Automated Labeling of Intracranial Arteries with Uncertainty Quantification Using Deep Learning

Accurate anatomical labeling of intracranial arteries is essential for cerebrovascular diagnosis and hemodynamic analysis but remains time-consuming and subject to interoperator variability. We present a deep learning-based framework for automated artery labeling from 3D Time-of-Flight Magnetic Resonance Angiography (3D ToF-MRA) segmentations (n=35), incorporating uncertainty quantification to enhance interpretability and reliability. We evaluated three convolutional neural network architectures: (1) a UNet with residual encoder blocks, reflecting commonly used baselines in vascular labeling; (2) CS-Net, an attention-augmented UNet incorporating channel and spatial attention mechanisms for enhanced curvilinear structure recognition; and (3) nnUNet, a self-configuring framework that automates preprocessing, training, and architectural adaptation based on dataset characteristics. Among these, nnUNet achieved the highest labeling performance (average Dice score: 0.922; average surface distance: 0.387 mm), with improved robustness in anatomically complex vessels. To assess predictive confidence, we implemented test-time augmentation (TTA) and introduced a novel coordinate-guided strategy to reduce interpolation errors during augmented inference. The resulting uncertainty maps reliably indicated regions of anatomical ambiguity, pathological variation, or manual labeling inconsistency. We further validated clinical utility by comparing flow velocities derived from automated and manual labels in co-registered 4D Flow MRI datasets, observing close agreement with no statistically significant differences. Our framework offers a scalable, accurate, and uncertainty-aware solution for automated cerebrovascular labeling, supporting downstream hemodynamic analysis and facilitating clinical integration.

💡 Research Summary

This paper presents a comprehensive deep learning framework for the automated anatomical labeling of intracranial arteries from 3D Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) segmentations, with integrated uncertainty quantification to enhance reliability and interpretability.

The study addresses the critical yet labor-intensive task of labeling cerebral vasculature, which is essential for cerebrovascular diagnosis and hemodynamic analysis. The framework was developed and evaluated using a cohort of 35 subjects, comprising 25 patients with intracranial atherosclerotic disease (ICAD) and 10 healthy controls. From the 3D TOF-MRA images, binary vessel segmentations were created, and centerlines were extracted and manually annotated for nine major arterial segments: Basilar Artery, left/right Internal Carotid Arteries, Middle Cerebral Arteries, Anterior Cerebral Arteries, and Posterior Cerebral Arteries. These centerline labels were then propagated to create voxel-wise ground truth masks for training.

A core contribution of the work is the systematic evaluation and comparison of three state-of-the-art convolutional neural network architectures: a residual encoder UNet (serving as a common baseline), CS-Net (which incorporates channel and spatial attention mechanisms for better curvilinear structure recognition), and nnUNet (a self-configuring framework that automates preprocessing, architecture design, and training based on dataset characteristics). The models were trained using a hybrid loss function combining Dice and cross-entropy losses with dynamic weighting and evaluated via 5-fold stratified cross-validation. Results demonstrated that nnUNet achieved the highest performance, with an average Dice score of 0.922 and an average surface distance of 0.387 mm, outperforming the other architectures and showing improved robustness in anatomically complex regions.

To address the black-box nature of deep learning models and build trust for clinical use, the authors implemented an uncertainty quantification method based on Test-Time Augmentation (TTA). During inference, multiple predictions are generated by applying random rotations and translations to the input. The variance across these predictions is then used as a pixel-wise uncertainty measure. A key innovation introduced is a “coordinate-guided” transformation strategy designed to mitigate interpolation errors that typically occur when inverting geometric transformations on discrete label maps. By applying transformations to coordinate grids and mapping back to the original space, this method preserves label integrity and yields more reliable uncertainty estimates. The resulting uncertainty maps were shown to effectively highlight regions of anatomical ambiguity, pathological variation, or inconsistencies in the manual ground truth.

Finally, the clinical utility of the automated labeling pipeline was validated in a downstream hemodynamic analysis task. Using co-registered 4D Flow MRI data from the same subjects, blood flow velocities were calculated within vessels defined by both the automated labels and the manual ground truth labels. Statistical analysis revealed no significant differences between the flow metrics derived from the two methods, confirming that the automated framework produces anatomically accurate labels suitable for subsequent quantitative hemodynamic assessments.

In conclusion, this research provides a scalable, accurate, and uncertainty-aware solution for automated intracranial artery labeling. By combining a high-performing, self-configuring deep learning model with a robust uncertainty quantification technique and demonstrating its validity in a clinical application, the framework represents a significant step towards facilitating the integration of automated cerebrovascular analysis into clinical workflows.