Deep Learning-Assisted Three-Dimensional Segmentation of Vertebrobasilar Artery Calcification in Cone Beam Computed Tomography.

VBAC assessment is crucial for stroke risk evaluation, but manual segmentation is time-consuming and subject to inter-observer variability. We developed a novel deep learning model specifically optimized for small vascular structure detection. We designed ImprovedVertebroV5, a 3D U-Net architecture incorporating focal attention mechanisms, small-object detection modules, and pyramid pooling for context aggregation. The model was trained and evaluated on a clinical dataset of CBCT scans from patients undergoing vertebrobasilar evaluation. The V5 model achieved a mean Dice score of 0.7312 ± 0.2080 and an IoU value of 0.6100 ± 0.2180 across 20 test patients. Precision was 0.7785 ± 0.1347, recall was 0.7457 ± 0.2526, specificity was 0.9997 ± 0.0003, and small-object sensitivity was 0.9417 ± 0.1816. Compared with the V4 benchmark, V5 showed statistically significant improvements in Dice score, IoU, recall, F1 score, and specificity. The proposed small object-optimized architecture demonstrated promising internal validation performance for automated VBAC segmentation on CBCT images and may support the future development of opportunistic screening tools in dental imaging.