Rethinking Multi-center Semi-supervised Breast Cancer Ultrasound Image Segmentation: An Intermediate-domain Perspective

⭐ This code has been completely released ⭐

Paper Link

Overview

Fig. 1. (A) Overview of the CFG framework. IDG bidirectionally combines labeled and unlabeled samples to generate intermediate domain samples. Best viewed in color. (B) Overview of the IDG. Cut denotes dividing the sample into foreground and background based on M_IDG, while Mix involves merging the two parts into an intermediate sample. (C) Overview of the Swin MAE.

Experimental Results

Fig. 3. Visual segmentation results of CFG framework using different labeled ratio samples from the BUSI-WHU domain.

Fig. 4. Visual comparison results from four domains. Green represents the segmentation results and red represents the labels.

Requirements

This repository is based on PyTorch 2.4.0, CUDA 12.6 and Python 3.10.9.

Dataset Construction

The dataset needs to be divided into three folders for training, validating, and testing. The training data should be in the format of the "data/BreastSlice/1-WHU/train/image" folder.

Train

To pre-train the Swin MAE model

python code/networks/Swin_MAE/train.py --batch_size ... --epoch ... --mask_ratio ...

Then, train on the CFG framework

python code/train.py --domain_num ... --lb_domain ...

Test

python test.py --dataset

Visualization

python code/visualization.py

Citation

If you find the code helpful in your research or work, please cite the following paper:

@article{ye2026cfg,
  title={Rethinking Multi-center Semi-supervised Breast Cancer Ultrasound Image Segmentation: An Intermediate-domain Perspective},
  author={Ye, Zhaoyi and Zhang, Yimin and Huang, Jin and Wang, Du and Liu, Sheng and Mei, Liye and Lei, Cheng},
  journal={IEEE Journal of Biomedical and Health Informatics},
  year={2026},
  publisher={IEEE}
}