A Hybrid Deep Learning and Anomaly Detection Model for Interpretable Analysis of Dental OPG
Dental radiography, Deep learning, Anomaly detection, Explainable AI, OPG analysis, medical image classification
AuthorsAbstractDental panoramic radiography (orthopantomogram or OPG) is a vital tool in diagnosing a range of dental pathologies, yet manual interpretation is timeconsuming and subject to variability. This study proposes a hybrid framework that combines deep convolutional neural networks (CNNs) with statistical anomaly detection and explainable artificial intelligence (XAI) to enhance both diagnostic accuracy and clinical interpretability. A fine-tuned ResNet-50 model was trained to extract contextual features from OPG images, which were then fused with point anomaly scores generated by an Isolation Forest algorithm. The system was evaluated on a dataset comprising six diagnostic categories, including rare conditions like fractured teeth and infections. Compared to a baseline CNN, the hybrid model demonstrated higher test accuracy (43.26% vs. 35.12%), macro-F1 score (0.21 vs. 0.10), and macro-AUC (0.70 vs. 0.61). XAI tools-Grad-CAM, SHAP, and saliency maps were employed to visualise decision-critical regions, providing transparent, multi-angle explanations aligned with clinical reasoning. The results confirm that the proposed hybrid approach enhances both performance and trustworthiness, making it a practical solution for AI-assisted dental diagnostics. Future research will explore model generalizability using larger datasets and multi-modal imaging.
1. Introduction Orthopantomography (OPG), also known as dental panoramic radiography (DPR), is a vital type of imaging in dental diagnostic studies, providing a twodimensional image of the entire maxillofacial area. It is regularly employed in detecting a wide variety of conditions, including caries, impacted teeth, infections, fractured roots and developmental anomalies. These images, however, are heavily dependent on clinical expertise in their interpretation and are prone to both inter- and intra-observer variability. To address this problem the past few years, have experienced a significant increase in the integration of artificial intelligence (AI) into dental diagnostics, especially due to the popularity of deep learning methods. CNNs have played a role in the development of automated dental image recognition. Their capability of extracting hierarchical spatial features has made them very applicable in use in dental pathology detection, such as caries detection, tooth segmentation and anomaly classification. Research conducted recently has shown that CNNs are useful in the detection of dental restorations and cavities using panoramic radiographs with promising accuracy and reliability (1,2). Moreover, the latest architectures, such as transformer-based networks, have facilitated the division of complicated anatomical features in OPG images, which provided a better definition of dental components (3). The systematic reviews verify a rapid development of AI-based tools in the field of dentistry and their increasing potential in clinical adoption (4). Simultaneously, a number of studies investigated frameworks that can use deep learning to identify and classify abnormalities in dental images, as well as rare and subtle lesions (5). Such developments predetermined the more intelligent, automated, and reproducible diagnostic systems. •••••••••••••••••••••••••••••••• ejprd.org - Published by Dennis Barber Journals. Barber Ltd. All rights reserved
EJPRD
Copyright ©2025 by Dennis