Deep Learning-Based Diagnosis of ASD Using Pre-Trained Convolutional Neural Network

Autism Spectrum Disorder (ASD) is a developmental condition affecting the brain, characterized by challenges in social interaction, communication abilities, and repetitive behaviors. Diagnosing ASD is complex due to its varied manifestations in behavior, onset, response to treatment, and coexisting conditions. To address this, our study introduces a methodological approach to diagnose ASD using structural magnetic resonance imaging (sMRI) data and deep learning algorithms. For our analysis, we obtained the necessary dataset from publicly accessible ABIDE-I and ABIDE-II databases. The sMRI data underwent pre-processing using a standardized pipeline, and we focused on the ±5 sagittal slices from the mid-sagittal plane. These slices were then utilized as input for the inception v3 deep learning model. The results of our experiment produced an average 10-fold cross-validation accuracy = 78.83%. Moreover, we achieved a sensitivity = 79.68%, specificity = 77.93%, precision = 79.47%, and f1-score = 79.44%. These metrics indicate the effectiveness of our proposed process pipeline in accurately classifying ASD-like neurodevelopmental disorders. In conclusion, our study provides a valuable methodology that may contribute to improved diagnostic classification of ASD and similar neurodevelopmental conditions using sMRI and advanced deep learning techniques.