Noise Suppression With Similarity-Based Self-Supervised Deep Learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Image denoising is crucial for medical imaging tasks.
• Low-dose and photon-counting CT require effective denoising to balance diagnostic performance and radiation exposure.
• Existing self-supervised methods struggle with correlated noise common in CT scans.

Purpose of the Study:

• To develop a self-supervised deep denoising method capable of handling correlated noise in CT images.
• To address the limitation of paired data requirements in supervised denoising.
• To introduce Noise2Sim, a similarity-based approach for advanced image denoising.

Main Methods:

• Proposed Noise2Sim, a novel similarity-based self-supervised deep denoising approach.
• Implemented a nonlocal and nonlinear processing strategy.
• Validated the method on low-dose and photon-counting CT datasets.

Main Results:

• Noise2Sim effectively suppresses both independent and correlated noises.
• The method achieves performance comparable to or exceeding supervised methods on practical CT datasets.
• Visual, quantitative, and statistical analyses confirm the effectiveness of Noise2Sim.

Conclusions:

• Noise2Sim is the first similarity-based self-supervised deep denoising method for CT.
• It offers a powerful alternative to supervised methods when paired data is unavailable.
• The approach shows significant potential for diverse applications in medical image analysis.