An Improved dynamic functional connectivity and deep neural network model for Autism Spectrum Disorder Classification

Abstract

Brain disorders such as autism spectrum disorder (ASD) is still difficult to diagnose. In the recent years, different novel deep learning algorithms have been applied to detect ASD. Most studies use the functional connectivity (FC) pattern to represent the brain activities. However, it has been investigated that dynamic functional connectivity (dFC) which represent more features than FC can characterize the intrinsic brain organization changes over time. The goal of this paper is to determine that dFC features are more successful than FC features in the classification of ASD using deep learning. In this paper, we propose a classification model using dFC and deep neural network. Firstly, we used windowed k-means (WKM) approach to compute the sub-state of the brain and extract the main features of the functional magnetic resonance imaging(fMRI). Then, two stacked denoising autoencoders were applied to extract the features and reduce the dimension. At last, the MLP was utilized to complete the classification task and do fine-tuning based on the autoencoder encoders weights. The experiments were carried out on the Autism Brain Imaging Data Exchange (ABIDE) datasets. Result shows that we acquired a mean accuracy of 68.51%. Overall, our proposed classification is effective and provide evidence that dFC contains more brain states features.