Implementing EfficientNetB0 for Facial Recognition in Children with Down Syndrome

Abstract

Early detection of Down Syndrome in children is crucial to provide more appropriate medical and educational interventions. This study aims to build and evaluate a deep learning-based classification model using the EfficientNetB0 architecture to distinguish facial images of children with Down Syndrome and healthy children. The dataset used consists of two classes (Down Syndrome and healthy), which have gone through an augmentation process to increase data diversity and prevent overfitting. The model was trained using the Adam algorithm with a learning rate of 0.0001 and a sparse categorical crossentropy loss function for 10 epochs. The training results showed that the model achieved a validation accuracy of 93.94%, with the lowest validation loss value of 0.2390. Further evaluation was carried out using a confusion matrix, which showed that the model was able to properly classify 312 out of 333 Down Syndrome images and 309 out of 330 healthy children images, resulting in an overall accuracy of 94%. In addition, the precision, recall, and f1-score values ​​for both classes were in the range of 0.94, indicating a balanced and strong performance. Visual analysis of the misclassified images indicates that some misclassifications occur on healthy children’s faces with certain expressions, angles, or lighting conditions that resemble Down syndrome. Conversely, some children with Down syndrome are also predicted as healthy when their facial features are not too prominent or similar to normal children under certain lighting conditions. This shows that despite the high performance of the model, sensitivity to facial feature variations remains a challenge.