Artificial Intelligence-Powered Spatial Analysis for Interpretable Bone Loss Assessment in Cone-Beam Computed Tomography

Periodontitis is a chronic inflammatory condition affecting nearly 3.5 billion people globally and is a primary driver of adult tooth loss with significant systemic health complications. While radiographic assessment is central to diagnosis, conventional 2D imaging is limited by anatomical superimposition and high inter-examiner variability. This study presents a hybrid deep-learning and rule-based (HDLRB) framework for automated, interpretable detection and staging of periodontal bone loss using cone-beam computed tomography (CBCT). To mitigate dependence on opaque deep learning models for critical healthcare tasks, our approach combines nnU-Net-based segmentation of dental and alveolar structures with a novel cylinder-based projection mechanism to quantify bone loss based on clinical landmarks: the cemento-enamel junction (CEJ), the alveolar bone surface, and the root length. Evaluated against an expert-adjudicated gold standard of 109 tooth-level sites, the tool achieved a binary detection accuracy of 96.3% and a sensitivity of 98.7%. For periodontal staging according to the 2018 AAP/EFP framework, the system demonstrated excellent agreement (quadratic weighted κ  = 0.879), with 96.3% of sites classified within one stage of expert consensus. Subgroup analysis confirmed robust performance across anatomical regions, with anterior teeth exhibiting higher continuous-measurement precision than posterior teeth. By extracting precise periodontal findings from CBCT volumes acquired for other indications, such as third-molar evaluation, this HDLRB tool provides an opportunistic screening solution that maximises diagnostic performance without additional radiation exposure, offering a transparent and auditable path for integrating AI into clinical periodontal workflows.