Novel Transfer Learning Based Deep Features For Diagnosis Of Down Syndrome In Children Using Facial Images

The early and accurate diagnosis of Down syndrome in children is critical for effective intervention and support. This study presents a novel approach to Down syndrome diagnosis using facial images through advanced transfer learning techniques and deep feature extraction methods. We propose a multi-faceted approach that integrates VNL-Net and a MobileNet + SVM hybrid model to enhance diagnostic accuracy and computational efficiency.

Our primary methodology involves VNL-Net, which combines the VGG16 model for initial spatial feature extraction with Non-Negative Matrix Factorization (NMF) for dimensionality reduction and refined feature extraction. The extracted features are then further enhanced using the Light Gradient Boosting Machine (LGBM). This robust feature generation method is followed by classification using Logistic Regression, with the model's performance rigorously evaluated through k-fold cross-validation.

To extend our approach for practical deployment, especially on mobile and edge devices, we introduce a MobileNet + SVM hybrid model. MobileNet's efficient feature extraction capabilities are leveraged to process facial images, producing lightweight yet high-performance features. These features are then classified using a Support Vector Machine (SVM), aimed at distinguishing between Down syndrome and healthy children effectively.

Our proposed methods demonstrate improved accuracy in Down syndrome detection, leveraging the strengths of advanced transfer learning models and hybrid classification approaches. This research not only contributes to the field of automated medical diagnosis but also addresses the need for efficient, real-time solutions suitable for mobile and edge computing environments.

Keywords: Down syndrome, facial images, transfer learning, VNL-Net, VGG16, Non-Negative Matrix Factorization, Light Gradient Boosting Machine, MobileNet, Support Vector Machine, classification, logistic regression, feature extraction, deep learning.