Three dimensional segmentation of patients with vertebrobasilar artery calcification using cone beam computed tomography

Aim: 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. Material and Methods: 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. Results: Our V5 model achieved a mean Dice score of 0.7312 ± 0.1842 and an IoU value of 0.6230 across 20 test patients; this represents a 31.7% improvement compared to previous models. The detection success rate was determined to be 100% (20/20 patients). The average inference time was 372.2 ms. In addition, precision was 0.8460 ± 0.1347, recall was 0.7890 ± 0.2526, specificity was 0.9997 ± 0.0003, and AUC was 0.8737. Conclusions: The proposed small object-optimized architecture demonstrates state-of-the-art performance for vertebrobasilar artery segmentation, with potential for clinical deployment as a diagnostic aid. Keywords: 3D Segmentation, Artificial Intelligence, Cerebrovascular Event, Cone Beam Computed Tomography, Deep Learning, Vertebrobasilar Artery Calcification