Comparing Deep Learning Models for Identifying Maxillary Transverse Deficiency from Intraoral Photographs.

Area of Science:

• Orthodontics
• Artificial Intelligence
• Medical Imaging

Background:

• Maxillary transverse deficiency (MTD) diagnosis typically relies on cone-beam computed tomography (CBCT), which involves significant radiation exposure, cost, and clinical workload.
• Frontal intraoral photographs are a routine diagnostic tool in orthodontics, offering a potentially more accessible alternative for MTD assessment.

Purpose of the Study:

• To develop and validate deep learning (DL) models for identifying MTD using only frontal intraoral photographs.
• To compare the performance of various DL models in detecting MTD based on established diagnostic criteria.

Main Methods:

• Trained multiple DL models (DenseNet 121, ResNet 18, EfficientNet, MobileNetV3) on frontal intraoral photographs from 826 internal and 192 external patients.
• Utilized CBCT-derived transverse measurements and two labelling schemes (University of Pennsylvania analysis and Yonsei transverse analysis) for model training and validation.
• Evaluated model performance using accuracy, precision, recall, F1 scores, confusion matrices, and receiver operating characteristic curves, with DeLong's test for statistical comparison.

Main Results:

• ResNet 18 achieved 90.62% accuracy on the external test set using the UPA labelling scheme.
• DenseNet 121 and ResNet 18 achieved 96.88% accuracy on the external test set using the YTA labelling scheme.
• DenseNet 121 and ResNet 18 demonstrated the best overall performance across all datasets and labelling schemes, with no statistically significant difference between them.

Conclusions:

• Deep learning models show significant potential for analyzing frontal intraoral photographs to detect MTD.
• This DL-based approach offers a feasible, cost-effective adjunctive tool for clinicians to identify MTD and improve CBCT imaging selection.