RESEARCH ARTICLE
Deep Learning–Based Classification of Early Occlusal Caries from QLF Fluorescence Images: A Conceptual Validation Study
Eun-Ha Jung
Department of Dental Hygiene, College of Science and Technology, Kyungpook National University, Sangju, 37224, Korea
Corresponding Author: Eun-Ha Jung, E-mail: jeunha725@knu.ac.kr
Volume 14, Number 1, Pages 61–68, March 2026.
Korean J Clin Dent Hyg 2026;14(1):61–68. https://doi.org/10.12972/kjcdh.2026.14.1.6
Received on January 12, 2026, Revised on January 27, 2026, Accepted on January 30, 2026, Published on March 31, 2026.
Copyright © 2026 by Korean Society of Infection Control and Prevention in Dental Hygiene.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License(http://creativecommons.org/licenses/by-nc/4.0/).
Abstract
Objective: This study explored the preliminary feasibility of automatically classifying early- stage occlusal caries using deep learning applied to QLF fluorescence images. We examined whether the QS-Occlusal classification system could be implemented within an artificial intelligence framework by learning fluorescence-derived image patterns. Methods: QLF images were acquired from a single extracted tooth and classified into four QS-Occlusal stages (QS 0–3) according to fluorescence loss and red fluorescence characteristics. A total of 100 images were included; 83 were used for training and 17 for evaluation. A DenseNet-201 convolutional neural network pretrained on ImageNet was fine-tuned for four-class classification. Model performance was evaluated using confusion matrix analysis, accuracy, precision, recall, and F1-score. Results: Most samples were correctly classified in the confusion matrix. Agreement was highest for QS 0 and QS 3, while moderate misclassification was observed between QS 1 and QS 2, likely due to overlapping fluorescence patterns in early enamel lesions. Overall performance metrics indicate stable and reliable classification. Conclusions: Within the limitations of a single-tooth, proof- of-concept dataset, this study demonstrates the feasibility of deep learning–based QS-Occlusal classification using QLF images. This approach has the potential to support standardized assessment of early occlusal caries and provides a foundation for further validation using larger and more diverse datasets.
Keywords
Automatic classification, Deep learning, Early caries, QLF, Quantitative evaluation