Biophysical compartment models for single-shell diffusion MRI in the human brain: a model fitting comparison

• Published in: Physics in medicine & biology Vol. 67; no. 5; pp. 55009 - 55023
• Main Authors: Davis, Andrew D, Hassel, Stefanie, Arnott, Stephen R, Hall, Geoffrey B, Harris, Jacqueline K, Zamyadi, Mojdeh, Downar, Jonathan, Frey, Benicio N, Lam, Raymond W, Kennedy, Sidney H, Strother, Stephen C
• Format: Journal Article
• Language: English
• Published: England IOP Publishing 07.03.2022
• Subjects: Anisotropy; Bayes Theorem; biophysical modelling; Brain - diagnostic imaging; compartment model; Diffusion Magnetic Resonance Imaging - methods; diffusion tensor imaging; diffusion-weighted imaging; DTI; Humans; Image Processing, Computer-Assisted - methods; magnetic resonance imaging; neuroimaging; White Matter - diagnostic imaging; magnetic resonance imaging; compartment model; diffusion-weighted imaging; DTI; biophysical modelling; diffusion tensor imaging; neuroimaging
• ISSN: 0031-9155; 1361-6560; 1361-6560
• DOI: 10.1088/1361-6560/ac46de

Clinically oriented studies commonly acquire diffusion MRI (dMRI) data with a single non-zero b -value (i.e. single-shell) and diffusion weighting of b = 1000 s mm −2 . To produce microstructural parameter maps, the tensor model is usually used, despite known limitations. Although compartment models have demonstrated improved fits in multi-shell dMRI data, they are rarely used for single-shell parameter maps, where their effectiveness is unclear from the literature. Here, various compartment models combining isotropic balls and symmetric tensors were fitted to single-shell dMRI data to investigate model fitting optimization and extract the most information possible. Full testing was performed in 5 subjects, and 3 subjects with multi-shell data were included for comparison. The results were tested and confirmed in a further 50 subjects. The Markov chain Monte Carlo (MCMC) model fitting technique outperformed non-linear least squares. Using MCMC, the 2-fibre-orientation mono-exponential ball and stick model ( BS ME 2 ) provided artifact-free, stable results, in little processing time. The analogous ball and zeppelin model ( BZ 2 ) also produced stable, low-noise parameter maps, though it required much greater computing resources (50 000 burn-in steps). In single-shell data, the gamma-distributed diffusivity ball and stick model ( BS GD 2 ) underperformed relative to other models, despite being an often-used software default. It produced artifacts in the diffusivity maps even with extremely long processing times. Neither increased diffusion weighting nor a greater number of gradient orientations improved BS GD 2 fits. In white matter (WM), the tensor produced the best fit as measured by Bayesian information criterion. This result contrasts with studies using multi-shell data. However, in crossing fibre regions the tensor confounded geometric effects with fractional anisotropy (FA): the planar/linear WM FA ratio was 49%, while BZ 2 and BS ME 2 retained 76% and 83% of restricted fraction, respectively. As a result, the BZ 2 and BS ME 2 models are strong candidates to optimize information extraction from single-shell dMRI studies.