A Circle of Willis-Guided Cerebral Artery Segmentation Model for Intracranial Aneurysm Localization on Brain Magnetic Resonance Angiography

Purpose: To develop a cerebral artery (CA) segmentation method for fully automated intracranial aneurysm (IA) localization that integrates Circle of Willis (CoW)-based anatomical priors with AI systems on 3D Time-of-Flight Magnetic Resonance Angiography (TOF-MRA).
Methods and Materials: We collected 470 3D TOF-MRA data and their corresponding ground truth (GT) for IA annotations from Ajou University Medical Center, approved by the institutional review board. Given a TOF-MRA input, we first segment the CoW using a deep neural network, which partitions into five major arteries: internal CA (ICA), basilar artery (BA), anterior CA (ACA), posterior CA (PCA), and middle CA (MCA). The resulting CoW map is then added with a binary vessel map to construct an initial vessel structure, including both labeled CoW regions and unlabeled peripheral arteries. This fused map undergoes two core steps: (i) connected component analysis to isolate unsegmented vascular branches and (ii) K-Nearest Neighbor search to assign each component to the most probable CoW-based artery class. Each candidate component is subsequently repainted via its Euclidean distance to the CoW regions, allowing correction or refinement of misassigned or unsegmented arteries. Finally, we perform IA localization by evaluating the spatial overlap between the GTs and segmented arteries. The algorithm further ranks candidate arteries for each IA based on overlap scores, providing the top-N most likely associated vessels. This allows for interpretable and accurate IA location labeling.
Results: To evaluate localization performance, we compute the confusion matrix and four standard metrics: accuracy, precision, recall, and specificity. For Top-1 localization, the model achieved high average recall and specificity across all artery classes, demonstrating reliable classification. When evaluating Top-2 localization, performance improved substantially, with the model achieving near-perfect scores across all evaluation metrics, highlighting its robustness in identifying the correct CAs within the top two candidates.
Conclusion: We proposed a fully automated CA segmentation model for IA localization. This framework not only enhances the interpretability of aneurysm localization through anatomically meaningful segmentation but also shows strong potential for integration into clinical workflows. It can assist radiologists by automatically prioritizing candidate artery regions for further review, potentially reducing diagnostic time and improving detection consistency.