V-DDPM: MRI Rician Noise Removal Model Based on VST and DDPM

Abstract

Magnetic resonance imaging (MRI) often contains Rician noise. Unlike additive Gaussian noise, the distribution of Rician noise is related to the data of the image, making it more difficult to remove. It has been shown that the variance stabilizing transformation (VST) can transform the Rician noise distribution into a variance stabilized Gaussian distribution. Utilizing this property, combined with the diffusion model, we proposed an algorithm for the removal of Rician noise from MRI. The algorithm first performs VST on the MRI containing Rician noise and then carries out diffusion denoising. A Gaussian noise sequence is added to the diffusion process; afterward, the diffusion process is reversed to provide different denoising levels through Markov chain modeling. Finally, the denoised image is obtained through the inverse of VST. Experimental results demonstrate the performance of our method in removing Rician noise from magnetic resonance images compared to DDPM denoising alone, while preserving detailed information better. Noise removal was also performed for MRI with a simple structure.