Abstracts

Rationale: Over the past decades, FCD characterization has been driven by histology that classifies this developmental malformation into discrete histological subtypes. Type-II FCD, the most common form, combines cortical dyslamination together with cell overgrowth and morphological aberrations, including dysmorphic neurons (IIA) and balloon cells (IIB) (Blumcke et al., 2011, Epilepsia). Although histological grading is a well-defined framework to interpret pathology, recent studies have shown co-expression of multiple subtypes within a given FCD as well as heterogeneity of cellular features across FCDs within the same subtype (Iffland and Crino, 2017, Ann Rev Path Mech). Moreover, recent studies have identified components of the MTOR gene that cause FCD via somatic mutations, revealing a genetic continuum not linked to discrete FCD subtypes (Marsan and Baulac, 2018, Neuropath App Neurobio). Here, we combined multimodal MRI and unsupervised machine learning to model mesoscopic structural variability within and across FCD lesions to capture a wider histopathological spectrum beyond current FCD subtypes. Methods: We studied 44 patients with histologically-verified FCD Type-II and 40 healthy controls using 3D T1-weighted and T2-weighted (FLAIR) MRI at 3T. Two experts manually segmented FCD lesions. Using MRI postprocessing, we generated multiple surfaces running through the cortical mantle and underlying white matter (Hong et al., 2017, Neurology). We intersected FCD labels with these surfaces to define lesional vertices and sampled features of FCD pathology (Hong et al., 2017, Neurology), including cortical thickness (to model GM thickening), GM-WM intensity gradient (interface blurring), normalized FLAIR intensity (gliosis) and T1w/FLAIR ratio (abnormal myelin content). Features were smoothed using a 2D quadratic diffusion kernel (2 mm FWHM) and z-normalized with respect to controls. We applied spectral clustering on 10,000 bootstrap-subsets of lesional vertices with number of clusters from K=2 to 5. The clustering results were combined into a consensus affinity matrix storing pairwise probabilities of lesional vertices to belong to the same cluster. Spectral clustering on this matrix identified FCD classes that emerged consistently across bootstraps. To select optimal K, we computed percent agreement of the clustering results across bootstraps to assess stability of fit. Results: Clustering with K=4 provided the highest stability of fit (Fig. 1A). The 4 FCD classes had distinct structural profiles which were independent from IIA/IIB histological grading (Fig. 1B-C): Class-1 characterized by microstructural anomalies in the subcortical WM; Class-2 driven by cortical thickening; Class-3 defined by extensive GM-WM boundary blurring; Class-4 characterized by intracortical gliosis and hypomyelination. Individual analysis revealed structural variability within and across lesions, each co-expressing varying degrees of FCD classes with unique composition (Fig. 1D). Classes were anatomically contiguous and appeared as patches within each lesion regardless of size (Fig. 1E). Conclusions: Our analysis revealed four lesional classes with distinct structural MRI profiles that aggregate to form a given FCD Type-II but can also be found across lesions. Our data-driven framework effectively models the mesoscopic variability within and across FCDs, thereby setting a novel basis for phenotype-genotype association and automated lesion detection that harnesses wider pathological spectrum beyond current histological FCD gradings. Funding: Savoy Epilepsy FoundationCanadian Institutes of Health Research