Automated detection of marginal bone loss levels in implant brands using deep learning on periapical radiographs

In contemporary dental practice, implants are the standard solution for edentulism. However, the wide variety of implant brands and the prevalence of peri-implantitis present significant diagnostic hurdles for clinicians. This study evaluated an automated hybrid AI framework designed to simultaneously identify implant brands, determine clinical treatment stages, and classify peri-implant bone loss severity using periapical radiographs, aiming to address the efficiency limitations of existing single-function AI models.

A dataset comprising 708 periapical radiographs with 3i and Xive implants was utilized. We employed a YOLOv8 model to localize implants and exclude background noise precisely. Subsequently, a custom implant segmentation algorithm and an automated alveolar crest detection method based on two-stage clustering were applied. EfficientNet-B3 served as the backbone for a multi-task classification of 12 composite classes, integrating implant brand, exposure status, and bone loss status.

The YOLOv8 model demonstrated exceptional performance with 99.39% precision and 98.63% sensitivity. In the complex 12-class classification, the system achieved an overall accuracy of 97.42%, with specific categories such as Xive/Prothesis/Diseased achieving 98.28%. Clinical feasibility tests revealed the framework significantly outperformed manual expert evaluation, drastically reducing average assessment time from 15.5 to 0.16 s while elevating diagnostic accuracy from 90.73% to 97.38%.

The proposed hybrid AI framework successfully consolidates brand identification, staging, and bone loss assessment into a unified, efficient workflow. By offering superior accuracy and speed, it serves as a reliable second opinion to support clinical decision-making and improve diagnostic consistency in dentistry.