An Interpretable Deep Learning Framework for Multi-Class Dental Disease Classification from Intraoral RGB Images

Abstract

Dental anomalies and diseases are among the most prevalent health concerns world-wide, and their early and precise diagnosis is critical to ensuring effective treatment and improved patient outcomes. Traditional diagnostic approaches, particularly conventional radiography, are often time-consuming and may not provide sufficient diagnostic accuracy. To address these limitations, this study proposes a robust deep learning framework for the automated classification of dental conditions from intraoral RGB images. Three publicly available datasets—Oral Diseases (six classes), Oral Infection (six classes), and Teeth Dataset (five classes)—covering a broad spectrum of dental anomalies and exhibiting notable class imbalance were utilized. Five state-of-the-art convolutional neural network (CNN) architectures, namely Efficient-NetB3, EfficientNetB0, ResNet50, DenseNet121, and InceptionV3, were systematically evaluated using a unified transfer learning pipeline. Techniques such as stratified 5-fold cross-validation, ensemble inference, focal loss, class weighting, and label smoothing were employed to enhance generalization and mitigate class imbalance. EfficientNetB3 emerged as the optimal model, achieving accuracies of 95.4%, 89.9%, and 99.3% on the three datasets, with Kappa values reaching 0.989. Grad-CAM visualizations confirmed clinically meaningful feature localization, strengthening interpretability. The proposed framework demonstrates strong potential for integration into intelligent clinical decision-support systems, offering an optimal balance between diagnostic accuracy, computational efficiency, and transparency to assist dental practitioners in timely and reliable decision-making.