Anatomy-informed multi-task framework for visceral artery aneurysm screening on non-contrast CT.

Abstract

Visceral artery aneurysm (VAA) screening on non-contrast CT (NCCT) is fundamentally a structure-constrained representation learning problem under weak vascular contrast. In NCCT, visceral arteries lack explicit lumen enhancement, and aneurysms often appear only as subtle local dilatations embedded in thin and branching vascular structures. This makes conventional intensity-driven segmentation unreliable and demands models that can preserve vascular continuity, capture subtle lesion morphology, and leverage anatomical context at the same time. To address this problem, we propose an anatomy-informed multi-task framework for NCCT-based VAA screening. Built on a Mamba-based segmentation backbone, our method reshapes token propagation according to vascular structure, enabling anatomy-consistent long-range dependency modeling without changing the original state space dynamics. To further strengthen the learned representation, we incorporate hierarchical anatomical supervision through global vessel-location classification and local patch-wise lesion prediction, so that the embedding is guided by both coarse anatomical semantics and fine-grained lesion evidence. We further introduce a prototype-based inference strategy that uses feature-space similarity to suppress morphologically confusing false positives. Extensive experiments on internal and external cohorts show that the proposed framework improves lesion segmentation, reduces false positives, and generalizes robustly across multiple cohorts. These results demonstrate that geometry-aware and anatomy-informed representation learning provides an effective and practical solution for VAA opportunistic screening on NCCT.