Automated implant-level assessment of peri-implant bone loss severity on panoramic radiographs

Peri-implant marginal bone loss (MBL) is a critical radiographic parameter for evaluating implant prognosis; however, its assessment on panoramic radiographs remains time-consuming and prone to observer variability. This study aimed to develop and internally validate a two-stage deep learning framework for implant localization and cross-sectional radiographic classification of implant-level peri-implant bone-loss severity on panoramic radiographs. In this retrospective single-center study, 184 annotated panoramic radiographs, each obtained from a unique patient, were analyzed. In Stage 1, implants were localized using a ResNet-34-based U-Net segmentation model trained on full panoramic images. In Stage 2, implant-centered regions of interest were generated by combining implant bounding boxes with associated bone-loss annotations and applying a 25% contextual margin to preserve the surrounding crestal bone region. The resulting regions were classified as Normal, Mild, or Severe using a DenseNet-169 model. The initial train/validation/test split was performed at the panoramic image level before region extraction, ensuring that all implant-level samples from the same patient remained within the same subset. Stage-2 performance was primarily evaluated using five-fold GroupKFold cross-validation with the panoramic image/patient as the grouping unit. Bootstrap resampling was used to estimate 95% confidence intervals. Calibration analysis, ablation experiments, and Grad-CAM visualization were also performed. Performance was assessed using Dice coefficient, intersection over union, accuracy, balanced accuracy, macro-F1 score, and multiclass ROC-AUC. The Stage-1 segmentation model achieved an implant-class Dice score of 0.9368 and an IoU of 0.8811. In five-fold GroupKFold validation, the Stage-2 classifier achieved an accuracy of 0.848 ± 0.035, balanced accuracy of 0.817 ± 0.050, macro-F1 score of 0.806 ± 0.040, and one-vs-rest ROC-AUC of 0.935 ± 0.021. Bootstrap analysis of pooled out-of-fold predictions yielded an accuracy of 0.849 (95% CI: 0.822–0.874), balanced accuracy of 0.811 (95% CI: 0.780–0.841), macro-F1 score of 0.804 (95% CI: 0.773–0.834), and ROC-AUC of 0.934 (95% CI: 0.920–0.948). The highest class-wise performance was observed for the Normal category, whereas the Mild category showed the greatest classification difficulty. Most misclassifications occurred between adjacent severity categories. The proposed framework showed promising internal validation performance for implant localization and radiographic bone-loss severity classification. However, the model should be interpreted as a preliminary radiographic decision-support tool rather than a standalone diagnostic system. External validation, observer reliability assessment, and integration of clinical parameters are required before clinical implementation.