Automated measurement of orthodontic digital dental arch models based on convolutional neural networks

Accurate automation of digital dental model analysis is critical for orthodontic diagnosis and planning, yet the current workflows depend on labor-intensive manual steps. Considering this requirement, a fully automated, measurement-ready model based on convolutional neural networks was developed in this study. This study aimed to develop and evaluate an automated software tool, named Fusion Analyzer, for the rapid and accurate analysis of digital orthodontic digital dental arch. Fusion Analyzer integrates a deep graph convolutional network (DeepGCN)-based tooth instance segmentation with multi-view convolutional neural network (MV-CNN) landmarking. Its performance was evaluated on 337 segmentation cases and 3,887 tooth meshes. Accuracy metrics included mean surface deviation for the jaws and teeth, as well as landmark L2 error distributions. The throughput was benchmarked against manual segmentation. Clinical validation compared intraoral and plaster scans, with intra-/inter-observer reliability (ICCs), arch-width differences (and relative error), and Curve of Spee agreement. Mean surface deviations were 0.39 mm for jaws and 0.20 mm for teeth. Landmark L2 errors were 0.31–0.33 mm (SD ≈ 0.22 mm), with >80% <0.3 mm and 2.04% >1 mm. Inference required ~ 25 s/arch vs. ~ 5 h for manual segmentation. Clinical validation revealed excellent reproducibility (intra-/inter-observer ICCs >0.997). Mean arch-width differences were ~ 0.30–0.66 mm (relative error 1.09–1.51%), and no significant differences were detected for the Curve of Spee. Fusion Analyzer achieves submillimeter accuracy with orders-of-magnitude faster processing, substantially alleviating the manual burden while preserving the measurement fidelity. These findings support reliable, scalable integration of automated model analysis into routine digital orthodontic workflows.