Abstract
BACKGROUND: Marginal Bone Loss (MBL) around dental implants is a key indicator of peri-implantitis. Existing literature evaluates MBL at a single time point, neglecting baseline comparisons and the temporal nature of disease progression. Conventional radiographic assessments remain limited by subjectivity in angulation and magnification. This study aimed to develop and validate a cascade of six Deep Learning (DL) models for automated detection and quantification of MBL on sequential Periapical (PA) radiographs. It further assessed the system’s accuracy against dental specialists and deployed it as a web-based tool. METHODS: A total of 6,453 PAs of dental implants were collected from institutional centers and an open-access dataset. After applying exclusion criteria, 2,960 PAs remained. Sample size calculation was not applicable due to the multi-model architecture; therefore, each model was trained iteratively until convergence was achieved using these PAs. This Artificial Intelligence (AI)-cascade was designed to perform: rotation correction, implant alignment, Region of Interest (ROI) detection, implant classification, mesial-distal detection, and MBL estimation. Each model was assessed using performance metrics appropriate for its specific task. Thereafter, MBL was estimated on forty pairs of PAs by periodontists and AI-cascade; these were compared against a Reference Standard (RS) established by dental specialists. Agreement between AI-derived and RS measurements was evaluated using the Mann–Whitney U test (p < 0.05). RESULTS: Rotation correction model achieved 98.0% accuracy and an F1 score of 99.0%. Implant alignment model demonstrated a Mean Absolute Error of 3.02 degrees. ROI detection model achieved a mean Average Precision (mAP) of 99.4%, with 98.3% precision and 98.2% recall. Implant classification model achieved a mAP of 93.9%, with 100% precision and 98.0% recall. Mesial-distal detection model achieved 78.0% accuracy and an F1 score of 78.0%. Finally, the MBL estimation model attained a mAP of 75.0%, with 73.0% precision and 69.8% recall. No significant differences were found in MBL severity scores between the RS and AI-cascade (p > 0.05). CONCLUSIONS: The AI-cascade demonstrated reliable performance in detecting and quantifying MBL across sequential PAs, comparable to expert evaluation. Its deployment as a web-based application enables objective peri-implant monitoring, representing a significant advancement in clinical translation of AI for implant diagnostics.