ISCL: Interdependent Self-Cooperative Learning for Unpaired Image Denoising

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning-based image denoising traditionally requires paired clean-noisy data.
• Existing blind denoising methods often rely on assumptions about noise characteristics, limiting their use in medical imaging.
• Unpaired learning offers a more feasible approach for real-world medical image denoising by relaxing these assumptions.

Purpose of the Study:

• To propose a novel unpaired image denoising scheme, Interdependent Self-Cooperative Learning (ISCL).
• To address the limitations of existing blind denoising methods in the medical domain.
• To develop a method that does not require paired training data or strict assumptions on noise characteristics.

Main Methods:

• Leveraging unpaired learning by combining cyclic adversarial learning with self-supervised residual learning.
• Designing two complementary architectures within ISCL that enhance each other's learning process.
• Utilizing a novel scheme that avoids matching data distributions across domains, unlike other unpaired methods.

Main Results:

• ISCL demonstrated superior performance in denoising various biomedical images, including electron microscopy (EM) and low-dose computed tomography (CT) scans.
• The method effectively handled noise arising from physical device characteristics and structural noise.
• Image quality achieved by ISCL surpassed conventional and current state-of-the-art deep learning-based unpaired denoising techniques.

Conclusions:

• Interdependent Self-Cooperative Learning (ISCL) provides an effective solution for blind image denoising in medical applications.
• The proposed unpaired learning approach significantly improves image quality without requiring paired data or restrictive noise assumptions.
• ISCL shows great potential for enhancing the diagnostic accuracy and utility of medical imaging modalities.