@Article{csse.2025.068616,
AUTHOR = {Sari Masri, Ahmad Hasasneh},
TITLE = {Explainable Transformer-Based Approach for Dental Disease Prediction},
JOURNAL = {Computer Systems Science and Engineering},
VOLUME = {49},
YEAR = {2025},
NUMBER = {1},
PAGES = {481--497},
URL = {http://www.techscience.com/csse/v49n1/64044},
ISSN = {},
ABSTRACT = {Diagnosing dental disorders using routine photographs can significantly reduce chair-side workload and expand access to care. However, most AI-based image analysis systems suffer from limited interpretability and are trained on class-imbalanced datasets. In this study, we developed a balanced, transformer-based pipeline to detect three common dental disorders: tooth discoloration, calculus, and hypodontia, from standard color images. After applying a color-standardized preprocessing pipeline and performing stratified data splitting, the proposed vision transformer model was fine-tuned and subsequently evaluated using standard classification benchmarks. The model achieved an impressive accuracy of 98.94%, with precision, recall and F1 scores all greater than or equal to 98% for the three classes. To ensure interpretability, three complementary saliency methods, attention roll-out, layer-wise relevance propagation, and LIME, verified that predictions rely on clinically meaningful cues such as stained enamel, supragingival deposits, and edentulous gaps. The proposed method addresses class imbalance through dataset balancing, enhances interpretability using multiple explanation methods, and demonstrates the effectiveness of transformers over CNNs in dental imaging. This method offers a transparent, real-time screening tool suitable for both clinical and tele-dentistry frameworks, providing accessible, clarity-guided care pathways.},
DOI = {10.32604/csse.2025.068616}
}