Point cloud deep learning-based automatic grading system for tooth wear

This study aimed to construct a machine learning model for automatic grading of the degree of wear of each tooth.

Intraoral scan models from 200 patients (4,556 teeth) were retrospectively collected and independently annotated by three calibrated dentists using the 2001 Lobbezoo tooth wear classification system. After exclusions, 4,256 teeth were used for model training and 300 for testing (balanced 100 per wear grade: no, mild, moderate)AQ. The SnowflakeNet method with point cloud deep learning was applied to extract morphological features and construct a three-class classification model. The SnowflakeNet method with snowflake point deconvolution was used to generate point clouds, extract the morphological features of each tooth, and construct a machine-learning model. The performance of the classification model was evaluated using a confusion matrix; receiver operating characteristic and precision recall curves; and area under the curve (AUC).

The overall accuracy of the Point-MAE model with pre-training was 73.7%. Class-wise performance varied: moderate wear achieved the highest precision (91.4%) and specificity (96.5%), while mild wear-the most clinically critical category for early intervention-showed lower precision (61.6%) and F1-score (0.65). The area under the curve (AUC) values were 0.88, 0.80, and 0.92 for no, mild, and moderate wear, respectively. In external validation on 50 patients (1,363 teeth), the weighted kappa coefficient for agreement between AI and dentists was 0.827 (95% CI: 0.802-0.852), indicating strong consistency.

The point cloud learning model could achieve automatic intraoral scan model grading, demonstrating potential as an assistive tool that requires further optimization and validation for grading teeth with no, mild, and moderate wear. However, its grading efficacy must be further optimizeAQd.