Classification of periapical radiographic findings for root canal therapy decision support using deep neural networks.

Abstract

Objectives: Recent advancements in deep learning and its applications in oral and dental health have garnered considerable interest. A crucial component of the diagnostic process in everyday clinical practice involves the examination of dental radiographs which artificial intelligence can be a suitable auxiliary tool.

Methods: This study develops a deep learning-based diagnostic support system for classifying periapical radiographic patterns, including findings associated with potential need for endodontic evaluation. We emphasize that radiographic assessment is only one component of comprehensive endodontic diagnosis, which must include clinical examination, patient history, and pulp vitality testing. This study compares seven cutting-edge pre-trained deep learning models (YOLO 11, ResNet-18, VGG-19, AlexNet, EfficientNet, MobileNet V2, LeNet-5) alongside a custom convolutional neural network for classifying periapical radiographs into three categories: Healthy, Decay (caries approaching pulp), and Periapical pathology (RCT-indicated) (periapical pathology suggesting need for root canal treatment).

Results: The dataset comprised 2,641 periapical radiographs from 2,600 patients, split at the patient level into training (65%, n = 1,717), validation (10%, n = 264), and test (25%, n = 660) sets. Model robustness was further confirmed through 5-fold stratified patient-level cross-validation and leave-one-center-out (LOCO) evaluation. YOLO 11 achieved the highest performance with 83.90% accuracy (95% CI: 81.20-86.40%), 83.70% macro F1-score, 84.60% macro precision, and 84.20% macro recall on the independent test set, significantly outperforming VGG-19 (78.10%), MobileNet V2 (77.20%), and other architectures (p < 0.01). The 5-fold patient-level cross-validation yielded a mean accuracy of 84.55% ± 0.85% for YOLO 11, confirming model stability.

Conclusion: This study demonstrates the feasibility of deep learning-based diagnostic support for periapical radiograph classification. While YOLO 11 shows promising performance, this study is limited to data from three imaging centers within a single geographic region (Kermanshah, Iran), and clinical deployment requires external validation on multi-institutional, geographically diverse datasets and prospective trials to establish true generalizability and clinical utility. System should be used only as a second-opinion tool to assist, not replace, comprehensive clinical diagnosis.