A Review on Deep Learning Frameworks for Dental Anomaly and Disease Classification

Abstract

Oral anomalies and dental diseases affect billions of people worldwide, yet diagnosis often relies on manual interpretation of radiographs and clinical images, which is time-consuming and prone to variability. Advances in deep learning (DL) have opened new opportunities for accurate, efficient, and scalable dental diagnostics. This review examines state-of-the-art DL frameworks applied to dental imaging modalities, in-cluding intraoral RGB photographs, bitewing and periapical radiographs, panoramic radiography, and cone-beam computed tomography (CBCT). The analysis covers pre-processing pipelines, backbone architectures (convolutional neural networks and vi-sion transformers), task designs (classification, detection, segmentation, hybrid mod-els), and strategies for addressing data imbalance, calibration, and uncertainty. Find-ings reveal that modality-specific preprocessing enhances reliability, hybrid CNN-Transformer models improve performance for wide-field or complex tasks, and segmentation-assisted classification increases sensitivity to subtle lesions. Moreover, calibrated probability outputs, robust evaluation metrics (ROC-AUC, PR-AUC), and external validation are essential for clinical readiness. The review identifies critical gaps—limited cross-site generalization, under-reported calibration, and scarce re-al-world validation—and outlines future directions such as label-efficient learning, federated training, and calibration-first pipelines. With these safeguards, DL-based systems can evolve from experimental tools to trustworthy clinical aids that strength-en diagnostic accuracy and decision support in dentistry.