Abstract
BACKGROUND: Accurate anatomical labeling of intracranial arteries is critical for cerebrovascular diagnosis and hemodynamic analysis, but remains time-consuming and prone to inter-operator variability. While deep learning provides an automated solution, its clinical adoption is limited by the lack of confidence measures. Incorporating uncertainty quantification into automated labeling could enhance interpretability by identifying ambiguous or abnormal regions and support clinical trust, yet this aspect remains underexplored. METHODS: To address this gap, we introduce an uncertainty-aware deep learning framework for automated artery labeling from 3D Time-of-Flight Magnetic Resonance Angiography (3D ToF-MRA) segmentations (n = 35). Three convolutional neural network architectures were evaluated: (1) UNet with residual encoder blocks, (2) CS-Net, an attention-augmented UNet with spatial attention, and (3) nnUNet, a self-configuring framework that adapts architecture and training to dataset characteristics. Confidence was modeled via test-time augmentation (TTA) combined with a novel coordinate-guided strategy to reduce interpolation errors during inference. Generalizability was assessed by evaluating a subset of the public TubeTK ToF-MRA dataset (n = 20). RESULTS: Voxelwise uncertainty maps highlighted anatomical ambiguities, pathological variations, and inconsistencies in manual references, providing intuitive confidence indicators. nnUNet achieved the highest performance (average Dice score 0.93; clDice 0.94; average surface distance 0.35 mm; 95th percentile of Hausdorff distance 4.51 mm), demonstrating robustness in complex vascular regions. On the TubeTK dataset, nnUNet maintained robust generalization (average Dice score 0.87; clDice 0.87; average surface distance 0.42 mm; 95th percentile of Hausdorff distance 5.85 mm). Validation against co-registered 4D flow MRI showed close agreement between flow velocities derived from automated and manual labels, with no significant differences. CONCLUSION: The proposed framework delivers a scalable, accurate, and uncertainty-aware solution for intracranial artery labeling. By integrating uncertainty quantification, it offers a transparent and clinically trustworthy tool to facilitate cerebrovascular imaging workflows and support subsequent hemodynamic analyses.