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

Background Facial analysis tools can assist in diagnosing rare genetic syndromes, but their accuracy is limited in ultra-rare conditions and underrepresented ethnicities due to small, biased datasets. Synthetic facial images could enrich training data and improve equity in diagnostic performance. Methods We developed a synthetic face generation pipeline using diffusion models (DreamShaper XL Turbo), enhanced with LoRA-based syndrome-specific domain adaptation and pose conditioning via ControlNet. A total of 4432 synthetic faces were generated across ten rare syndromes, balanced by age (0-18 years), sex (50/50), and ethnicity (33% Caucasian, Afro-Caribbean, Asian). Synthetic and real data were used in four machine learning designs to train and test ArcFace R-100, a phenotyping algorithm for syndromic classification. Results Synthetic faces generated from real patient data achieved high phenotypic realism, with classification performance reaching top-1 accuracy of 0.823 and AUC of 0.991. Adding synthetic images to real training datasets increased accuracy on real test images from 0.766 to 0.869 and improved AUC from 0.988 to 0.993. Performance gains were most significant for ultra-rare syndromes and Asian individuals (top-1: 0.971). Training with synthetic images alone yielded lower accuracy (top-1: 0.606), underscoring their complementary role. Conclusion This study demonstrates that diffusion-generated synthetic faces can enhance inclusivity and accuracy in AI-based dysmorphology tools. These synthetic datasets provide a scalable, ethical solution to data scarcity, with applications in clinical training and telemedicine, especially in underserved regions.