Deep Learning Framework for Pediatric Dental Pathology Detection in Panoramic Radiographs

The interpretation of pediatric panoramic radiographic images presents significant challenges due to dynamic anatomical changes during tooth development. This study addresses the critical gap in pediatric dental datasets by developing a deep learning framework for dental disease detection. A custom dataset of 106 pediatric panoramic radiographs from patients aged 2-13 years was combined with 2,586 adult images to train and validate deep neural network (DNN) architectures. The framework employs image preprocessing, segmentation, and Chi-square (x^2)-based feature selection, reducing 1,000 extracted features to 13 discriminative vectors for classification. The research contributions are creation of the first publicly available pediatric dental panoramic radiograph dataset, application of x^2-based feature selection for dimensionality reduction, and comparative evaluation across multiple DNN architectures. The best-performing model (DNN-A) achieved 97.9% accuracy on segmented adult images and 80.1% on segmented pediatric images. However, performance on raw pediatric radiographs was limited to 31.1% accuracy, highlighting the need for automated segmentation integration. The model demonstrated 98% accuracy in detecting periapical infections when trained on adult segmented data. These results demonstrate both the potential and current limitations of deep learning approaches in pediatric dental diagnostics, emphasizing the necessity for larger pediatric datasets and end-to-end automated pipelines to achieve clinical utility.