Clinically applicable deep learning model for segmentation of the mandibular bone and inferior alveolar canal in CBCT cross-sectional images

BACKGROUND AND OBJECTIVES: With the increased reliance on using CBCT scans for dental implant procedures, the need for accurate and efficient segmentation of mandibular anatomical structures has intensified, posing a significant burden on dentomaxillofacial radiologists. This study addresses the clinical applicability of automated deep learning-based segmentation by comparing three advanced models and proposing a solution. MATERIALS AND METHODS: This study evaluated the performance of three state-of-the-art segmentation models (YOLOv8-seg, nnUNet, and SwinUNETR) on cross-sectional CBCT images for segmenting alveolar bone and inferior alveolar canal. YOLOv8-seg, a single-stage CNN detector with segmentation capacities, was trained on a curated dataset and benchmarked against the other models using standard metrics. RESULTS: The YOLOv8-seg model achieved superior segmentation accuracy, with a DSC of 0.962, an IoU of 0.929, and a mean average precision (mAP50) of 0.952. Its inference time (0.00586 sec/image) makes it over 100 times more efficient than the conventional models. Despite some false-positive canal segmentations in the anterior regions, YOLOv8-seg demonstrated strong generalization and clinical promise. CONCLUSION: With further validation and dataset refinement, YOLOv8-seg demonstrates potential as a clinically applicable tool for CBCT image segmentation, offering high-accuracy parameters and significant computational efficiency. Its integration into real-world dental implant planning workflows may reduce clinician workload and improve consistency in decision-making.