Hippocampal Radiomic Signatures in Multiple Sclerosis Subtypes: A Machine Learning-Based MRI Study

This study aimed to investigate whether three-dimensional (3D) hippocampal magnetic resonance imaging (MRI) radiomic features could differentiate multiple sclerosis (MS) subtypes using machine learning models, while establishing a reproducible workflow potentially adaptable to other neuroanatomical and neurodegenerative imaging studies. Brain MRI examinations from 267 patients with MS were included: 99 with relapsing-remitting MS (RRMS), 81 with primary progressive MS (PPMS), and 87 with secondary progressive MS (SPMS). The right and left hippocampi were analyzed in separate hemisphere-specific datasets, each containing 534 hippocampal regions of interest. Hippocampal segmentation was performed from 3D T1-weighted MRI using FreeSurfer/SynthSeg, and radiomic features were extracted using 3D Slicer/SlicerRadiomics. The machine learning algorithms were developed with Orange Data Mining. Random forest achieved the highest area under curve (AUC) values in the right and left hippocampal datasets, with AUCs of 0.790 and 0.786. Gradient boosting demonstrated comparable performance, with AUCs of 0.770 and 0.763 for the right and left hippocampi, respectively, and no significant differences from random forest across the evaluated metrics. Support vector machine showed lower discrimination, with corresponding AUCs of 0.693 and 0.725. Twelve of the 15 highest-ranked radiomic features were common to both hippocampi. T1-weighted MRI-derived three-dimensional hippocampal radiomic features demonstrated moderate internal discrimination among RRMS, PPMS, and SPMS. The integration of automated segmentation with 3D radiomic analysis provides an exploratory imaging-informatics framework that may also be adapted to investigate in other neurological and neurodegenerative disorders, although disease specific and multicenter validation is required.