Deep Learning-based Classification of Autism Spectrum Disorder Using Resting State fMRI Data

Autism spectrum disorder (ASD) is a neurodevelopmental disorder which impacts individuals in various ways. It is distinguished by confined and repetitive behaviors as well as difficulties in interaction and social communication. Neuroimaging techniques such as resting-state functional magnetic resonance imaging (rs-fMRI) have impacted our knowledge of the brain’s function by enabling researchers to view brain activity with respect to time. These technologies have offered useful information on how brain regions are engaged in numerous cognitive processes. Functional connectivity (FC) features extracted from the rs-fMRI scans are widely used for classifying individuals with ASD using deep learning methods. However, developing an effective approach to increase ASD classification accuracy from typically developing controls remains an important challenge. Therefore, we introduced a new deep learning architecture for ASD classification. The architecture is composed of four main steps: 1. Convolutional-max pooling 2. First fully connected layers 3. Concatenating 4. Second fully connected layers. We have evaluated the architecture for classification using rs-fMRI data from the publicly available dataset Autism Brain Imaging Data Exchange I (ABIDE I) with a 10-fold cross-validation method. We utilized the Pearson correlation coefficient to calculate the FC matrices, which served as the input for the architecture. Then, the proposed architecture classifies the subjects with ASD. We achieved an average classification accuracy of 72.46%, which outperformed the existing methods in ASD classification.