Classification of pediatric dental diseases from panoramic radiographs using natural language transformer and deep learning models

Introduction: Accurate classification of pediatric dental diseases from panoramic radiographs is essential for early diagnosis and effective treatment planning. While deep learning models traditionally operate directly on image data, text-based representations generated from radiographs may provide an alternative strategy for disease classification. Methods: This study proposed a text-driven framework in which a natural language transformer was used to generate structured textual descriptions from panoramic radiographs. These descriptions were subsequently classified for binary disease detection using three deep learning architectures: a one-dimensional convolutional neural network (1D-CNN), a long short-term memory (LSTM) network, and a pretrained Bidirectional Encoder Representations from Transformer (BERT) model. Model performance was evaluated and compared against three pretrained convolutional neural networks trained directly on radiographic images. Results: The 1D-CNN achieved the highest performance with 84% accuracy, demonstrating balanced classification across disease categories. The BERT model reached 77% accuracy, showing strong performance in detecting periapical infections but comparatively lower sensitivity for caries identification. The LSTM model performed substantially worse, achieving 57% accuracy. Both the 1D-CNN and BERT text-based approaches outperformed the three image-based pretrained CNN models. Discussion: These findings suggest that text-based classification of panoramic radiographs is a potential alternative to conventional image-based deep learning methods. Language-driven models show promise for radiographic interpretation; however, challenges remain in achieving consistent generalizability across disease types. Future research should focus on improving radiograph-to-text generation quality, developing hybrid architectures that integrate textual and visual features, and validating performance on larger and more diverse datasets to strengthen clinical applicability.