AIDY (Artificial Intelligence for Dysmorphology): Automated Detection of Genetic Syndromes from 2D Photographs
Seance of wednesday 16 april 2025 (Intelligence Artificielle et Chirurgie)
DOI number : 10.26299/hmex-c760/2025.15.03
Abstract
Introduction
Characteristic craniofacial features are present in 30 to 40% of rare diseases, of which over 7,000 have been identified globally. Their recognition generally relies on the subjective expertise of clinical dysmorphologists. To democratise and broaden access to facial dysmorphology diagnostics, we developed AIDY, a novel mobile application that leverages artificial intelligence (AI) for the automated recognition of genetic syndromes using standard 2D facial photographs.
Methods
Our dataset was derived from the photographic archives of the Maxillofacial Surgery, Plastic Surgery, and Medical Genetics departments at Necker–Enfants Malades Hospital (AP-HP), comprising over one million images from more than 30,000 patients seen since 1976. The dataset was divided into 80% for training, 10% for validation, and 10% for testing. A deep learning model (ArcFace R-100, pre-trained on Glint360K) was fine-tuned on this dataset, which included over 500 genetically confirmed syndromes as well as non-syndromic controls. Model performance was evaluated using the area under the curve (AUC), along with top-1 and top-3 classification accuracy. In parallel, we employed a probabilistic diffusion model to generate representative synthetic facial composites for each syndrome, incorporating ethnically diverse features.
Results
The model achieved a top-1 classification accuracy of 71.3% (95% CI: 61.8–79.6) and a top-3 accuracy of 93.5%. Synthetic facial representations were successfully generated and validated for each syndrome, providing an innovative resource for training and database augmentation, with the option of adjustable ethnic variability.
Conclusion
We present a functional AI-based mobile application capable of automatically analysing facial phenotypes from standard 2D images. By integrating robust syndrome classification with synthetic image generation, AIDY offers both a clinical decision support tool and a powerful educational resource in the field of dysmorphology. Its potential for global deployment opens new perspectives for the diagnosis of rare diseases, particularly in resource-limited healthcare settings.
Quentin Hennocq, Olivier Lienhard, Ludovic Bénichou, Luan Breton, Valérie Cormier-Daire, Antoine Ferry, Ana Julia Hidalgo Bravo, Stanislas Lyonnet, Arnaud Picard, Marlène Rio, Ahmed Zaiter, Nicolas Garcelon, Roman H. Khonsari
Necker–Enfants Malades Hospital, AP-HP
Imagine Institute
Characteristic craniofacial features are present in 30 to 40% of rare diseases, of which over 7,000 have been identified globally. Their recognition generally relies on the subjective expertise of clinical dysmorphologists. To democratise and broaden access to facial dysmorphology diagnostics, we developed AIDY, a novel mobile application that leverages artificial intelligence (AI) for the automated recognition of genetic syndromes using standard 2D facial photographs.
Methods
Our dataset was derived from the photographic archives of the Maxillofacial Surgery, Plastic Surgery, and Medical Genetics departments at Necker–Enfants Malades Hospital (AP-HP), comprising over one million images from more than 30,000 patients seen since 1976. The dataset was divided into 80% for training, 10% for validation, and 10% for testing. A deep learning model (ArcFace R-100, pre-trained on Glint360K) was fine-tuned on this dataset, which included over 500 genetically confirmed syndromes as well as non-syndromic controls. Model performance was evaluated using the area under the curve (AUC), along with top-1 and top-3 classification accuracy. In parallel, we employed a probabilistic diffusion model to generate representative synthetic facial composites for each syndrome, incorporating ethnically diverse features.
Results
The model achieved a top-1 classification accuracy of 71.3% (95% CI: 61.8–79.6) and a top-3 accuracy of 93.5%. Synthetic facial representations were successfully generated and validated for each syndrome, providing an innovative resource for training and database augmentation, with the option of adjustable ethnic variability.
Conclusion
We present a functional AI-based mobile application capable of automatically analysing facial phenotypes from standard 2D images. By integrating robust syndrome classification with synthetic image generation, AIDY offers both a clinical decision support tool and a powerful educational resource in the field of dysmorphology. Its potential for global deployment opens new perspectives for the diagnosis of rare diseases, particularly in resource-limited healthcare settings.
Quentin Hennocq, Olivier Lienhard, Ludovic Bénichou, Luan Breton, Valérie Cormier-Daire, Antoine Ferry, Ana Julia Hidalgo Bravo, Stanislas Lyonnet, Arnaud Picard, Marlène Rio, Ahmed Zaiter, Nicolas Garcelon, Roman H. Khonsari
Necker–Enfants Malades Hospital, AP-HP
Imagine Institute