Fine-Tuned Hybrid EfficientNetB2–ViT-B16 Architecture Model on fMRI Brain Data for ASD Classification

Abstract

Autism, or autism spectrum disorder (ASD), is a neurodevelopmental condition characterized by atypical development of neurons in the brain. Diagnosing ASD remains a major challenge due to the complexity and heterogeneity of the symptoms associated with this disorder. Early detection is essential, as rapid treatment significantly improves patients’ quality of life. However, traditional diagnostic methods are often subjective, time-consuming, and resource-intensive. To defeat these limitations, this study proposes an improved deep learning-based framework for the early detection of ASD from functional magnetic resonance imaging (fMRI) data. For this study, a pretrained Convolutional Neural Network (CNN) model EfficientNetB2 and a pretrained Vision Transformer (ViT) model ViT-B16 have been fine-tuned and then get combined to an hybrid model. All these models will be evaluated its performance on the pre-processed ABIDE dataset. Data augmentation techniques were also implemented to overcome the limitations of the dataset and improve the model’s generalization ability. The experimental results show that the fine-tuned EfficientNetB2 model achieves 96% accuracy, while the fine-tuned ViT-B16 model realises 95% and the hybrid model reaches 97%. These results were promising, highlight the potential of transfer learning models optimized for clinical applications, demonstrating their effectiveness to aid in the early and accurate diagnosis of ASD. This study thus contributes to the development of diagnostic systems compatible with the state of the art, paving the way for better management of neurodevelopmental disorders.