Artificial intelligence for facial attractiveness assessment in orthodontics: model development and evaluation.

BACKGROUND

Achieving valid and reproducible facial attractiveness assessment in orthodontic practice remains challenging. Artificial intelligence (AI) has shown potential in orthodontics and may help reduce inconsistencies in the evaluation of facial and smile aesthetics. This study aimed to develop and internally validate an AI model for assessing facial attractiveness in a Caucasian population.

MATERIALS AND

METHODS

The study sample was derived from a subset of the SCUT-FBP5500 V2 dataset and included Caucasian subjects, comprising 750 males and 750 females with human-rated facial attractiveness scores on a 1-5 scale. The model was based on a ResNet-50 backbone pretrained on ImageNet. Training was performed using the Adam optimizer (learning rate 0.0001) and mean squared error loss for up to 100 epochs. Stratified data splits of 70% for training, 10% for validation, and 20% for testing were created according to facial attractiveness score and sex. During training, data augmentation techniques, including resizing, random horizontal flipping, rotation, colour jittering, and tensor conversion, was applied, whereas the validation and test sets underwent resizing only. All experiments were conducted on a Linux-based system with an NVIDIA RTX 2080 GPU, using Python 3.12 and PyTorch v2.9. Model performance was assessed using MAE and RMSE across the test set, by sex, and across facial attractiveness-score quintiles.

RESULTS

The model demonstrated predictive performance on the test set, with an MAE of 0.215 (95% CI: 0.196-0.234) and an RMSE of 0.276 (95% CI: 0.251-0.300). Sex-specific analysis showed marginally better performance in males, with an MAE of 0.201 (95% CI: 0.177-0.225), compared with 0.228 (95% CI: 0.198-0.264) in females (P = 0.662). The highest accuracy was observed in the densely represented segment of the dataset, where the MAE was 0.156 (95% CI: 0.119-0.199). Error metrics varied across attractiveness-score ranges and were highest in the (3.02-3.63] range.

CONCLUSIONS

The proposed model closely approximated human ratings of facial attractiveness and demonstrated predictive accuracy. These findings support the potential of AI-based tools to provide more consistent aesthetic assessments in orthodontics. Future research should include diverse populations to improve generalizability and clinical applicability.