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

INTRODUCTION AND AIMS: Maxillary transverse deficiency (MTD) is conventionally evaluated using cone-beam computed tomography (CBCT), which entails increased radiation exposure, cost, and clinical workload. Frontal intraoral photographs are routinely obtained in orthodontic practice. Using CBCT-derived transverse measurements as reference standards, we developed, validated, and compared multiple deep learning (DL) models to assess the feasibility of identifying MTD from frontal intraoral photographs.

METHODS

This study included 826 internal and 192 external patients who underwent paired frontal intraoral photographs and CBCT. MTD was determined based on the University of Pennsylvania analysis (UPA) and Yonsei transverse analysis (YTA) labels. DenseNet 121, ResNet 18, EfficientNet B0/B3, and MobileNetV3 Small/Large were trained separately on photographs using UPA- and YTA-based labels. Five-fold cross-validation was employed, and performance was evaluated on the internal and external test sets using accuracy, precision, recall, and F1 scores, along with confusion matrices and areas under the receiver operating characteristic curves. DeLong's test assessed the differences between the models.

RESULTS

In the external test set under the UPA labelling scheme, ResNet 18 achieved the highest accuracy (90.62%). Under the YTA labelling scheme, DenseNet 121 and ResNet 18 achieved the highest accuracy (96.88%). Across all internal and external test sets using both labelling schemes, DenseNet 121 and ResNet 18 yielded the best overall performance, and no statistically significant difference was observed between the two models (P > .05).

CONCLUSIONS

The DL models demonstrated strong potential for analysing frontal intraoral photographs to detect MTD. These findings provide initial insights into the use of DL models to identify MTD from frontal intraoral photographs for orthodontic purposes.

CLINICAL RELEVANCE: This study demonstrates the feasibility of using DL-based recognition of frontal intraoral photographs to identify MTD. As a cost-effective adjunctive tool, the proposed approach may assist clinicians in identifying MTD and help improve case selection for CBCT imaging.