Radiomics-Prior SAM for Multi-Region Brain Tumor Substructure Segmentation

Abstract

Different tumor subregions exhibit distinct morphological and radiological signatures, yet many segmentation models treat them uniformly. RP-SAM incorporates radiomics-derived priors that encode intensity, shape, and texture cues associated with edema, non-enhancing core, and enhancing tumor. These priors guide SAM’s mask generation process to favor subregion-consistent boundaries and reduce misclassification between visually similar regions. On BraTS2020 (369 subjects), RP-SAM improves enhancing-tumor Dice from 0.743 to 0.802 (+7.9%), tumor core Dice from 0.786 to 0.841 (+7.0%), and whole-tumor Dice from 0.892 to 0.923 (+3.1%).