AI-Based Automated Segmentation of the Mandibular Cortex and Mental Foramen for Radiomorphometric Measurement on...

Background/Objectives: This study evaluated the feasibility of a YOLOv5-based model for automated segmentation of these structures and automated measurement of mandibular cortical width (MCW) and panoramic mandibular index PMI. Methods: Five hundred panoramic radiographs were retrospectively collected. The mandibular cortex and the right and left mental foramina were annotated with polygon labels as three separate classes and verified by an experienced radiologist; intra- and inter-observer reliability was evaluated on a random 10% subset. A YOLOv5s-seg model was trained for 500 epochs and evaluated on an independent, patient-level test set. Model performance was quantified with precision, recall, F1 score, Dice coefficient, mean intersection over union (mIoU), and mean average precision (mAP@0.5 and mAP@0.5:0.95). Results: The model achieved a precision of 0.94, a recall of 0.82, and an F1 score of 0.88 (mAP@0.5 = 0.86; mAP@0.5:0.95 = 0.52). The Dice coefficient was 0.93 for the mandibular cortex and 0.80 and 0.75 for the right and left mental foramina, respectively. In the side-level paired agreement analysis (62 measurable sides from 44 test-set patients), the ICC was 0.89 (95% CI: 0.84-0.93) for MCW, with a mean bias of +0.08 mm and 0.87 for PMI, with a mean bias of +0.01. Conclusions: YOLOv5 showed high agreement with expert observers for mandibular cortex segmentation and moderate agreement for the mental foramen, supporting the technical feasibility of automated MCW and PMI measurement. Because missed mental foramen detections reduce the number of measurable sites, automated measurement should be combined with a verification step.