From Image to Insight: Using Vision Transformers to Revolutionize Dental Caries Assessment in Radiographic Imaging

Abstract

Background: Dental caries is a prevalent chronic disease affecting people of all ages and socioeconomic backgrounds. Early detection and accurate diagnosis are crucial for effective treatment planning and preventing further progression. Artificial intelligence advancements, including CNNs and deep learning technologies like Vision Transformers, are improving diagnostic accuracy for carious lesions in medical image analysis, capturing global dependencies and contextual information.This study evaluates a Vision Transformer-based approach for automated dental caries classification using dental radiographs. The model improves diagnosis efficiency, reduces manual examination time and cost, and enhances access to dental care by addressing challenges like structure complexity and image quality, using advanced image preprocessing techniques.
Methods: This study used 110 intraoral periapical (IOPA) images from public databases, with 70 assigned for pulp dental caries and 40 for non-pulp dental caries. The dataset consisted of 104 images, with 82 for training and 22 for validation. Labels were assigned automatically from class subdirectories, and the default worker configuration optimized data loading and throughput.vision transformer model, pre-trained on the ImageNet21k dataset, was utilized for binary classification with two classes and evaluated for accuracy metrics .
Results: The model achieved a maximum of 98% and 89.50% in training and validation, with a final training loss of 0.0069 and a final validation loss of 0.0101. It showed moderate performance, with a PR AUC of 0.537, and an overall accuracy of 52.0%. The model showed consistent decrease in loss curves, steady improvement in accuracy, stable learning progression, and no significant overfitting.The model achieved a maximum of 98% and 89.50% in training and validation, with a final training loss of 0.0069 and a final validation loss of 0.0101. It showed moderate performance, with a PR AUC of 0.537, and an overall accuracy of 52.0%. The model showed consistent decrease in loss curves, steady improvement in accuracy, stable learning progression, and no significant overfitting.
Conclusion: The Vision Transformer model, while predicting and classifying dental caries involving pulp, has limitations in generalizing to unseen data. Future improvements should focus on optimizing the model architecture, expanding the dataset diversity, and implementing advanced techniques for improved diagnostic support.