Evaluation of a New Inclusive Next-Generation Synthetic Face Tool for Dysmorphology

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Genetics

Background:

• Facial analysis tools for rare genetic syndromes face accuracy limitations due to small, biased datasets, particularly for ultra-rare conditions and underrepresented ethnicities.
• Synthetic facial images offer a potential solution to augment training data and promote equitable diagnostic performance.

Purpose of the Study:

• To develop and evaluate a synthetic face generation pipeline using advanced AI models to enhance rare genetic syndrome diagnosis.
• To assess the impact of synthetic data augmentation on the accuracy and inclusivity of AI-driven phenotyping algorithms.

Main Methods:

• A synthetic face generation pipeline was created using diffusion models (DreamShaper XL Turbo) with LoRA-based domain adaptation and ControlNet for pose conditioning.
• 4,432 synthetic faces across ten rare syndromes were generated, balanced for age, sex, and ethnicity.
• The performance of an ArcFace R-100 phenotyping algorithm was evaluated using real and synthetic data in various machine learning configurations.

Main Results:

• Synthetic faces exhibited high phenotypic realism, achieving top-1 accuracy of 0.823 and AUC of 0.991.
• Augmenting real training datasets with synthetic images improved accuracy from 0.766 to 0.869 and AUC from 0.988 to 0.993.
• Significant performance gains were observed for ultra-rare syndromes and Asian individuals, highlighting improved inclusivity.

Conclusions:

• Diffusion-generated synthetic faces can enhance the inclusivity and accuracy of AI-based dysmorphology tools.
• Synthetic datasets offer a scalable and ethical approach to overcome data scarcity in rare disease diagnostics.
• This technology has potential applications in clinical training, telemedicine, and improving healthcare in underserved regions.