Automated classification of sagittal, vertical, and transverse malocclusions from 3D intraoral scans via a multi-head CNN framework

Abstract

This study aimed to develop and evaluate a multi-task deep learning model for the simultaneous classification of sagittal, vertical, and transverse malocclusion components, as well as midline deviation, using three-dimensional (3D) intraoral scan (IOS) data. A total of 801 anonymized 3D IOS datasets were retrospectively collected. Surface mesh (STL) files were standardized and converted into six-view two-dimensional (2D) representations. This approach preserved morphological information while enabling the use of transfer learning-based convolutional neural networks (CNNs). A shared-backbone, multi-head architecture was designed to perform simultaneous classification tasks. Six CNN backbones were evaluated using accuracy, precision, recall, F1-score, and ROC-AUC. EfficientNetV2B3 achieved the highest performance in anterior bite classification (accuracy: 80%, ROC-AUC approximate to 0.89). ResNet50 provided the most balanced results for sagittal and midline classification. Transverse occlusion classification showed comparatively lower performance across all models. Multi-view representations derived from 3D intraoral scans enable simultaneous and clinically relevant classification of multiple malocclusion components. While strong performance was observed for anterior bite and midline tasks, sagittal and transverse classifications remain more challenging. IOS-based multi-task AI systems may serve as supportive tools in orthodontic diagnosis and digital workflows.