Application of 3D neural networks and explainable AI to classify ICDAS detection system on mandibular molars

Considerable variations exist in cavity preparation methods and approaches. Whether the extent and depth of cavity preparation because of the extent of caries affects the overall accuracy of training deep learning models remains unexplored.

The purpose of this study was to investigate the difference in 3-dimensionsal (3D) model cavity preparations after International Caries Detection and Assessment System (ICDAS) classification performed by different practitioners and the subsequent influence on the ability of a deep learning model to predict cavity classification.

Two operators prepared 56 restorative cavities on simulated mandibular first molars according to 4 ICDAS classifications, followed by 3D scanning and computer-aided design processing. The surface area, virtual volume, Hausdorff distance (HD), and Dice Similarity Coefficients were computed. Multivariate analysis of variance was used to assess cavity size and operator proficiency interactions, and 1-way ANOVA was used to evaluate HD differences across 4 cavity classifications (α=.05). The 3D convolutional neural network (CNN) predicted the ICDAS class, and Saliency Maps explained the decisions of the models.

Operator 1 exhibited a cavity preparation surface area of 360.55 ±15.39 mm², and operator 2 recorded 355.24 ±10.79 mm². Volumetric differences showed operator 1 with 440.41 ±35.29 mm³ and operator 2 with 441.01 ±35.37 mm³. Significant interactions (F=2.31, P=.01) between cavity size and operator proficiency were observed. A minimal 0.13 ±0.097 mm variation was noted in overlapping preparations by the 2 operators. The 3D CNN model achieved an accuracy of 94.44% in classifying the ICDAS classes with a 66.67% accuracy when differentiating cavities prepared by the 2 operators.

Operator performance discrepancies were evident in the occlusal cavity floor, primarily due to varying cavity depths. Deep learning effectively classified cavity depths from 3D intraoral scans and was less affected by preparation quality or operator skills.