Anatomy-Aware Segmentation of Cerebral Arteries for Automated Intracranial Aneurysm Localization on 3D Time-of-Flight Magnetic Resonance Angiography

Introduction: Accurate localization of intracranial aneurysms (IAs) on 3D Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) requires precise whole vascular segmentation along with anatomical context to identify the involved arteries. However, the state-of-the-art method, express IntraCranial Arteries Breakdown (eICAB), primarily targets the Circle of Willis (CoW), offering limited coverage of peripheral branches and relying on time-intensive preprocessing procedures.

To address these limitations, we propose a fully automated CoW-oriented Refinement Architecture (CoRA) that integrates CoW-based anatomical priors with a repainting artery module (RAM). Our approach enables comprehensive segmentation of both central and peripheral cerebral arteries (CAs) while assigning meaningful labels to each vascular segment. By ranking arteries via their spatial overlap with aneurysm locations, CoRA provides accurate and interpretable IA localization without preprocessing requirements.

Methods: A total of 470 TOF-MRA scans with annotated IA locations were retrospectively collected from an internal database under IRB approval. The proposed pipeline begins with the segmentation of CoW using a pretrained deep learning (DL) model (i.e., nnUNet), which identifies five major arteries. The segmented CoW is added with a binary vessel map to generate an initial vascular structure that includes both labeled and unlabeled arterial branches. To classify peripheral vessels, we apply the RAM, which consists of (i) connected component analysis to isolate unlabeled vascular segments and (ii) K-Nearest Neighbor classification to associate each segment with one of the CoW-derived artery classes. Each segment is then repainted based on its Euclidean distance from the CoW's center point, enabling correction of misclassifications and incorporation of previously unlabeled vessels. Finally, IA localization is performed by computing spatial overlaps between segmented arteries and annotated IA regions. Candidate arteries are ranked for each IA by calculating their spatial overlap with the lesion, yielding the top-N most likely associated vessels.

Results: We assessed localization performance via a confusion matrix and classification metrics, including accuracy, precision, recall, and specificity. In Top-1 localization, CoRA demonstrated a notable performance gap across all artery types compared to eICAB, indicating reliable identification of the aneurysm-associated artery. Notably, Top-2 evaluation yielded near-perfect metrics, emphasizing the CoRA’s robustness in correctly ranking the parent artery within the top two predictions.

Discussion & Conclusion: We developed a DL-based CA segmentation model for accurate IA localization. Our CoRA not only achieved high performance but also enhanced interpretability through anatomically meaningful vessel labeling. Its modular design supports easy integration into broader neurovascular analysis pipelines. Clinically, it can assist radiologists by identifying likely IA-associated arteries, thereby improving diagnostic consistency and reducing interpretation time.