TMI2020 paper 'DoFE: Domain-oriented Feature Embeddingfor Generalizable Fundus Image Segmentationon Unseen Datasets'. cellgraph Schematic diagram of our proposedDoFEframework to utilize multi-source domain datasets {D1,...,DK} forgeneralizable segmentation on the unseen dataset D_{K+1}. We adopt the Domain Knowledge PoolMpoolto learn and memorizethe multi-source domain prior knowledge. Importantly, our framework dynamically enriches the image semantic featurehswith domain-oriented aggregated featureˆhaggextracted fromMpoolto improve the expressiveness of the semantic feature.

Requirements

conda create -n DOFE python=3.6.8 

Usage

  1. Clone the repository and download the dataset into your own folder and change --data-dir correspondingly.

  2. Train the model.

    python train.py -g 0 --datasetTrain 1 2 3 --datasetTest 4 --batch-size 16 --resume ./pretrained-weight/test4-epoch40.pth.tar # You need to pretrain a model
    
  3. Test the model.

    python test.py --model-file ./logs/test4/lam0.9/20201120_215812.079473/checkpoint_80.pth.tar --datasetTest 4 -g 0
    
    

Citation

If DoFE is useful for your research, please consider citing:

@article{wang2020dofe,
  title={DoFE: Domain-oriented Feature Embedding for Generalizable Fundus Image Segmentation on Unseen Datasets},
  author={Wang, Shujun and Yu, Lequan and Li, Kang and Yang, Xin and Fu, Chi-Wing and Heng, Pheng-Ann},
  journal={IEEE Transactions on Medical Imaging},
  year={2020},
  publisher={IEEE}
}