Related Experiment Video
Updated: Aug 10, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
ACT: Semi-supervised Domain-adaptive Medical Image Segmentation with Asymmetric Co-Training
Xiaofeng Liu1, Fangxu Xing1, Nadya Shusharina2
1Gordon Center for Medical Imaging, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, 02114.
Medical Image Computing and Computer-Assisted Intervention : MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention
|February 13, 2023
Summary
Semi-supervised domain adaptation (SSDA) improves medical image segmentation by using limited labeled target data. The novel asymmetric co-training (ACT) framework enhances performance significantly, even with few target samples.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Machine Learning
Background:
- Unsupervised domain adaptation (UDA) struggles with significant domain shifts in medical imaging.
- Existing UDA methods show limitations when source and target domains differ greatly.
- Limited labeled target data can yield substantial performance gains in domain adaptation.
Purpose of the Study:
- To develop a semi-supervised domain adaptation (SSDA) framework for medical image segmentation.
- To effectively utilize labeled source data, unlabeled target data, and a small amount of labeled target data.
- To address the underexplored area of SSDA in medical image segmentation.
Main Methods:
- Proposed a novel asymmetric co-training (ACT) framework for SSDA.
- Decoupled SSDA into semi-supervised learning (SSL) and UDA sub-tasks using two segmentors.
- Integrated knowledge from segmentors adaptively via confidence-aware pseudo-labeling and an exponential MixUp decay scheme.
Main Results:
- ACT demonstrated marked improvements over UDA and state-of-the-art SSDA methods.
- Significant performance gains were achieved even with a limited number of labeled target samples.
- The proposed method approached the performance upper bound of supervised joint training.
Conclusions:
- The asymmetric co-training (ACT) framework is effective for SSDA in medical image segmentation.
- ACT successfully leverages diverse data subsets to overcome domain shifts.
- This approach offers a promising direction for improving segmentation accuracy with limited labeled target data.
