Abstract
BACKGROUND: Medication-related osteonecrosis of the jaw (MRONJ) is a serious pathology that can cause bone necrosis in the jaws of patients using antiresorptive or antiangiogenic drugs, often triggered by conditions such as dental infections, periodontal diseases, surgical procedures, or trauma. The purpose of this study was to evaluate the feasibility and segmentation performance of an nnU-Net v2–based deep learning model for the automatic segmentation and voxel-level localization of MRONJ lesions on cone-beam computed tomography (CBCT) images. METHODS: In this study, cone beam computed tomography images of 52 patients with MRONJ were used. All images were manually annotated by experienced oral and maxillofacial radiologists using 3D Slicer software to generate ground truth segmentations. The nnU-Net v2 3D low-resolution architecture, implemented within an automated training pipeline, was trained for 1000 epochs. Model evaluation was performed within the framework of the default five-fold cross-validation strategy of nnU-Net v2. Model performance was evaluated using accuracy, recall (sensitivity), precision, Dice coefficient (DC), Intersection over Union (IoU) and 95% Hausdorff distance (95% HD) metrics. RESULTS: The nnU-Net v2 algorithm generated predictions on the evaluated cases of MRONJ patients. The algorithm achieved average accuracy, recall, and precision scores of 0.999, 0.707, and 0.743, respectively. In addition, the DC, 95% HD, and IoU values were 0.716, 4.045 mm, and 0.569, respectively. CONCLUSIONS: This study demonstrates the feasibility of nnU-Net v2–based automatic MRONJ segmentation on CBCT images. Although limited by dataset size, the results suggest potential clinical utility as a supportive tool rather than a standalone diagnostic system.