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

A research team led by Jine Xie and Zhicheng Zhang has published a novel AI method for denoising diffusion-weighted magnetic resonance images (dMRI), a critical tool in neuroscience and clinical diagnostics. The paper, "Unsupervised Denoising of Diffusion-Weighted Images with Bias and Variance Corrected Noise Modeling," addresses a fundamental limitation: dMRI's inherently low signal-to-noise ratio (SNR), especially under high diffusion weighting, degrades image quality and impairs analysis. Current self-supervised denoising methods often fail to account for the non-Gaussian, Rician noise characteristics of dMRI magnitude data, leading to systematic bias and unreliable metrics. The researchers' breakthrough is an unsupervised framework that corrects this noise model directly within the training process.

The technical innovation lies in two new, mathematically derived loss functions integrated into a Deep Image Prior (DIP) framework. The first loss function, derived from the first-order moment, removes mean bias, while the second, from the second-order moment, corrects squared-signal bias. Both include adaptive weighting to handle variance heterogeneity (heteroscedastic variance). Crucially, this method is image-specific and requires no clean reference data for training. Comprehensive experiments on simulated and in-vivo dMRI data show the proposed losses effectively reduce Rician bias and suppress noise fluctuations, yielding superior image quality and more reliable diffusion metrics than state-of-the-art baselines. This work underscores the importance of precise, bias-aware noise modeling for robust medical image analysis, paving the way for more accurate diagnostics and research in low-SNR imaging scenarios.