Related Experiment Video
Updated: May 8, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.6K
Source-free domain transfer algorithm with reduced style sensitivity for medical image segmentation
Jian Lin1, Xiaomin Yu2, Zhengxian Wang2
1Sichuan Academy of Medical Science and Sichuan Provincial People's Hospital, Chengdu, China.
Plos One
|December 27, 2024
Summary
This study introduces a new source-free domain transfer algorithm (SFDT-RSS) for medical image segmentation. It significantly improves accuracy by reducing style sensitivity and enhancing generalization without needing source data.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Unsupervised transfer learning for medical image segmentation faces challenges with error propagation due to inaccessible source domain data.
- Existing algorithms struggle to generalize across different medical imaging styles and datasets without direct access to the original training data.
Purpose of the Study:
- To develop a novel source-free domain transfer algorithm (SFDT-RSS) that minimizes style sensitivity and enhances generalization for medical image segmentation.
- To adapt pre-trained models to new target domains without requiring access to the source domain data, addressing a key limitation in current methods.
Main Methods:
- SFDT-RSS employs a generalization strategy for initial source domain model pre-training.
- It utilizes an interpatch style transfer (ISS) strategy with a Transformer architecture for self-training to reduce style sensitivity and improve generalization.
- A model-agnostic adaptive confidence regulation (ACR) loss is used during the domain transfer phase to fine-tune the source model.
Main Results:
- The proposed SFDT-RSS algorithm demonstrated significant improvements in segmentation accuracy across five public datasets for unsupervised cross-domain organ segmentation.
- Specific accuracy improvements ranged from 2.64% to 3.32% compared to existing unsupervised domain transfer algorithms.
- The method effectively enhances generalization capability and reduces reliance on specific image styles.
Conclusions:
- SFDT-RSS offers a robust solution for unsupervised cross-domain medical image segmentation by effectively addressing style sensitivity and data accessibility issues.
- The algorithm's performance highlights the potential of source-free domain transfer methods in medical imaging applications.
- The combination of ISS and ACR loss contributes to improved segmentation accuracy and model generalizability.
