Integrating image processing and deep learning framework for automatic supernumerary tooth detection on occlusal radiographs.

Background/purpose: Supernumerary tooth is the most common form of hyperdontia in children and is typically identified through radiographic examination. Although occlusal radiographs are recommended for assessing anterior abnormalities, anatomical superimposition and variability in pediatric imaging make early detection challenging, thereby increasing diagnostic inconsistency. Artificial intelligence has shown strong potential in dental imaging, but applications involving pediatric occlusal radiographs remain limited due to low clinical utilization and insufficient datasets. Therefore, a reliable automated tool is needed to assist clinicians in improving supernumerary tooth screening accuracy. Materials and methods: We collected 600 pediatric occlusal radiographs from children aged 6-12 years. An image-processing algorithm was developed to automatically segment the tooth-bearing region and improve feature visibility. These processed images were used to train five convolutional neural networks (CNNs) for the classification of supernumerary tooth. A fuzzy control strategy was employed to integrate the outputs of the five models, utilizing majority voting and weighted confidence scoring to generate the final diagnostic decision and mitigate model-specific bias. Results: All five CNN models achieved accuracy above 90 %. EfficientNet performed best and reached 96.12 % accuracy and 96.06 % F1 score. Statistical comparison further showed that EfficientNet achieved the highest agreement with ground truth, achieving R Conclusion: The proposed hybrid image-processing and deep learning framework provides accurate and interpretable supernumerary tooth detection on occlusal radiographs, offering a reliable tool to support pediatric diagnosis and improve clinical decision-making.