Object Detection Models Facilitating Self-Monitoring Of White Spot Lesions

This study evaluates mainstream deep learning models for detecting post-orthodontic white spot lesions, incorporating explainability analysis to reveal decision-making processes. Results show YOLOv8 excels in accuracy and matches orthodontists' assessments, highlighting its potential for enabling patient self-monitoring through reliable AI-driven detection. 529 intra-oral photographs with white spot lesions were collected and annotated. SqueezeNet, RetinaNet, FasterRCNN, and YOLOv8 models were fine-tuned based on pre-trained model weights. The performances of models were evaluated with metrics including mean average precision, f1-score, and area under ROC curve. Ablation-CAM technique was used for explainability analysis. YOLOv8′s performance was the best among all 4 models ([email protected]:0.95: 0.80 for tooth, 0.48 for white spot lesions, [email protected]: 0.98 for tooth, 0.75 for white spot lesions, Cohen's Kappa: 0.82 for tooth, 0.76 for white spot lesions). FasterRCNN's performance metrics were similar to YOLOv8, but it focused on a different subset of white spot lesion targets: 20% of white spot lesions could only be detected by FasterRCNN and RetinaNet. Explainability analysis using Ablation-CAM demonstrated that YOLOv8 has the ability to implicitly learn the distribution pattern of white spot lesions, which contributes to its superior performance. For the detection of white spot lesions, YOLOv8 demonstrated the best accuracy metrics and substantial agreement with professional orthodontics as indicated by Cohen's Kappa metric.