AI-based classification of pediatric sagittal skeletal patterns using multi-view facial photographs

The increasing global demand to assess pediatric skeletal malocclusions poses a growing challenge, as current screening methods are dependent on availability of orthodontists and X-rays, which limits accessibility. This study aimed to develop and validate an artificial intelligence (AI) model based on facial photographs to predict pediatric sagittal skeletal pattern classifications and to explore its feasibility for radiation-free early screening of pediatric skeletal malocclusion.

Facial photographs and lateral cephalograms were collected retrospectively from a cohort of 2033 pediatric patients at a stomatology hospital between 2013 and 2023. The dataset was divided into a training (n = 1789) and an independent test set (n = 244). Data were annotated, preprocessed, and augmented. Six backbones were compared, and ConvNeXt-T was selected as the optimal backbone. An innovative primary- auxiliary feature fusion strategy was adopted to identify the primary feature inputs and we first trained four single-view models. Subsequently, the primary and auxiliary features from other views were fused to construct four multi-view feature fusion models. Performance was evaluated by using five-fold cross-validation, with Grad-CAM visualizing key feature areas. Finally, the optimal model was clinically validated against practicing orthodontists on the independent test set.

The 90° profile model achieved the highest accuracy of 86.36% among all single-view models. The optimal multi-view model further improved the accuracy to 88.30% (recall=81.79%, specificity=91.22%). Using the independent test set, the model's agreement with the gold standard was 9.25% higher than junior level orthodontists and 7.43% higher than intermediate orthodontists (both P < 0.05), and closely aligns with senior professionals (difference: 2.70%). Grad-CAM confirmed focus on the nasolabial and chin contour regions.

An innovative method was successfully established to predict pediatric sagittal skeletal patterns with high accuracy (>88%) using multi-dimensional facial characteristics. This lays the groundwork for a non-clinical, photographic screening approach for identify pediatric skeletal malocclusions.

This method provides a feasible, radiation-free solution for early large-scale screening in non-clinical settings, supporting initial referral decisions while promoting early interventions and stratified management of children's oral health. It should be emphasized that this tool is intended primarily to assist screenings and referrals, rather than replacing the essential work of expert professionals.