Autism Spectrum Disorder Deep Learning Approaches: A Comparative Study

Autism Spectrum Disorder (ASD) is a neurodevelopment disorder that is complicated and associated with social, behavioral, and communication issues. It should also be diagnosed early enough so that the appropriate treatment is administered in time and better results are realized. The recent tendencies in the field of deep learning (DL) allowed to implement neuroimaging, in the present case, the magnetic resonance imaging (MRI) to ASD diagnosis, which can be automated, to the given field. The publicly available ABIDE MRI data were used in this research and assisted in comparing some of the latest DL models, such as a blank Convolutional Neural Network (CNN), ResNet50, EfficientNet, and Vision Transformer (ViT). Data normalization, skull stripping, and data augmentation were used as preprocessing. The models were trained using Adam optimizer and categorical cross-entropy loss and assessed according to accuracy, precision, recall, F1-score and AUC-ROC. The best and highest performance one was the Vision Transformer (ViT) that achieved the highest accuracy of 97.1% and 0.99 AUC- ROC, which shows the superiority in the ASD detection in the ABIDE MRI data. These results favor the fact that the transformer-based models are powerful diagnostic tools of ASD detection.