Classification of ASD Subtypes Based on Coherence Features of BOLD Resting-state fMRI Signals

Abstract

Resting-state brain functional connectivity (FC) patterns play an essential role in the development of autism spectrum disorder (ASD) classification models based on functional magnetic resonance imaging (fMRI) data. Due to the limited number of models in the literature for identifying ASD subtypes, a multiclass classification is introduced in this study. The aim of this study is to develop an ASD diagnosis model using convolutional neural networks (CNN) with dynamic FC as inputs. The rs-fMRI dataset used in this study consists of 35 individuals from multiple sites labeled based on autistic disorder subtypes (ASD, APD, and PDD-NOS) and normal control (NC). The Atlas for Automated Anatomical Labeling (AAL) is selected as the brain atlas for defining brain nodes. The BOLD signals of the nodes are extracted and then the dynamic FC between brain nodes is determined using our new metric wavelet coherence (WCF), where WCF quantifies the overall variability of coherence in specific low-frequency scales over the time. Based on the statistical analysis of WCF values between ASD and NC, 6 pairwise nodes are identified. Classification algorithm is developed using CNN, and wavelet coherence maps (scalogram) of pairwise nodes. The training and testing of the CNN is using a cross-validation framework. The results of the multiclass classification provided an average accuracy of 88.6%. The results of this study illustrate the good potential of the wavelet coherence technique in analysing dynamics FC and open up possibilities for its application in diagnostic models, not only for ASD but also for other neuropsychiatric disorders.