Thermal noise lowers the accuracy of rotationally invariant harmonics of diffusion MRI data and their robustness to experimental variations

Abstract

Purpose: Rotational invariants (RIs) are at the root of many dMRI applications.Among others, they are presented as a sensible way of reducing the dimension-ality of biophysical models. While thermal noise impact on diffusion metrics hasbeen well studied, little is known on its effect on spherical harmonics-based RI(RISH) features and derived markers. In this work, we evaluate the effect of noiseon RISH features and downstream Standard Model Imaging (SMI) estimates. Theory and Methods: Using simulated and test/retest multishell MRI data,we assess the accuracy and precision of RISH features and SMI parameters inthe presence of thermal noise, as well as its robustness to variations in protocoldesign. We further propose and evaluate correction strategies that bypass theneed of rotational invariant features as an intermediate step. Results: Both RISH features and SMI estimates are impacted by SNR-dependentRician biases. However, higher-order RISH features are susceptible to a sec-ondary noise-related source of bias, which not only depends on SNR, but alsoprotocol and underlying microstructure. Rician bias-correcting techniques areinsufficient to maximize the accuracy of RISH and SMI features, or to ensureconsistency across protocols. SMI estimators that avoid RISH features by fit-ting the model to the directional diffusion MRI data outperform RISH-basedapproaches in accuracy, repeatability, and reproducibility across acquisitionprotocols. Conclusions: RISH features are increasingly used in dMRI analysis, yet theyare prone to various sources of noise that lower their accuracy and reproducibil-ity. Understanding the impact of noise and mitigating such biases is critical tomaximize the validity, repeatability, and reproducibility of dMRI studies.