Accuracy of transfer learning in predicting peri-implantitis onset from intraoral radiograph: A multi-center longitudinal retrospective study.

backgroundThe growing application of artificial intelligence in dentistry has opened new opportunities for predicting peri-implantitis (PI) using convolutional neural networks (CNNs). Transfer learning, where knowledge from a model pre-trained on a large dataset is adapted to a new, domain-specific task, may enhance performance, particularly when medical datasets are limited. This study evaluated the ability of several pre-trained CNN architectures applied to baseline 2D intraoral radiographs to predict peri-implantitis onset.

methodsA retrospective multi-center study was conducted at the University of Michigan and King's College London. A total of 195 implants (146 from Michigan and 49 from London) placed in 108 maintained patients were included. For each implant, a baseline periapical radiograph taken up to one year after prosthetic restoration was analyzed. Five pre-trained CNN architectures (ResNet50, DenseNet, VGG16, InceptionV3, EfficientNet) were combined with four transfer learning strategies: feature extraction (hybrid and non-hybrid), partial fine-tuning (50% of layers), and full fine-tuning. Models were trained using patient-wise splits to avoid leakage, and performance was evaluated on an independent test set using accuracy, area under the ROC curve (AUC), sensitivity, and specificity.

resultsAll CNN architectures showed variable performance across transfer learning strategies. The non-hybrid feature extraction approach provided the most consistent discrimination, with the best models achieving AUC values between 0.74 and 0.83. InceptionV3 reached the highest AUC (0.83). Fine-tuning strategies yielded modest or low performance, with AUCs generally ranging from 0.42 to 0.69 across architectures.

conclusionNon-hybrid feature extraction applied to pre-trained CNNs showed the most promising, though moderate, performance for predicting PI onset from baseline intraoral radiographs. These findings indicate that transfer learning may assist early risk stratification in implant dentistry, but larger prospective studies are needed before clinical adoption. CLINICAL RELEVANCE: Deep learning models may serve as adjunctive tools to identify implants at higher risk for future PI based on routine baseline radiographs, supporting more individualized maintenance protocols and earlier preventive interventions.