Related Experiment Video
Updated: Jul 24, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
444
A novel one-to-multiple unsupervised domain adaptation framework for abdominal organ segmentation
Xiaowei Xu1, Yinan Chen2, Jianghao Wu3
1SenseTime Research, Shanghai, China; School of Mechanical and Electrical Engineering, University of Electronic Science and Technology of China, Chengdu, Sichuan, China.
Medical Image Analysis
|July 8, 2023
Summary
This study introduces OMUDA, a novel framework for efficient one-to-multiple unsupervised domain-adaptive segmentation of abdominal organs in multi-sequence MRI. OMUDA improves training efficiency while maintaining segmentation accuracy across different imaging domains.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Abdominal multi-organ segmentation in multi-sequence MRI is crucial for pre-operative planning but is hindered by time-consuming manual annotation.
- Existing unsupervised domain adaptation (UDA) methods struggle with domain gaps, anatomical consistency, and one-to-one adaptation, limiting efficiency for multiple target domains.
Purpose of the Study:
- To develop a unified framework (OMUDA) for efficient one-to-multiple unsupervised domain-adaptive segmentation of abdominal organs across various MRI sequences.
- To address limitations of existing UDA methods by improving anatomical consistency and reducing computational cost.
Main Methods:
- Proposed OMUDA framework utilizing disentanglement of content and style for efficient image translation across multiple target domains.
- Incorporated generator refactoring and style constraints to maintain cross-modality structural consistency and minimize domain aliasing.
Main Results:
- OMUDA achieved competitive segmentation performance (e.g., 91.38% DSC on CHAOS dataset) compared to CycleGAN.
- Significantly reduced computational load, with ~87% fewer floating-point calculations during training and ~30% during inference compared to CycleGAN.
Conclusions:
- OMUDA demonstrates effective and efficient one-to-multiple unsupervised domain-adaptive segmentation for abdominal multi-organ MRI.
- The framework's balance of segmentation accuracy and computational efficiency makes it suitable for practical applications like early-stage product development.
