Assessment of Dissemination in Time in Multiple Sclerosis Using a Single 3D T2-FLAIR magnetic resonance imaging Radiomic Signature

Background: In multiple sclerosis (MS), dissemination in time (DIT) is an important component of demonstrating disease activity and establishing the chronology of lesion development. DIT is routinely assessed from contrast-enhanced or follow-up magnetic resonance imaging (MRI). Although proved useful, these methods are not without caveats motivating alternative approaches to study lesion chronology. We investigated whether single T2-FLAIR MRI contains sufficient information to distinguish temporally distinct MS lesion categories using radiomic features. Methods: In a retrospective longitudinal cohort of 273 participants with MS (1,311 MRI acquisitions), persistent brain lesions were temporally stratified using serial T2-FLAIR MRI. Long-standing established (LS-E) lesions were defined as observed for six years at least, and recent incident (R-I), as lesions that appeared over the span of two years in a follow-up MRI and were present until the last scan available. Radiomic features were extracted from lesion-level T2-FLAIR MRI. Feature selection was performed in an independent discovery cohort (Group 1, N=55). Selected features were then used to train a support vector machine classifier in a separate development cohort with repeated group-wise cross-validation (Group 2, N=163). Final performance was assessed in an held-out test cohort (Group 3, N=55). An exploratory analysis was then performed on baseline MRI from 64 subjects with clinically isolated syndrome (CIS), acquired after their first clinical demyelinating event. Results: Longitudinal tracking identified 814 classifiable lesions in Group 1, 3203 in Group 2, and 984 in Group 3. In Group 1, 18 radiomic features met predefined selection criteria (FDR-adjusted p smaller than 0.05 and absolute Cohen's d greater than 0.8), spanning intensity, shape, and texture domains. In the independent test cohort, the support vector machine classifier achieved an area under the curve of 0.78, with correct classification rates of 66% for R-I lesions and 73% for LS-E lesions. In the exploratory CIS analysis, predicted lesion temporal composition showed substantial inter-patient heterogeneity, with approximately one in ten participants presenting radiomic features compatible with LS-E lesions at their first clinical event. Conclusions: Single T2-FLAIR MRI contains measurable information related to MS lesion chronology. This proof-of-concept framework supports the feasibility of a line of research based on radiomics temporal lesion characterisation without contrast administration or dense longitudinal follow-up MRI, with potential relevance for studying lesion evolution and supporting DIT assessment.