Deep Learning-Based Automated Detection and Severity Classification of Dental Caries on Panoramic Radiographs

Abstract

This study aimed to develop and evaluate a deep learning-based model for automated detection and severity classification of dental caries on panoramic radiographs.

A total of 500 panoramic radiographs were collected and divided into training (n = 400), validation (n = 50), and testing (n = 50) datasets. Two separate annotation sets were prepared to develop a caries detection and a severity classification model based on the International Caries Classification and Management System (ICCMS). Image annotation was performed using Roboflow and validated by two radiologists. Four variants of the YOLOv5 architecture (s, m, l, and x) were trained on Google Colab Pro for both tasks. Model performance was evaluated using confusion matrices at five confidence thresholds (10%, 20%, 30%, 40%, and 50%). Accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and false case distribution were then analyzed.

The YOLOv5s and YOLOv5x variants achieved the highest overall performance for the 1-class caries detection model (recall 73.8%, precision 78.8%, mean average precision at IoU 0.50 [mAP50] 74.3%) and the 2-classes severity classification model (recall 55.8%, precision 61.0%, mAP50 54.8%), respectively. The 1-class detection model outperformed the severity classification model, with average accuracy, sensitivity, specificity, PPV, and NPV of 94.32%, 49.40%, 97.70%, 72.75%, and 95.99%, respectively, compared with 93.41%, 45.90%, 96.86%, 62.07%, and 95.91% for the 2-class model.

The YOLOv5 model demonstrated reliable performance in the automated detection of dental caries and classification of lesion severity on panoramic radiographs, with stronger performance observed in detection, highlighting their potential as a supportive tool in clinical decision-making.