Beyond Overlap: Topology-Aware Evaluation for Biomedical Image Segmentation

Authors:
Publisher website
Google Scholar
Abstract
Continuity preservation is a critical concern in biomedical image segmentation, ensuring that anatomical structures are represented in a clinically meaningful and computationally robust manner. While traditional pixel or voxel-wise evaluation metrics, such as the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), only capture spatial overlap, they fail to capture the structural coherence and topological integrity of segmented anatomies. Such limitations are critical in biomedical contexts, as they can compromise further analyses, lead to misinterpretation of structures, and reduce clinical reliability. To address these limitations, we evaluate continuity and topology-aware metrics with conventional overlap measures. We employ a mandibular canal dataset and retinal vessel dataset for evaluation purposes, allowing the assessment of segmentation performance for both tubular and branching anatomies, and across 2D and 3D modalities. In addition, we propose a measure that enables the identification of the nature of the discontinuities. Experimental results demonstrate that topology-aware evaluation exposes critical segmentation errors overlooked by conventional metrics, offering more comprehensive and clinically relevant insights. This work highlights the necessity of moving beyond overlap-based metrics toward continuity and topology-preserving methods, thereby enhancing the reliability of biomedical imaging workflows in diagnostic, surgical, and computational applications.

This publication has 14 references indexed in Scilit: