Centerline Boundary Dice Loss for Vascular Segmentation

TL;DR

This work tackles vascular segmentation by jointly preserving topology and capturing geometric detail across vessels of varying diameters. It introduces cbDice, a boundary-aware, radius-informed extension of centerline-based losses, and a unified family of clDice variants (cl-X-Dice) that incorporate diameter information through skeleton radii and distance maps. The approach is validated on three diverse 2D/3D datasets (DRIVE, Parse Challenge, TopCoW), showing superior Dice and topology metrics over standard Dice, clDice, and B-DoU losses, with particularly strong results on TopCoW; theoretical analysis in the supplementary material clarifies the geometric sensitivities of the different variants. The authors provide publicly available code and offer guidance on hyperparameter choices to balance topology and geometry across tasks, signaling practical impact for clinical vascular segmentation pipelines.

Abstract

Vascular segmentation in medical imaging plays a crucial role in analysing morphological and functional assessments. Traditional methods, like the centerline Dice (clDice) loss, ensure topology preservation but falter in capturing geometric details, especially under translation and deformation. The combination of clDice with traditional Dice loss can lead to diameter imbalance, favoring larger vessels. Addressing these challenges, we introduce the centerline boundary Dice (cbDice) loss function, which harmonizes topological integrity and geometric nuances, ensuring consistent segmentation across various vessel sizes. cbDice enriches the clDice approach by including boundary-aware aspects, thereby improving geometric detail recognition. It matches the performance of the boundary difference over union (B-DoU) loss through a mask-distance-based approach, enhancing traslation sensitivity. Crucially, cbDice incorporates radius information from vascular skeletons, enabling uniform adaptation to vascular diameter changes and maintaining balance in branch growth and fracture impacts. Furthermore, we conducted a theoretical analysis of clDice variants (cl-X-Dice). We validated cbDice's efficacy on three diverse vascular segmentation datasets, encompassing both 2D and 3D, and binary and multi-class segmentation. Particularly, the method integrated with cbDice demonstrated outstanding performance on the MICCAI 2023 TopCoW Challenge dataset. Our code is made publicly available at: https://github.com/PengchengShi1220/cbDice.

Figures (4)

• Figure 1: Challenges in vascular segmentation across diverse datasets. Key features are highlighted: (1) green centerlines for vessel connectivity; (2) blue circles for morphological characteristics; and (3) light orange arrows within light blue frames indicating branches with significant diameter differences.
• Figure 2: This figure displays a segmented 2D retinal vessel, offering a visual example of the different variables mentioned in Table \ref{['tab:comprehensive_metrics_comparison']}. A 2D schematic depicts the transition from clDice to cbDice, visualized as a bottle with varying radius.
• Figure 3: (a) With translation-only perturbations, cb-Dice metric sensitivity to cl-M-Dice variations, increasing alongside translation distance, is comparable to B-DoU, while clDice remains near 1. (b) In uniform scaling (enlargement or reduction), cbDice-Dice pairing ensures more consistent evaluations than clDice-Dice, effectively adapting to scale changes. (c) For diameter imbalances, cbDice-Dice consistently assesses varied diameter branches, outperforming clDice-Dice.
• Figure 4: Comparative visualization of results on the TopCoW 2023 dataset. Yellow arrows mark areas of segmentation false negatives, green arrows point to false positives, and red arrows identify areas of misclassification.

Theorems & Definitions (6)

• theorem thmcountertheorem
• proof
• theorem thmcountertheorem
• proof
• theorem thmcountertheorem
• proof