Comparison of VGG16 and VGG19 Models in the Classification of Down Syndrome in the European Region with Transfer Learning

Abstract

Down syndrome detection by utilizing facial images as the main data has been widely developed through deep learning approaches, especially Convulutional Neural Network (CNN). However, most studies only classify the disorder without paying attention to regional factors. This has limited the effectiveness of the model in the classification of Down syndrome, especially in populations in European regions that have different morphological characteristics. This study examines the performance of two pretrained CNN models, namely VGG16 and VGG19, in classifying facial images of children from Europe who are divided into 2 categories of Down Syndrome and Healthy. The dataset used in the study consists of 1,543 images from the Down syndrome class 671 images and the Healthy class 872 images. It was then expanded to 1570 images to balance the data between both Down syndrome and Healthy classes. The evaluation results of this research by applying augmentation show that the VGG16 model has superior performance compared to VGG19, with accuracy reaching 94%. Meanwhile, the VGG19 model obtained an accuracy of 90%. This difference shows that the VGG16 model has a more stable performance in detecting both categories with a better balance between precision and recall. This research is limited to European children's image data and still does not exist for ethnic teenagers or the elderly. This provides a basis for the development of facial image-based early detection systems, particularly for clinical applications or early screening in areas with similar populations.