Influence of Intraoral Scanners, Operators, and Data Processing on Dimensional Accuracy of Dental Casts for Unsupervised Clinical Machine Learning: An In Vitro Comparative Study

Abstract

Purpose. This study assessed the impact of intraoral scanner type, operator, and data augmentation on the dimensional accuracy of in vitro dental cast digital scans. It also evaluated the validation accuracy of an unsupervised machine-learning model trained with these scans. Methods. Twenty-two dental casts were scanned using two handheld intraoral scanners and one laboratory scanner, resulting in 110 3D cast scans across five independent groups. The scans underwent uniform augmentation and were validated using Hausdorff’s distance (HD) and root mean squared error (RMSE), with the laboratory scanner as reference. A 3-factor analysis of variance examined interactions between scanners, operators, and augmentation methods. Scans were divided into training and validation sets and processed through a pretrained 3D visual transformer, and validation accuracy was assessed for each of the five groups. Results. No significant differences in HD and RMSE were found across handheld scanners and operators. However, significant changes in RMSE were observed between native and augmented scans with no specific interaction between scanner or operator. The 3D visual transformer achieved 96.2% validation accuracy for differentiating upper and lower scans in the augmented dataset. Native scans lacked volumetric depth, preventing their use for deep learning. Conclusion. Scanner, operator, and processing method did not significantly affect the dimensional accuracy of 3D scans for unsupervised deep learning. However, data augmentation was crucial for processing intraoral scans in deep learning algorithms, introducing structural differences in the 3D scans. Clinical Significance. The specific type of intraoral scanner or the operator has no substantial influence on the quality of the generated 3D scans, but controlled data augmentation of the native scans is necessary to obtain reliable results with unsupervised deep learning.