An interpretable machine learning model to explain the interplay between brain lesions and cortical atrophy in multiple sclerosis.

Multiple Sclerosis (MS) is the most common cause, (after trauma) of neurological disability in young adults in Western countries. While several Magnetic Resonance Imaging (MRI) studies have demonstrated a strong association between the presence of cortical grey matter atrophy and the progression of neurological impairment in MS patients, the neurobiological substrates of cortical atrophy in MS, and in particular its relationship with white matter (WM) and cortical lesions, remain unknown. The aim of this study was to investigate the interplay between cortical atrophy and different types of lesions at Ultra-High Field (UHF) 7 T MRI, including cortical lesions and lesions with a susceptibility rim (a feature which histopathological studies have associated with impaired remyelination and progressive tissue destruction). We combined lesion characterization with a recent machine learning (ML) framework which includes explainability, and we were able to predict cortical atrophy in MS from a handful of lesion-related features extracted from 7 T MR imaging. This highlights not only the importance of UHF MRI for accurately evaluating intracortical and rim lesion load, but also the differential contributions that these types of lesions may bring to determine disease evolution and severity. Also, we found that a small subset of features [WM lesion volume (not considering rim lesions), patient age and WM lesion count (not considering rim lesions), intracortical lesion volume] carried most of the prediction power. Interestingly, an almost opposite pattern emerged when contrasting cortical with WM lesion load: WM lesion load is most important when it is small, whereas cortical lesion load behaves in the opposite way.Clinical Relevance- Our results suggest that disconnection and axonal degeneration due to WM lesions and local cortical demyelination are the main factors determining cortical thinning. These findings further elucidate the complexity of MS pathology across the whole brain and the need for both statistical and mechanistic approaches to understanding the etiopathogenesis of lesions.