Title: Realistic Noise Synthesis Mitigates Bias and Enhances Tissue Microstructure Estimation in Supervised Machine Learning

Accurate parameter estimation in tissue microstructure probing via diffusion MRI is often compromised by noise-related artifacts. When supervised machine learning models are trained on simulated data, a covariate shift frequently occurs: the noise characteristics of the synthetic signals diverge from those of the actual acquired data, causing the input signal distribution to differ between the training and inference phases. To address this discrepancy, we propose a Realistic Noise Synthesis (RNS) framework and examined its impact on reducing bias in microstructure parameter estimation.

The RNS approach integrates both the Rician expectation and the effective post-processing noise variance into the simulated training signals. Specifically, the Rician expectation is modeled using a noise standard deviation calculated via MPPCA, while the effective standard deviation is extracted from the spherical harmonic residuals of the preprocessed data. We validated this method using the cylinder-zeppelin and SANDI models across various signal-to-noise ratio (SNR) levels on simulated datasets, as well as on in vivo diffusion data featuring repeated acquisitions. Additionally, we assessed the method’s sensitivity to inaccuracies in noise estimation.

Our results indicate that neglecting magnitude-induced noise effects during the training phase leads to systematic, SNR-dependent bias in parameter estimates, a problem that is especially pronounced at low SNR. By incorporating the Rician expectation, we significantly reduced this bias, bringing it down to levels comparable to those achieved by noise-aware nonlinear least-squares fitting. Furthermore, modeling the effective standard deviation yielded additional improvements in precision. While the performance remained largely consistent regardless of the regression architecture employed, it was highly dependent on the accuracy of the noise estimation. These findings underscore that incorporating realistic noise models into simulated training data is crucial for eliminating signal-domain covariate shift and ensuring unbiased supervised microstructure estimation, particularly in low-SNR conditions typical of high b-values or high spatial resolutions.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC