Related Experiment Video
Updated: Mar 3, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.6K
Domain adaptation for low-dose CT denoising via pretraining and self-supervised fine-tuning.
Simiao Yuan1,2, Haipeng Lv3, Zhedian Zhou2
1School of Biomedical Engineering (Suzhou), Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, China.
Journal of X-Ray Science and Technology
|March 2, 2026
Summary
This study introduces a self-supervised method to improve low-dose CT (LDCT) denoising across different datasets. The approach adapts pretrained models without needing new labeled data, enhancing cross-domain generalization for medical imaging.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep learning excels at low-dose CT (LDCT) denoising but struggles with domain shifts.
- Existing domain adaptation methods often require target-domain labeled data, limiting practical use.
Purpose of the Study:
- To develop a self-supervised fine-tuning method for LDCT denoising.
- To address the domain gap in LDCT denoising without requiring target-domain labels.
Main Methods:
- A self-supervised fine-tuning strategy using pixel shuffle preprocessing.
- A two-stage fine-tuning approach to manage pretraining-finetuning input misalignment.
- Utilizing a dual-scale SwinIR model for effective source domain knowledge capture.
Main Results:
- The proposed method successfully bridges the domain gap in LDCT denoising.
- Achieved effective denoising performance and robust cross-domain generalization.
- Demonstrated the efficacy of self-supervised learning without target-domain labels.
Conclusions:
- Self-supervised fine-tuning is a viable approach for cross-domain LDCT denoising.
- The developed method enhances model adaptability and performance on unseen datasets.
- Publicly available code and models facilitate further research and application.