Robust Facial Classification of Down Syndrome using Lightweight CNNs

Abstract

Down Syndrome (DS) is a genetic disorder caused by trisomy 21 and is characterized by distinctive facial features that can support early screening. However, access to conventional diagnostic tools remains limited, particularly in resource-constrained regions. This study presents a comparative evaluation of two lightweight convolutional neural network (CNN) architectures, EfficientNet-B1 and MobileNetV3-Large, for facial image-based DS classification. A curated dataset of 3,030 facial images underwent quality control and image enhancement processes applied exclusively to the training data, resulting in 2,620 images. The dataset was split into training, validation, and test sets at a 70:20:10 ratio. Both models were fine-tuned using ImageNet-pretrained weights and evaluated based on accuracy, precision, recall, and F1-score. Performance robustness and statistical significance between models were assessed using five-fold cross-validation and one-way ANOVA. The experimental results demonstrate that both architectures achieved high classification performance; however, EfficientNet-B1 exhibited superior stability, more balanced class predictions, and lower fold-to-fold variability. Furthermore, Grad-CAM visualization confirmed that both models focused on clinically relevant facial regions, with EfficientNet-B1 showing more consistent and interpretable attention patterns. These findings suggest that EfficientNet-B1 is a robust and interpretable model for facial-based DS screening, offering significant potential for deployment in resource-limited healthcare settings.