Unsupervised Denoising of Diffusion-Weighted Images with Bias and Variance Corrected Noise Modeling

Researchers developed new unsupervised denoising methods for diffusion magnetic resonance imaging that correct for Rician noise bias and variance issues. The techniques use bias-corrected training objectives within a Deep Image Prior framework to improve image quality in low signal-to-noise ratio conditions without requiring clean reference data.

• →New denoising method addresses non-Gaussian noise characteristics in diffusion MRI magnitude data that previous methods ignored.
• →Two alternative loss functions are proposed: one correcting mean bias and another correcting squared-signal bias from Rician statistics.
• →The approach works within existing network architectures and doesn't require clean reference images for training.
• →Comprehensive testing on simulated and real MRI data shows superior performance compared to state-of-the-art denoising methods.
• →The technique enables more reliable diffusion metrics and higher image quality in challenging low-SNR medical imaging scenarios.