AutoCBCT: An Automated CBCT Bone Thickness Analysis Using Deep Learning

The impact of jawbone diseases extends far beyond the mouth. Worldwide, cancer patients completing chemotherapy often have routine dental checkups, with about 70% of patients with breast and prostate cancer and 30–40% of those with lung and other solid tumours being diagnosed with bone metastases. Oncologists treating these patients usually prescribe bisphosphonates to protect their bones, a common and necessary treatment. Yet determining whether the jawbone is starting to deteriorate is something neither the patient nor the dentist can easily detect. Studies show that medication-related osteonecrosis of the jaw (MRONJ) affects a growing share of the millions of patients on antiresorptive therapies across Egypt, the Gulf, and North Africa (up to 15%, compared to just 0.01% in the general population). Nevertheless, by the time it becomes clinically visible, the bone damage is often already irrecoverable. Patients recovering from head and neck radiotherapy, elderly patients with chronic bone loss, and those living with metabolic bone disorders face the same invisible progression. Moreover, all experience the same diagnostic gap, as the primary imaging tool, CBCT, has major drawbacks: its interpretation relies heavily on visual inspection, making conclusions highly subjective, along with the shortage of specialists in many areas, causing patients’ conditions to deteriorate between visits. Globally, more than 12.5 million Cone Beam Computed Tomography (CBCT) scans are performed annually, with total imaging volume increasing by over 50% in the last five years. Additionally, in the Middle East and Africa alone, the CBCT imaging market is projected to grow at a compound annual rate of 12.91% through 2032. This growth creates an expanding diagnostic workload that current practices are unable to meet, highlighting the need for new automated and reliable imaging models. In this work, we propose AutoCBCT, an automated CBCT model that combines Attention U-Net segmentation — which learns to focus on anatomically relevant structures while ignoring noise — with Euclidean Distance Transform-based thickness mapping to produce a spatial heatmap of the entire jaw. Indicators for MRONJ, osteoradionecrosis, fibrous dysplasia, and resorption are based on established clinical criteria. The proposed framework serves as a support tool for clinical decision-making through resorption grading (based on the Cawood and Howell classification) and automated detection of abnormal bone density patterns. The proposed approach is evaluated using 443 CBCT scans from the ToothFairy international dataset, obtained from three commercial scanner platforms with voxel spacings of 0.160–0.300 mm. Bone segmentation achieved a mean Dice coefficient of 0.884 ± 0.020 (range 0.798–0.938), with all 443 cases exceeding the clinical acceptance threshold of 0.7. The thickness estimation of the bone showed a mean absolute error of 0.209 ± 0.094 mm, with 99% of patients below 0.5 mm and every patient below 1.0 mm. The total mean thickness of the bone was 2.797 ±0.410 mm. The clinical data showed that 73.6% of the patients required augmentation or reconstruction. There was no abnormal bone in this dataset.