Deep Learning-Based Multi-Class Prediction of Enamel Caries Severity Using EfficientNet and DenseNet with CLAHE Enhancement

Saja Abdullah Al-Baghdadi · Ali A. Al-Kharaz

Dental caries is one of the most frequent chronic diseases in dentistry, and hence it needs to be diagnosed early for the prevention of lesion growth and enhanced treatment success. This paper introduces a machine learning method to classify dental caries lesions into different severity levels based on intraoral pictures. Four different categories have been considered in this study: healthy teeth without any caries, early low severity caries, early high severity caries, and advanced caries. Different deep learning models such as Convolution Neural Networks were utilized under the three optimization methods including Adam, AdamW, and RMSprop. Moreover, data balancing techniques and Contrast Limited Adaptive Histogram Equalization (CLAHE) have been used to enhance the data and obtain better results. Based on the experimental work conducted, EfficientNet-B1 along with AdamW optimizer and CLAHE gave the best results among all other models, obtaining maximum accuracy and ROC-AUC values with the minimum number of errors.