To leverage radiomics and machine learning to identify clusters of focal acute MS lesions exhibiting similar spatial and/or textural patterns on conventional cross-sectional (pre- and post-contrast) T1-weighted, T2-weighted, PD-weighted and FLAIR brain MRIs.

Multiple sclerosis (MS) lesions exhibit substantial heterogeneity in terms of underlying pathology, anatomic location, and longitudinal evolution. Conventional MRI is sensitive to some aspects of lesion heterogeneity. Unsupervised clustering approaches may identify subtypes of acute MS lesions with specific pathologic features and help improve disease characterization and prognostication.

Brain MRIs from the ADVANCE trial (1,512 patients with relapsing-remitting MS, NCT00906399) were retrospectively analysed. For each scan, focal acute (gadolinium-enhancing and/or new or substantially enlarging) T2 lesions (N=3058) were identified in the white matter (WM). Each lesion was represented via its center of mass in standard ICBM-152 space and radiomics-based textural properties. Similar lesions were grouped together via k-means clustering, applied separately across the spatial and each sequence-specific textural domain. Cluster count was optimized based on the silhouette score, Davies-Bouldin index and Variance Ratio Criterion.

Optimal clustering quality resulted in 8 prototypical WM sites of lesion emergence, 3 prototypical textural patterns on each of post-contrast T1-, PD-weighted and FLAIR MRI and 5 patterns on each of pre-contrast T1- and T2-weighted MRI. Statistically significant cross-sectional associations were found between the occurrence of these lesion subtypes and patient age, sex, years since disease onset and EDSS score, via logistic regression analysis (p < 0.05).

Specific spatial and textural patterns of acute MS lesions were identified using unsupervised clustering. Further work will evaluate the clinical relevance of identified acute MS lesion patterns towards predicting clinical and MRI endpoints of disease progression. The longitudinal dynamics of each lesion pattern will be investigated via measurements of future lesion repair or transition towards chronic active/inactive lesion states.