A unique AI-based tool for automated segmentation of pulp cavity structures in maxillary premolars on CBCT

This is the authors' abstract. We don't add key points for this paper.

To develop and validate an artificial intelligence (AI)-driven tool for the automatic segmentation of pulp cavity structures in maxillary premolars teeth on cone-beam computed tomography (CBCT). One hundred and eleven CBCT scans were divided into training (n = 55), validation (n = 14), and testing (n = 42) sets, with manual segmentation serving as the ground truth. The AI tool automatically segmented the testing dataset, with errors corrected by an operator to create refined 3D (R-AI) models. The overall AI performance was assessed by comparing AI and R-AI models, and thirty percent of the test sample was manually segmented to compare AI and human performance. Time-efficiency of each method was recorded in seconds (s). Statistical analysis included independent and paired t-tests to evaluate the effect of tooth type on accuracy metrics and AI versus manual segmentation. One-way ANOVA with Tukey's post hoc test was used for time efficiency analysis. A 5% significance level was used for all analyses.The AI tool demonstrated excellent performance with Dice similarity coefficients (DSC) ranging from 88% ± 7 to 93% ± 3 and 95% Hausdorff distances (HD) from 0.13 ± 0.06 to 0.16 ± 0.06 mm. Automated segmentation of maxillary second premolars performed slightly better than that of maxillary first premolars in terms of intersection over union (p = 0.005), DSC (p = 0.008), recall (p = 0.008), precision (p = 0.02), and 95% HD (p = 0.04). The AI-based approach showed higher recall (p = 0.04), accuracy (p = 0.01), and lower 95% HD than manual segmentation (p < 0.001). AI segmentation (42.8 ± 8.4 s) was 75 times faster than manual segmentation (3218.7 ± 692.2 s) (p < 0.001). The AI tool proved highly accurate and time-efficient, surpassing human expert performance.