The study included skeletal Class III patients treated at the Department of Oral and Maxillofacial Surgery, Faculty of Dentistry, Ondokuz Mayıs University. All patients underwent isolated mandibular setback surgery without concomitant maxillary osteotomy. This approach was intentionally adopted to ensure a homogeneous study population and to eliminate potential confounding effects associated with bimaxillary procedures. All patients underwent presurgical orthodontic treatment, and a surgery-first protocol was not applied. Preoperative CBCT scans were obtained approximately one week prior to surgery, and postoperative CBCT imaging was performed 12–18 months after surgery to allow sufficient time for condylar remodelling to occur. No cases of progressive condylar resorption or clinically evident relapse or TMJ symptoms were observed during the follow-up period.

An end-to-end Python-based workflow was developed to automatically analyse condylar changes after mandibular setback surgery ( Fig. 1 ). Preoperative and postoperative CBCT datasets were imported in DICOM format, anonymised, quality-checked, and converted to a Neuroimaging Informantics Technology Initiative (NıfTI) format for standardised processing.

Schematic overview of the fully automated end-to-end cone-beam computed tomography (CBCT)-based workflow for three-dimensional quantification of condylar changes following mandibular setback surgery.

A total of 350 CBCT scans obtained from the Sivas Cumhuriyet University Faculty of Dentistry were used to train the deep learning segmentation model. Manual annotations were performed by two experienced researchers using 3D Slicer (v5.10.0).

CBCT volumes were resampled to 0.3 mm isotropic resolution and standardised using intensity normalisation and region of interest (ROI)-based cropping. Data were split into training, validation, and test sets in a patient-wise basis.

Segmentation was performed using nnU-Net v2, a self-configuring deep learning framework implemented on a PyTorch platform with CUDA-enabled GPU acceleration. A two-stage segmentation strategy was adopted: initial segmentation of the mandible to establish a stable anatomical reference and ROI, followed by isolation and segmentation of the condylar region from the mandibular mask.

Post-processing included the removal of small disconnected components, hole filling, and surface smoothing using connected-component analysis and morphological operations.

Segmented preoperative and postoperative condylar surfaces were exported as stereolithographic (STL) files and rigidly registered using the iterative closest point (ICP) algorithm, aligning postoperative models to preoperative references with a six-degree-of-freedom transformation. Registration quality was assessed using transformation parameters and the ICP-derived root mean square error (RMSE).

After registration, automatic surface cropping was applied to standardise the region of interest. The inferior 35% and anterior 50% of the condyle were excluded, isolating the superior–posterior condylar head and ensuring consistent spatial coverage across time points.

Point-based surface distance analysis was performed on registered condylar surfaces by computing the distance from each postoperative surface point to the nearest preoperative point. Maximum Hausdorff distance, 95th percentile Hausdorff distance (HD95), surface-based root mean square (RMS) distance, and ICP-derived RMSE were calculated to assess surface correspondence. ^, Volumetric parameters, including preoperative and postoperative volumes, appositional and resorptive volumes, and net volume change, were calculated using a voxel-based approach. In addition to absolute volumetric differences (mm ³), relative volume change (%) was calculated to improve clinical interpretability, using the following formula:

Segmentation overlap was evaluated using the Dice similarity coefficient and Jaccard index. All results were summarised per patient and exported in tabular format, with automated PDF reports generated.

Statistical analyses were performed using JASP software version 0.19.1.0 (JASP Team), with statistical significance set at α = 0.05. Descriptive data are expressed as mean (SD), with all values rounded to two decimal places.

Paired samples t-tests compared left and right condylar changes within patients according to rotation direction. One-way ANOVA assessed the effect of rotation direction on condylar morphometric changes and surface-based metrics. ANCOVA models were used to control for preoperative morphometric variables and sex when appropriate, with Bonferroni-corrected post hoc tests applied for multiple comparisons.

Dice, Jaccard, Hausdorff distance, HD95, and ICP-based RMSE values were compared across rotation groups using one-way ANOVA. Pearson correlation analyses evaluated associations between condylar volume changes, mandibular movement magnitude, and preoperative morphometric measurements, stratified by rotation direction.