Non-Cochrane · DTA · PM41562630
Diagnostic Accuracy of Deep Learning Models in Detecting Peri-Implant Marginal Bone Loss: A Systematic Review and Meta-Analysis.
Atieh Momen A, Shah Maanas, Hakam Abeer, et al. · Clinical oral implants research · 2026 · PubMed · doi:10.1111/clr.70090
Findings
Population Adults with dental implant radiographs (periapical or panoramic) evaluated for peri-implant marginal bone loss/peri-implantitis
20 findings
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Diagnostic odds ratio (DOR) | DOR 72 [26, 203] GRADE not reported 5 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled sensitivity Dataset size: <= 1200 images | Sensitivity 0.91 [0.83, 0.99] GRADE not reported 2 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled specificity Dataset size: > 1200 images | Specificity 0.92 [0.88, 0.96] GRADE not reported 3 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled sensitivity Unit of analysis: implant | Sensitivity 0.88 [0.78, 0.98] GRADE not reported 2 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled sensitivity | Sensitivity 0.88 [0.79, 0.93] GRADE not reported 5 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Post-test probability after a positive deep learning result | Post-test probability (%) 71 GRADE not reported 5 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled specificity Unit of analysis: image | Specificity 0.90 [0.83, 0.98] GRADE not reported 3 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Post-test probability after a negative deep learning result | Post-test probability (%) 3 GRADE not reported 5 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled specificity | Specificity 0.91 [0.86, 0.95] GRADE not reported 5 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled specificity Unit of analysis: implant | Specificity 0.91 [0.86, 0.97] GRADE not reported 2 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Positive likelihood ratio (LR+) | LR+ 9.8 [5.7, 16.7] GRADE not reported 5 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled sensitivity Dataset size: > 1200 images | Sensitivity 0.85 [0.75, 0.94] GRADE not reported 3 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled specificity Dataset size: <= 1200 images | Specificity 0.89 [0.82, 0.97] GRADE not reported 2 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Area under the SROC curve (AUC) | AUC 0.95 [0.93, 0.97] GRADE not reported 5 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Negative likelihood ratio (LR-) | LR- 0.14 [0.08, 0.24] GRADE not reported 5 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled sensitivity Imaging modality: panoramic | Sensitivity 0.88 [0.78, 0.98] GRADE not reported 2 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled specificity Imaging modality: panoramic | Specificity 0.91 [0.86, 0.97] GRADE not reported 2 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled sensitivity Imaging modality: periapical | Sensitivity 0.87 [0.78, 0.96] GRADE not reported 3 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled sensitivity Unit of analysis: image | Sensitivity 0.87 [0.78, 0.96] GRADE not reported 3 studies |
| Deep learning models applied to radiographic images | Diagnosis by experienced clinicians interpreting the same radiographic images (reference standard) | Pooled specificity Imaging modality: periapical | Specificity 0.90 [0.83, 0.98] GRADE not reported 3 studies |
Included primary studies 5
Open in Studies| Study | Year | Also included by | Outcomes it informs here | Trial data |
|---|---|---|---|---|
| Chen 2023 Improving Dental Implant Outcomes: CNN-Based System Accurately Measures Degree of Peri-Implantitis Damage on Periapical Film. PMID:37370571 | 2023 | this review only | not stated | not extracted yet |
| Kibcak 2025 DEEP LEARNING-DRIVEN SEGMENTATION OF DENTAL IMPLANTS AND PERI-IMPLANTITIS DETECTION IN ORTHOPANTOMOGRAPHS: A NOVEL DIAGNOSTIC TOOL. PMID:39947781 | 2025 | this review only | not stated | not extracted yet |
| Lee 2024 Establishing a novel deep learning model for detecting peri-implantitis. PMID:38618118 | 2024 | this review only | not stated | not extracted yet |
| Lee 2025 Performance of artificial intelligence-based diagnosis and classification of peri-implantitis compared with periodontal surgeon… PMID:40350773 | 2025 | this review only | not stated | not extracted yet |
| Liu 2022 A pilot study of a deep learning approach to detect marginal bone loss around implants. PMID:35034611 | 2022 | this review only | not stated | not extracted yet |
Review question
| Element | Planned (methods) | Found (results) |
|---|---|---|
| Population | Adults with dental implant radiographs evaluated for peri-implantitis | Radiographs of dental implants with and without marginal bone loss (12,545 radiographs, 5 studies) |
| Intervention | Deep learning models applied to radiographic images to detect marginal bone loss and peri-implantitis | Deep learning models (YOLOv2, YOLOv7, YOLOv8, AlexNet-based CNN, Faster R-CNN) (n=5) |
| Comparison | Diagnosis by experienced clinicians using radiographic images (reference standard) | Experienced clinicians or periodontists as reference standard on the same radiographs (n=5) |
| Outcomes | Sensitivity, specificity, PPV, NPV, F1 score, SROC curve (primary: sensitivity and specificity) | Pooled sensitivity, specificity, LR+, LR-, DOR, AUC, Fagan post-test probabilities |
| Study designs | Retrospective human studies using deep learning on periapical or panoramic radiographs; peer reviewed | 5 retrospective diagnostic accuracy studies (3 periapical, 2 panoramic; 3 image-level, 2 implant-level) |
| Timing | Not specified | Not specified |