A Transfer Learning Approach for Down Syndrome Detection Using Facial Image Features

Early detection of Down Syndrome (DS) is crucial for timely medical intervention, but current methods like karyotyping are invasive, costly, and time-consuming. This study proposes a non-invasive, transfer learning-based approach using facial image analysis for DS diagnosis. The proposed VNL-Net integrates VGG16 for feature extraction, Non-Negative Matrix Factorization for dimensionality reduction, and LightGBM for feature generation. K-fold cross-validation with Logistic Regression ensures reliable classification. A MobileNet + SVM hybrid model is also developed for deployment on mobile and edge devices. Using publicly available Kaggle datasets, the system achieves superior diagnostic accuracy compared to traditional methods. This cost-effective solution demonstrates the potential of deep learning in medical image analysis and aims to expand with explainable AI for greater clinical adoption. Keywords: Down Syndrome, Facial Recognition, Transfer Learning, VGG16, Non-Negative Matrix Factorization, LightGBM, Logistic Regression, MobileNet, SVM.