Klasifikasi Autisme dalam Implementasi Transfer Learning Menggunakan Model EfficientNet-B3

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that affects communication, social interaction, and behavior, and is often difficult to diagnose at an early stage. Conventional diagnosis still relies heavily on clinical observation, which may lead to potential errors. With the advancement of deep learning technology, early detection of autism can be carried out through facial image analysis. This study proposes the application of transfer learning using the EfficientNet-B3 architecture to classify images of autistic and non-autistic children. The dataset used consists of 2,940 facial images (1,470 autism and 1,470 non-autism) obtained from the Kaggle platform. The preprocessing stage includes image resizing, pixel normalization, and data augmentation to prevent overfitting and improve the model’s generalization. The EfficientNet-B3 model was modified by adding fully connected layers, batch normalization, and dropout layers to meet the requirements of binary classification.
The evaluation results show that EfficientNet-B3 achieved an accuracy of 92%, which is higher than the previous study using VGG19 with an accuracy of 84%. In addition to accuracy, the precision, recall, and F1-score of EfficientNet-B3 also indicate a more balanced performance in distinguishing autism and non-autism images. Thus, this study demonstrates that EfficientNet-B3 can be an effective approach to support early autism detection through digital image classification and has great potential for further development in clinical applications and diagnostic support systems.