TopoVASC: A Topology-Aware Vascular Segmentation Framework

Zhiwei Tan, Xin Wang, Meng Wang, Zixuan Liu, Yin Guo, Jiamin Xia, Niranjan Balu, Linda Shapiro, Chun Yuan, Mahmud Mossa-Basha

Keywords: Intracranial artery segmentation, Uncertainty-guided refinement;SAM

Abstract:

Accurate segmentation of intracranial arteries in CTA and MRA remains a highly challenging task due to fine-scale vascular morphology, heterogeneous modality-specific contrast, bone-induced artifacts in CTA, and low-signal ambiguities in MRA. These limitations are further exacerbated by local discontinuity regions, which significantly degrade the topological integrity needed for downstream clinical applications.
We introduce TopoVASC, a topology-aware vascular segmentation framework designed to preserve both global vascular connectivity and local structural continuity. TopoVASC employs a SAM-based ViT encoder equipped with modality-aware adapters, including a Frequency Adapter that enhances high-frequency vascular responses and a Vascular Adapter that incorporates multi-scale vesselness priors to emphasize tubular structures. On top of these representations, TopoVASC integrates a topology-aware decoder that jointly predicts vascular segmentation, centerline probability, local discontinuity maps, and voxel-level uncertainty. This multi-head formulation enables the model to explicitly target discontinuous regions while maintaining global topological coherence, resulting in improved branch preservation and reduced false positives in anatomically ambiguous areas.
Experiments on CTA and MRA datasets show that TopoVASC achieves higher accuracy, better structural continuity, and improved centerline preservation compared to several state-of-the-art methods. It consistently outperforms nnUNet, MedSAM, MedSAM2, and interactive SAM-based baselines. Source code will be released soon.

Primary Subject Area: Segmentation

Secondary Subject Area: Application: Radiology

Registration Requirement: Yes

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 118