Developing an artificial intelligence-based progressive growing GAN for high-quality facial profile generation and evaluation through turing test and aesthetic analysis
Abstract This study aimed to develop a Progressive Growing Generative Adversarial Network with Gradient Penalty (WPGGAN-GP) to generate high-quality facial profile images, addressing the scarcity of diverse training data in orthodontics. A dataset of 50,000 profile images, representing varied ages,...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Nature Portfolio
2025-07-01
|
| Series: | Scientific Reports |
| Subjects: | |
| Online Access: | https://doi.org/10.1038/s41598-025-11172-x |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract This study aimed to develop a Progressive Growing Generative Adversarial Network with Gradient Penalty (WPGGAN-GP) to generate high-quality facial profile images, addressing the scarcity of diverse training data in orthodontics. A dataset of 50,000 profile images, representing varied ages, genders, and ethnicities, was collected from two centers. The WPGGAN-GP model was trained to generate high-resolution images (1024 × 1024 pixels) using a progressive growing approach. Evaluation included both quantitative and qualitative assessments. The Sliced Wasserstein Distance (SWD) between real and generated images reached 0.026. A Turing test was conducted with 15 observers (orthodontists, surgeons, and laypersons), each assessing 100 images (50 real, 50 generated). Average classification accuracies were 0.58, 0.578, and 0.46 for orthodontists, surgeons, and laypersons, respectively. Aesthetic evaluation involved six key facial angles, with only the naso-frontal angle showing a statistically significant difference (p = 0.032). The intra-class correlation coefficient (ICC) for repeated measurements ranged from 0.952 to 0.968, and inter-rater ICC values exceeded 0.90, indicating excellent measurement consistency. Additionally, signal-to-noise ratio (SNR) analysis revealed no significant difference between real and generated images (p > 0.05), confirming comparable photometric quality. The results suggest that the WPGGAN-GP model effectively generates realistic facial profiles with both anatomical and perceptual fidelity. This approach offers valuable applications in orthodontic education, treatment simulation, and data augmentation, particularly where patient privacy and dataset balance are critical. Future research should explore conditional generation models for specific malocclusion types and further diversify training data to enhance clinical relevance. |
|---|---|
| ISSN: | 2045-2322 |