Automated Detection and Segmentation of Dental Caries in Panoramic Radiographs Using YOLOv8-Based Deep Learning Model.

OBJECTIVES

This study aimed to develop an automated dental caries localization system through detection and classification using a deep learning (DL) approach based on You Only Look Once Version 8 (YOLOv8).

MATERIALS AND

METHODS

YOLOv8 models were developed using 750 panoramic radiographs, which were divided into training, validation, and test sets at a ratio of 80%:10%:10%. The training and validation data were cropped into six regions to enhance the detection model performance. Confusion matrices were utilized to assess performance metrics, including accuracy, sensitivity, specificity, precision, and F1 score.

STATISTICAL ANALYSIS: Fleiss's kappa was used to evaluate overall agreement among the YOLOv8 predictions and the expert evaluations, followed by Cohen's kappa for each comparison category.

RESULTS

For caries detection, the YOLOv8-M model achieved an optimal accuracy of 95.22%, a sensitivity of 84.07%, a specificity of 98.18%, a precision of 92.46%, and an F1 score of 88.06%. For segmentation, the YOLOv8-X model achieved an optimal accuracy of 93.92%, a sensitivity of 78.30%, a specificity of 98.01%, a precision of 91.15%, and an F1 score of 84.24%. Fleiss's kappa analysis indicated substantial agreement of 0.766 (p < 0.001) between model predictions and expert evaluations.

CONCLUSION

The YOLOv8 models demonstrated strong performance in detecting and segmenting dental caries in panoramic radiographs. They also showed substantial agreement with expert assessments, highlighting their potential as a reliable tool for automated caries identification. Further development is recommended to enhance performance, particularly by reducing false positive and false negative cases.