Development and evaluation of a novel automated segmentation approach for mandibular cortical bone and teeth in cone-beam computed tomography using nnU-net.
Authors
Affiliations (7)
Affiliations (7)
- Department of Maxillofacial Orthognathics, Division of Maxillofacial and Neck Reconstruction, Graduate School of Medical and Dental Sciences, Institute of Science Tokyo, 1-5-45 Yushima, Bunkyo-ku, Tokyo 113-8549, Japan.
- Department of Orthodontics, Faculty of Dentistry, Chulalongkorn University, 34 Henri-Dunant Road, Wangmai, Pathumwan, Bangkok 10330, Thailand.
- Institute of Science Tokyo and Chulalongkorn University International Joint Degree Doctor of Philosophy Program in Orthodontics, 1-5-45 Yushima, Bunkyo-ku, Tokyo 113-8549, Japan.
- Department of Mechanical and Systems Engineering, Chukyo University, 101-2 Yagoto Honmachi, Showa-ku, Nagoya 466-8666, Japan.
- Department of Biomedical Informatics, Laboratory for Biomaterials and Bioengineering, Institute of Integrated Research, Institute of Science Tokyo, 2-3-10 Kanda-Surugadai, Chiyoda-ku, Tokyo 101-0062, Japan.
- Oral Science Center, Institute of Science Tokyo, 1-5-45 Yushima, Bunkyo-ku, Tokyo 113-8549, Japan.
- Department of Physiology, Faculty of Dentistry, Chulalongkorn University, 34 Henri-Dunant Road, Wangmai, Pathumwan, Bangkok 10330, Thailand.
Abstract
Characterizing mandibular cortical bone and adjacent tooth roots is essential for safe temporary anchorage device (TAD) placement. Manual cone-beam computed tomography (CBCT) segmentation is time-consuming, and existing automated pipelines output a single mandible label without separating cortical from cancellous bone. This study developed and evaluated an nnU-Net tool that segments mandibular cortical bone and individual lower teeth from CBCT. A total of 113 CBCT scans were split into 90 training and 23 evaluation cases. Manual segmentations of cortical bone and 16 lower teeth were produced by a single annotator after structured calibration with a senior rater on a 23-case subset, and served as a single-observer reference standard. A Residual-Encoder nnU-Net was trained with 5-fold cross-validation and ensemble at inference. Performance was evaluated by precision, recall, Dice similarity coefficient (DSC), 95th-percentile Hausdorff distance (HD95) and Normalized Surface Dice (NSD), and compared with human inter-rater agreement. Calibration achieved a post-washout Dice score of 95.53%. Against the reference standard, the model achieved 98.1% DSC for cortical bone and 95.9% for lower teeth, with HD95 of 0.25 mm and 0.74 mm respectively. Cortical bone NSD was 99.5% at 0.5 mm and 92.9% at 0.3 mm tolerance. Agreement between the model and an independent rater did not differ significantly from that between two calibrated raters. Inference took ∼13 s per scan, or 70 s with the five-model ensemble. Failures concentrated in severe metal artifact and thin cortical bone. The proposed nnU-Net tool produced segmentations of mandibular cortical bone and lower teeth from CBCT in close agreement with a single-observer reference standard. By segmenting cortical bone as an explicit class, it enables visualization of relative cortical thickness at candidate TAD sites; the anatomical accuracy of thickness measurements derived from it has not yet been validated, and because CBCT grayscale is not standardized, such values are relative rather than absolute. These proof-of-concept results require multicenter, multivendor validation and improvements for thin cortex and metal-artifact regions before clinical adoption.