Deep Learning-Based Detection of Facial Dysmorphic Features for Rare Genetic Disorders in Pediatric Populations

Background: Facial dysmorphism is a critical phenotypic manifestation of many rare genetic disorders, yet systematic and interpretable computational approaches for identifying abnormal facial appearances in pediatric populations remain limited. Methods: We developed a deep learning-based framework for automated screening of pediatric facial anomalies using 434 frontal facial images (235 from genetically confirmed patients, 199 healthy controls). Facial regions were detected using YOLOv8 and processed through standardization and augmentation. Transfer learning with pretrained ResNet18, ResNet34, and ResNet50 architectures was applied for binary classification. Model interpretability was assessed using Gradient-weighted Class Activation Mapping (Grad-CAM). Syndrome-specific models were constructed for Williams–Beuren syndrome (WBS) and Achondroplasia (Ach). Results: On the validation set (n = 132), ResNet18 and ResNet50 achieved perfect classification performance (accuracy, F1 score, AUC = 1.000), while ResNet34 achieved AUC = 0.999. Grad-CAM analysis revealed that ResNet18 consistently focused on clinically relevant craniofacial regions in abnormal cases, with minimal diffuse attention in normal samples. WBS and Ach-specific models achieved AUC = 0.98 with distinct syndrome-specific attention patterns. Conclusion: This framework enables accurate and interpretable identification of pediatric facial anomalies. The syndrome-specific attention patterns demonstrate that the model captures clinically meaningful facial representations, supporting its potential as a computational tool for first-tier screening in rare genetic disorders.