AI and peri-implantitis: 93% accuracy in detecting bone loss

AI and peri-implant radiology: where does diagnostic accuracy stand?

The interpretation of peri-implant radiographs remains a daily challenge in the dental office, particularly for the early detection of marginal bone loss and the classification of peri-implantitis-related defects. While artificial intelligence (AI) has already widely entered the field of lesion detection, its actual reliability for quantitative measurement, severity grading, and, above all, its concrete clinical applicability, remained less documented until now.

This systematic review and meta-analysis, with searches conducted up to May 25, 2026, evaluates the performance of AI models specifically dedicated to the radiographic assessment of bone level variations and peri-implant defects. The objective was to synthesize data from twelve original studies using various imaging modalities, ranging from periapical and panoramic radiographs to images derived from CBCT.

The study did not test a single biological hypothesis, but sought to objectify the diagnostic capabilities of current architectures (notably Faster R-CNN, YOLO, U-Net, and Mask R-CNN). The analysis focused on measuring the accuracy, sensitivity, and internal specificity of the models. The authors aim here to determine whether these tools can establish themselves as autonomous diagnostic systems or if they should be limited to the role of radiographic decision support in the face of current clinical data variability.

Methodology: how AI screens for bone defects

This systematic review and meta-analysis, with a literature search extending up to May 25, 2026, examined the performance of artificial intelligence (AI) models in the radiographic assessment of marginal peri-implant bone loss. The authors queried several major databases, including PubMed/MEDLINE, Embase, Web of Science, Scopus, Cochrane Library, and IEEE Xplore, supplemented by a search on Google Scholar and citation tracking.

To be included, original studies had to use AI-based methods to assess marginal bone loss, bone level changes, or defects related to peri-implantitis on dental radiographs. In total, 12 studies were selected for the analysis.

• Imaging modalities: Periapical radiographs, panoramic (orthopantomograms), intraoral radiographs, and CBCT-derived imaging.
• AI architectures: Approaches varied according to the clinical objective: Faster R-CNN, YOLO, AlexNet, and CNN classifiers for detection; U-Net and Mask R-CNN for segmentation; modified R-CNN and YOLOv8-pose for measurements; and a multi-stage deep learning cascade workflow for longitudinal quantification.
• Statistical analysis: Random-effects meta-analyses were performed for internal performance metrics when at least three studies were comparable. Five studies allowed for the pooling of accuracy (mean: 0.934; 95% CI: 0.874–0.967), while three studies provided data on sensitivity (0.930) and specificity (0.961).

Due to substantial heterogeneity (I² ranging from 82.9% to 96.4%), the results were interpreted as descriptive syntheses of the models' internal performance rather than as measures of absolute diagnostic accuracy.

Internal performance indicators

This systematic review and meta-analysis, including 12 original studies, highlights promising performance of artificial intelligence models for the radiographic diagnosis of peri-implant bone loss. The marked heterogeneity of the data (I² ranging from 82.9% to 96.4%) nevertheless requires a cautious interpretation of these figures: they reflect internal processing capabilities rather than standardized clinical diagnostic accuracy.

Diversity of architectures and modalities

The analyzed models exploit a wide range of imaging: periapical radiographs, panoramic, intra-oral, and slices from CBCT. Architectures differ according to the complexity of the task:

• Detection and classification: Predominant use of Faster R-CNN, YOLO-based models, AlexNet, and CNN classifiers.
• Segmentation: Use of U-Net or Mask R-CNN.
• Measurement and quantification: Use of modified R-CNN, YOLOv8-pose, or dedicated image processing pipelines.

Only one protocol integrated a multi-stage deep learning cascade workflow for longitudinal quantification. Accuracy, severity classification, and keypoint-based measurements could not be statistically pooled due to the inconsistency of classification schemes and metrics reported across studies.

Clinical limitations

The major weakness lies in the methodology: almost all studies rely on internal validation. No study has provided robust external validation or prospective evaluation in a real clinical workflow. Current models therefore cannot claim the status of an autonomous diagnostic system for peri-implantitis.

Artificial intelligence and peri-implantitis: how much credit should be given to current models?

This meta-analysis, covering 12 studies, examines the capabilities of artificial intelligence (AI) in the radiographic assessment of marginal bone loss and peri-implant defects. With a pooled overall accuracy of 0.934, a sensitivity of 0.930, and a specificity of 0.961, the models show promising performance during internal validations. These figures suggest a real aptitude for automating screening, although significant statistical heterogeneity (I² up to 96.4%) requires major clinical caution.

The weak point of these works lies in the methodology. No model has undergone rigorous external validation or prospective evaluation in real-world workflow conditions. Unlike other areas of dental imaging where AI is beginning to standardize diagnoses, the assessment of peri-implant defects here suffers from a lack of uniform reference standards. These tools can therefore only be considered today as decision-support aids, and not as autonomous diagnostic systems.

Information intended for healthcare professionals. This content may contain errors or truncated summaries. We recommend always verifying with the original source article. Delynov disclaims all liability regarding the use of this information. This document is not intended for patients or the general public.