Deep learning-driven analysis of osteotomy gap healing following mandibular bilateral sagittal split osteotomy.
Authors
Affiliations (4)
Affiliations (4)
- Charité - Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin, and Humboldt-Universität zu Berlin, Department of Oral and Maxillofacial Surgery, Berlin, Germany.
- Department of Oral and Maxillofacial Radiology, Tokyo Dental College, Tokyo, Japan.
- Charité - Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin, and Humboldt-Universität zu Berlin, Department of Radiology, Berlin, Germany.
- Charité - Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin, and Humboldt-Universität zu Berlin, Department of Oral and Maxillofacial Surgery, Berlin, Germany. Electronic address: [email protected].
Abstract
Bone healing across the osteotomy gap after bilateral sagittal split osteotomy (BSSO) determines whether osteosynthesis material can be safely removed, but is traditionally assessed by subjective visual inspection of cone-beam computed tomography (CBCT) images. This study aimed to develop and validate a deep learning-based model for automated segmentation of BSSO osteotomy gaps on postoperative CBCT scans and to quantify the effect of training dataset size on segmentation performance. Seventy-two osteotomy gaps from 36 patients were manually segmented. Six patients (12 osteotomy sites) were randomly selected and reserved exclusively as an independent evaluation dataset, while the remaining 30 patients (60 osteotomy sites) constituted the full training dataset, from which a reduced training dataset of 15 patients (30 osteotomy sites) was drawn. An nnU-Net framework was trained on both datasets. Performance was assessed using voxel-wise precision, sensitivity, Dice-Sørensen coefficient (DSC), average symmetric surface distance (ASSD), and 95th percentile Hausdorff distance (HD95). The model achieved a mean DSC of 0.73, ASSD of 1.02 mm, and HD95 of 3.52 mm, with no significant difference between manual and automated volumetric measurements. Bland-Altman analysis showed minimal bias (+1.58 mm<sup>3</sup>) and narrow limits of agreement. Automated segmentation thus reproduced the manual reference within two to three voxels of the applied image resolution and without systematic volumetric bias. A reduced 15-case model showed comparable geometric accuracy but larger volumetric bias (+32.75 mm<sup>3</sup>). The proposed nnU-Net-based model enables accurate, reproducible, and fully automated segmentation of osteotomy gaps following BSSO, supporting objective and time-efficient assessment of postoperative bone healing.