Abstract
BACKGROUND: Third-dimensional binary localization, determining whether impacted maxillary canines (IMCs) requiring surgical exposure lie buccally or palatally, is among the key parameters influencing the choice of surgical approach. In 30–50% of cases, the tooth is non-palpable, necessitating further radiographic evaluation. This study investigated whether convolutional neural networks (CNNs) applied to panoramic radiographs (PRs) can accurately classify IMC position, enhancing their diagnostic value in preoperative assessment. METHODS: This retrospective study included 494 IMCs from 472 PRs. Surgical notes and CBCT findings served as reference standards. IMCs were annotated on PRs and classified using CNN architectures across multiple preprocessing pipelines. Deep feature embeddings were further evaluated using logistic regression (LR), support vector machines (SVM), and k-nearest neighbors (KNN). Model performance was assessed using accuracy, macro-averaged precision, recall, F1-score, and ROC-AUC. RESULTS: The ResNet50 + LR model yielded the highest performance (accuracy: 0.898, F1-score: 0.897, ROC-AUC: 0.945), outperforming InceptionV3 and all softmax-based models. Preprocessing had minimal effect on ResNet50 but improved InceptionV3 outcomes. Most misclassifications occurred in buccal cases. Feature-space analysis revealed favorable linear separability for LR. CONCLUSIONS: CNN-based analysis of PRs enables accurate IMC localization, with ResNet50 + LR demonstrating consistent performance. These findings support the potential of AI-assisted PR interpretation as a viable step toward enhancing clinical utility of two-dimensional imaging.