Detecting white spot lesions on post-orthodontic oral photographs using deep learning based on the YOLOv5x algorithm:...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Biomedical Imaging

Background:

• Deep learning models excel at identifying subtle visual differences.
• The YOLOv5x algorithm was assessed for detecting white spot lesions (WSLs) in post-orthodontic dental images.
• This research serves as a preliminary step towards fully automated clinical diagnostic tools.

Purpose of the Study:

• To evaluate the performance of the Convolutional Neural Network (CNN)-based YOLOv5x algorithm in detecting white spot lesions (WSLs) in post-orthodontic oral photographs.
• To assess the algorithm's effectiveness using a limited dataset.
• To lay the groundwork for future fully automated clinical integration.

Main Methods:

• 435 JPG images were labeled for white spot lesions using CranioCatch software.
• Images were resized to 640×320 while preserving aspect ratio for model training.
• The YOLOv5x algorithm was employed for deep learning, with performance analyzed via ROC analysis and confusion matrix (TP, FP, FN determined).

Main Results:

• The model achieved a precision of 0.786, recall of 0.618, and F1 score of 0.692 for WSL detection.
• The Area Under the Curve (AUC) from ROC analysis was 0.712.
• The mean Average Precision (mAP) from the Precision-Recall curve was 0.425.

Conclusions:

• The YOLOv5x model's accuracy and sensitivity for WSL detection were below practical application standards.
• However, the detection rate is considered promising and acceptable when compared to prior studies.
• Future improvements can be achieved by expanding the training dataset and refining the deep learning algorithm.