Distance-to-Ventricle Guided Deep Learning for Periventricular and Deep Multiple Sclerosis Lesion Segmentation
Year: 2026
Type: preprint
Abstract: Periventricular lesions are central to the diagnosis of multiple sclerosis (MS), but automatic segmentation is the least dependable. Whether a lesion is periventricular (PV) or deep is said by its three-dimensional distance to the ventricles. We show that on a thick-slice clinical MRI this decision depends on the ventricles in adjacent slices and on the slice spacing, so that an in-plane distance would show 30% of PV lesion voxels as deep. We therefore give the geometry explicitly. A 3D distance-to-ventricle map, computed in millimetres on the full volume, is added in as one of the input channel to a 2.5D U-Net that segments ventricles, normal white matter hyperintensities (WMH), PV lesions and deep lesions. On the 100-patient MS3SEG dataset, against an identically trained baseline, the distance model raised held-out Dice for PV lesions from 0.731 to 0.758 and for deep lesions from 0.702 to 0.738 (Holm-corrected p < 0.001) and reduced PV/deep label confusions by 71%, while the pooled lesion Dice was unchanged. Ablation and counterfactual analyses shows that the prior improves the structural assignment of lesions rather than their detection itself.
Source: Research Square
Institution: Vellore Institute of Technology University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235703424
