A Topology-Aware Deep Learning Approach for Automated Multi-Class Segmentation of the Circle of Willis in Modeling Applications

Abstract

Background: Cerebrovascular diseases, which often involve a disruption in blood flow in the Circle of Willis (CoW) and its branching arteries, pose a major global health risk. Computational fluid dynamics (CFD) analyses present an opportunity to study their pathophysiology but require high-quality vessel segmentations. Methods: To generate pseudo-labeled training data, computed tomography angiography (CTA) images were preprocessed and inference was run using two pretrained models: an nnU-Net for multi-class CoW segmentation and a DTUNet for binary cerebral vessel segmentation. These outputs were combined using a region-growing approach; the resulting pseudolabels were used to train an nnU-Net V2 with a topologyaware loss function. CFD analyses were performed on both a model-generated segmentation and a ground truth segmentation derived from a CTA scan which had been expert-labeled by a neuroradiologist. The resulting velocity, pressure and wall shear stress (WSS) profiles for both segmentations were compared across 10 cross-sections of the middle cerebral artery (MCA). Results: After filtering out inaccurate labels, 1,709 of 2,201 pseudo-labeled images were retained for training and testing. Common errors included over-segmentation of small vessels, under-segmentation of large vessels and poor separation of the anterior cerebral arteries when compared to expert-annotated ground truth segmentations. The proposed segmentation model was evaluated on the test set, which used pseudo-labels as a reference standard, and achieved a mean Dice score of 62%, clDice of 40%, IoU of 51%, a HD of 16.9 voxels and an ASD of 3.9 voxels. In terms of centerline-based metrics, the model achieved a mean overlap (OV) of 72% and an average ASCD of 4.26 voxels. In CFD simulations, the predicted segmentation yielded absolute errors of 48.96 �} 30.69 mm/s, 7.47 �} 6.07 Pa and 1.22 �} 0.81 Pa for blood flow velocity, pressure and WSS, respectively, compared to the expert-annotated reference (p < 0.05 for all). Conclusions: This study demonstrates that a deep learning model, trained using pseudo-labels, can successfully generate anatomically plausible multi-class segmentations of the CoW suitable for downstream CFD analysis. However, discrepancies in key hemodynamic metrics compared to expert-annotated data highlight the need for improved pseudo-label accuracy, especially in regions of complex vascular geometry.