Replace2Self: Self-Supervised Denoising Based on Voxel Replacing and Image Mixing for Diffusion MRI.

Area of Science:

• Medical Imaging
• Neuroimaging
• Biophysics

Background:

• Low signal-to-noise ratio (SNR) is a significant limitation in diffusion-weighted (DW) imaging.
• Noise in DW images complicates the analysis of tissue microstructure.

Purpose of the Study:

• To propose a novel self-supervised learning model, Replace2Self, for effective spatial correlated noise reduction in DW images.
• To enhance the accuracy of tissue microstructure analysis derived from DW imaging.

Main Methods:

• Developed a self-supervised learning model (Replace2Self) utilizing a voxel replacement strategy based on Q-space similar block matching.
• Implemented an image mixing strategy with complementary masks to generate diverse noisy inputs for network training.
• Introduced complementary mask mixing consistency and inverse replacement regularization losses to optimize denoising performance.

Main Results:

• Replace2Self demonstrated superior performance in reducing spatial correlated noise across various noise distributions, levels, and b-values.
• Achieved the highest Peak Signal-to-Noise Ratio (PSNR), exceeding the suboptimal method by at least 1.9% at a 10% noise level.
• Validated effectiveness through extensive simulations, real-world datasets, and ablation experiments, confirming superior generalization ability.

Conclusions:

• Replace2Self effectively reduces spatial correlated noise in DW images, overcoming a key limitation in the modality.
• The proposed method offers improved image quality and robust generalization for microstructure analysis in DW imaging.
• This approach holds promise for advancing quantitative analysis in neuroimaging and other fields utilizing DW imaging.