Deep Learning-Powered Down Syndrome Detection Using Facial Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Genetics

Background:

• Down syndrome (DS) is a common chromosomal disorder with characteristic features and health issues.
• Current genetic testing for DS is limited by cost and expertise, restricting access in underserved areas.
• There is a critical need for accessible, non-invasive screening methods for early Down syndrome detection.

Purpose of the Study:

• To develop and validate a deep learning model for early Down syndrome detection using infant facial images.
• To create an interpretable and robust AI tool for equitable DS screening.
• To establish a foundation for scalable digital health solutions in pediatric care.

Main Methods:

• A hybrid deep learning architecture combining RegNet X-MobileNet V3 and vision transformer (ViT)-Linformer for feature extraction.
• Adaptive attention-based feature fusion to focus on diagnostically relevant facial areas.
• Bayesian optimization with hyperband (BOHB) fine-tuned extremely randomized trees (ExtraTrees) for classification, with stratified five-fold cross-validation.

Main Results:

• The model achieved high performance on unseen data: 99.10% accuracy, 98.80% precision, 98.87% recall, 98.83% F1-score, and 98.81% specificity.
• The hybrid feature extraction and attention fusion effectively represented diagnostic facial features.
• The model demonstrated superior performance compared to existing DS classification methods.

Conclusions:

• The developed deep learning model is a reliable and accurate tool for early Down syndrome screening via facial imaging.
• This AI-driven approach enhances diagnostic accessibility, particularly in resource-limited settings.
• The study supports the development of trustworthy digital solutions for global pediatric healthcare.