Revisiting the MS paradox: automated 2.5D FLAIR segmentation and DTI microstructural profiling reveals dissociation between lesion burden and reference-like intralesional tract profiles.
Authors
Affiliations (4)
Affiliations (4)
- Departamento de Bioingeniería Translacional, Universidad de Guadalajara, Jalisco, México.
- Dirección de Enseñanza e Investigación en Salud, Unidad Médica de Alta Especialidad Hospital de Especialidades, Centro Médico Nacional de Occidente, Instituto Mexicano del Seguro Social, Guadalajara, México.
- Banco Central de Sangre, Unidad Médica de Alta Especialidad Hospital de Especialidades, Centro Médico Nacional de Occidente, Instituto Mexicano del Seguro Social, Guadalajara, México.
- Departamento de Farmacobiología, Centro Universitario de Ciencias de la Salud, Universidad de Guadalajara, Guadalajara, México.
Abstract
Magnetic resonance imaging (MRI) reveals macroscopic lesion burden in multiple sclerosis (MS) but correlates poorly with clinical phenotype, a phenomenon known as the clinico-radiological paradox. While diffusion tensor imaging (DTI) metrics assess microstructural damage, standard whole-lesion averaging may obscure microstructurally spared white matter within heterogeneous plaques. We propose a fully automated, multimodal pipeline to identify and quantify reference-like intralesional segments (RLIS) within MS lesions (i.e., the cluster exhibiting maximum feature-space similarity to subject-specific, lesion-spared reference tissue). A modified 2.5D U-Net, developed and evaluated on a cohort of 75 MS patients (115 series), performed automated FLAIR lesion segmentation. In an independent cohort of 100 MS subjects, intralesional DTI metrics (FA and MD) were extracted along streamlines reconstructed via Constrained Spherical Deconvolution tractography. Unsupervised clustering algorithms, validated against a subject-specific, lesion-spared reference CST streamlines (internal capsule), identified operationally defined reference-like segments. The 2.5D U-Net achieved robust segmentation performance (DSC: 0.732, VC: 0.954), exhibited higher nominal metrics in a contextual benchmark comparison (DSC: 0.682, VC: 0.864). Across the clinical cohort, lesion count (11-131) and volume (0.54-57.3 mL) varied widely. Among the evaluated clustering methods, K-Means and Gaussian Mixture Models proved to be the most conservative, identifying median reference-like segment proportions of 26% and 22%, respectively. Crucially, the proportion of reference-like intralesional segments showed no significant monotonic association with macroscopic lesion volume or count. This automated framework successfully integrates robust deep learning-based lesion segmentation with tract-specific analysis to quantify the proportion of reference-like intralesional segments. The observed dissociation between overall lesion burden and RLIS suggests that this index serves as a promising candidate metric for exploring microstructural features that may contribute to the clinico-radiological paradox.