Enhancing Caries Detection in Bitewing Radiographs Using YOLOv7.

Area of Science:

• Artificial Intelligence in Dentistry
• Radiology and Imaging Technology
• Machine Learning for Medical Diagnostics

Background:

• Dental caries detection from bitewing radiographs is crucial for timely intervention.
• Traditional methods can be subjective and time-consuming.
• Automated detection using deep learning models offers potential for improved efficiency and accuracy.

Purpose of the Study:

• To evaluate the impact of image size, Intersection over Union (IoU) thresholds, and confidence thresholds on YOLO model performance for dental caries detection.
• To compare the efficacy of YOLOv3 and YOLOv7 models in identifying carious lesions on bitewing radiographs.

Main Methods:

• Utilized 2575 bitewing radiographs annotated using the ICCMS™ radiographic scoring system.
• Employed YOLOv3 and YOLOv7 models with varying configurations (image sizes, IoU, confidence thresholds).
• Evaluated model performance using precision, recall, F1-score, and mean average precision (mAP).

Main Results:

• YOLOv7 demonstrated superior performance over YOLOv3, achieving higher precision, F1-score, and mAP.
• Increasing image size from 640x640 to 1280x1280 pixels did not significantly alter YOLOv7's mAP.
• Elevating IoU from 50% to 75% and confidence threshold from 0.001 to 0.5 improved precision but reduced recall and overall mAP.

Conclusions:

• YOLOv7 is a more effective model than YOLOv3 for automated dental caries detection in bitewing radiographs.
• Image size optimization is not a primary driver for performance enhancement in this context.
• Adjusting IoU and confidence thresholds requires careful balancing to optimize for both precision and recall in caries detection.