Background

Advanced diffusion-weighted MRI (DW-MRI) sequences, in combination with biophysical models, provide unprecedented information on the microstructural properties of both healthy and pathological brain tissue.

Nevertheless, it is nowadays challenging to identify the most accurate biophysical model to describe focal microstructural pathology in multiple sclerosis (MS) patients, due to the lack of appropriate comparative studies.

Objectives

To investigate the specificity and sensitivity of 124 independent features derived from 20 diffusion microstructural models to differentiate specific features of tissue alterations in white matter (WM) lesions compared to the surrounding normal-appearing WM (NAWM).

Methods

The study included 102 MS patients: RRMS: 66%, SPMS: 18%, PPMS: 16%, mean age 46±14; female 64%, disease duration 12.16±18.18 years, median Expanded Disability Status Scale (EDSS): 2.5.

DW-MRI data were acquired with 1.8mm isotropic resolution isotropic and with the b-values [0, 700, 1000, 2000, 3000] s/mm².

Lesion masks were generated with a deep learning network algorithm and manually corrected if required. Voxels of NAWM tissue were randomly chosen outside the lesion masks.

The following microstructural models were applied: DTI, Non-parametric DTI, DKI, Ball and Stick, Ball and Sticks, Ball and Rockets, NODDI-Watson, AMICO-NODDI, NODDI-Bingham, SMT-NODDI, NODDIDA, SMT, MCMDI, CHARMED, IVIM, sIVIM, Microstructure Fingerprinting, Microstructure Bayesian, DIAMOND, and DIAMOND isotropic-restricted.

The classification was performed using logistic regression on 300’000 voxels, equally divided in lesion and NAWM voxels. Features were scored according to the Area Under the Curve (AUC), sensitivity, and specificity.

Results

The intra-axonal signal fraction of the Microstructure Bayesian approach scored maximum with AUC=0.87, for threshold=0.5 sensitivity=0.79, sensitivity=0.83. AUC = 0.86 were attributed to the intra-axonal signal fraction of Ball and rockets, NODDI-Watson, AMICO-NODDI, NODDI-Bingham, SMT-NODDI and the extra-axonal perpendicular signal fraction of the Microstructure Bayesian approach. Low AUC scores (<0.75) were achieved by DTI and parameters not related to signal fractions, e.g. orientation dispersion.

Conclusions

Among available microstructural models, the Microstructure Bayesian appeared to best differentiate voxels with microstructural damage in WM lesions compared to NAWM. Very similar, albeit slightly lower accuracy, was achieved by NODDI-based models. In general, models with estimates intra-axonal signal fraction tend to perform better in this type of classification, showing that intra-axonal component may be the dominant factor in distinguishing the two types of tissue. Further analysis will explore the advantage of including combinations of independent features.