Microstructural biophysical models reconstructed from advanced diffusion MRI (dMRI) data provide quantitative measures (qMs), which inform about the brain tissue microenvironment, based on different assumptions.
To compare the sensitivity of available qMs to focal pathology in multiple sclerosis (MS), and to explore which qMs– or combinations of qMs – are best correlated with patients’ disability.
dMRI (1.8 mm isotropic resolution, 149 directions, b-values were 0, 700, 1000, 2000, 3000 s/mm2) was acquired from 67 relapsing-remitting and 33 progressive MS patients (median EDSS: 2.5). The qMs for the isotropic and intra-axonal compartments were derived from the following available models: Ball and Stick, NODDI, SMT-NODDI, MCMDI, NODDIDA, DIAMOND, Microstructure Bayesian approach (MB) and microstructure fingerprinting. In total, 13 qMs were included and subject-wise normalized within brain tissue (nqMs).
To identify the nqMs sensitive to focal pathology, an attention-based convolutional neural network (aCNN) was built to (a) classify randomly sampled WM lesion and perilesional WM patches and (b) generate attention weights (AWs) representing the relative importance of the qMs in the classification. Twenty patients were randomly selected in the test dataset (709 lesion patches and 746 perilesional WM patches), and the rest were in the cross-validation (CV) dataset (2925 lesion patches and 3176 perilesional WM patches). The performance metric was the area under the receiver operating characteristic curve (AUC). Because of the correlation between the nqMs, which may influence the relative AWs, we performed 10-fold CV and selected the nqMS that most contributed to the classification.
To assess which nqMS – or combination of nqMS was best correlated with EDSS, we used Spearman’s correlation coefficient (ρ) with two-sided 20000 permutation tests and followed by Bonferroni correction.
The test AUC was 0.911 indicating the aCNN learned the right AWs to differentiate lesions and perilesional WM. The most discriminating nqMs included isotropic and intra-axonal compartments from MB, the neural density index (NDI) from the NODDI and the intra-axonal compartment from MCMDI.
The sum of isotropic and intra-axonal compartments of the MB (sMB) showed the strongest correlation with EDSS (ρ=-0.40,corr. p<0.0001) followed by the sum of sMB and NDI (ρ=-0.30,corr. p<0.05), and the sum of sMB and intra-axonal compartment from MCMDI (ρ=-0.32,corr. p<0.05). None of the selected nqMs as a single measure and their other combinations correlated with EDSS.
By performing aCNN-aided selection of the openly available WM quantitative measures, we have identified the measures most sensitive to MS focal pathology; furthermore, we have derived a new contrast that – by combining the measures of isotropic and intracellular diffusion – strongly correlated with patients’ disability.