Tensor-derived diffusion MRI metrics and NODDI in multiple sclerosis classification

Multi-compartment diffusion models and multi-shell acquisitions are increasingly used to overcome limitations of conventional diffusion tensor imaging, but require longer scans and more complex processing. Meanwhile, diffusion MRI classification studies often rely on a narrow set of tensor-derived metrics, especially fractional anisotropy. We investigated whether broader use of tensor-derived features, including shape descriptors, could improve classification without increasing acquisition complexity.

Multi-shell diffusion MRI (b = 0, 1000 and 2000 s/mm²) was acquired in 220 participants, including 84 healthy controls and 136 patients with relapsing-remitting multiple sclerosis. Thirteen tensor-derived metrics and three neurite orientation dispersion and density imaging (NODDI) metrics were extracted from 98 automatically parcellated brain regions. Classification was performed using L2-regularized logistic regression within a repeated stratified cross-validation framework.

A four-metric set comprising fractional anisotropy, mean diffusivity, spherical and linear anisotropy reached an AUC of 0.967, above fractional anisotropy alone (0.922) and comparable to the full 13-metric representation. The point-estimate gain was concentrated in linear anisotropy; spherical anisotropy did not increase performance. NODDI showed no advantage at matched dimensionality; single-shell yielded similar point estimates, with no significant difference detected. Removing lesion voxels reduced performance only in white matter, where classification remained well above chance. Free-water correction and adjustment for age, sex and intracranial volume did not improve classification.

In this ROI-based diffusion MRI classification, broader use of conventional tensor-derived information showed similar within-cohort discrimination to more complex representations. Linear anisotropy, computed from eigenvalues already available, adds complementary information without requiring additional diffusion contrasts. These findings support exploiting conventional tensor-derived features more comprehensively before adopting more complex diffusion MRI frameworks.