Related Experiment Video
Updated: Apr 30, 2026
10:25
Deep Learning-Based Segmentation of Cryo-Electron Tomograms
Published on: November 11, 2022
8.7K
Self-supervised tomographic image noise suppression via residual image prior network.
Jiayi Pan1, Dingyue Chang2, Weiwen Wu1
1School of Biomedical Engineering, Sun Yat-Sen University, Shenzhen, Guangdong, China.
Computers in Biology and Medicine
|July 11, 2024
Summary
This study introduces Residual Image Prior Network (RIP-Net) for computed tomography (CT) denoising. RIP-Net effectively models residuals in similar noisy images, improving CT image quality without requiring paired clean data.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Image Processing
Background:
- Computed tomography (CT) denoising is crucial but challenging, especially in clinical settings where paired noisy-clean images are unavailable.
- Existing self-supervised methods for CT denoising have limitations, including ignoring image residuals and inadequate spectrum information learning.
Purpose of the Study:
- To propose a novel self-supervised CT denoising method, Residual Image Prior Network (RIP-Net), that addresses limitations of current approaches.
- To enhance the accuracy and robustness of CT denoising by modeling low-frequency residual image priors.
Main Methods:
- Developed RIP-Net, a network designed to model residuals between similar noisy image pairs.
- Introduced a novel regularization term for a low-frequency residual image prior.
- Designed a dual-path network to process high and low-frequency image components and incorporated context perception modules.
Main Results:
- Established a mathematical theorem on the non-equivalence of similar-image-based self-supervised learning and supervised learning.
- Demonstrated RIP-Net's superiority over existing unsupervised denoising methods on preclinical photon-counting CT, clinical brain CT, and low-dose CT datasets.
Conclusions:
- RIP-Net effectively models residuals and spectrum information for superior CT denoising.
- The proposed method advances self-supervised learning in medical image denoising, offering a viable alternative to supervised methods in data-scarce clinical environments.
