Document Type : Research Paper
Authors
1
Department of Basic Medical Sciences, College of Dentistry, University of Kerbala, Karbala, Iraq.
2
Department of Oral Diagnosis, College of Dentistry, University of Kerbala, Karbala, Iraq.
10.22059/jitm.2026.109978
Abstract
The process of dental image processing has gained more significance in the diagnosis of oral diseases since the dental imaging modalities produce complicated visual data that needs to be effectively and correctly processed by means of computation. The paper presents experimental analysis of the two approaches to the diagnosis of disease in dental images-classical and deep learning-based image processing-in a comparative manner. This is evaluated based on the segmentation and categorization of the major dental imaging modalities such as intraoral radiographs, panoramic radiographs, cone-beam computed tomography images, and bitewing images. To ensure consistency and improved model performance, all datasets went through a unified preprocessing pipeline of normalization, noise reduction, augmentation, and resizing. To compare in terms of Dice coefficient and Intersection over Union, thresholding, U-Net, and Attention U-Net were compared. To be classified, Basic CNN, ResNet50, and EfficientNet were tested with the help of accuracy, sensitivity, specificity, area under the curve, and F1-score. The results indicate that the methods that employ deep learning are always superior to the classical methods in the segmentation and classification of data. The best segmentation was achieved by U-Net, with the best classification accuracy and diagnostic discrimination in all dental image types being found with EfficientNet. These findings indicate that developed deep learning systems can be used with great potential to assist with automated dental diagnosis. Nevertheless, robustness, interpretability, computational efficiency, and wider real-world validation are also essential to successful clinical implementation.
Keywords