Geometric Shape Matching for Explainable and Accurate Medical Image Segmentation: A Post-Processing Refinement Framework

Deep learning models for medical image segmentation, while achieving remarkable performance, often produce anatomically implausible outputs that compromise clinical trust and adoption. We propose a novel inference-time refinement framework that leverages geometric shape matching against a curated library of high-quality organ segmentations to enhance TotalSegmentator predictions without requiring retraining or ground truth data. Our approach provides interpretable corrections by comparing predicted segmentations with anatomically plausible reference templates through a geometry-based matching framework. The framework operates as a modular post-processing layer, addressing TotalSegmentator's occasional anatomical hallucinations while maintaining compatibility with existing clinical workflows. Proof-of-concept experiments on liver segmentation using the CT-ORG dataset demonstrate an average 15\% improvement in Dice scores for poor-performing segmentations. This work presents a promising direction for improving segmentation reliability in clinical deployment while preserving the interpretability required for medical applications.