An Exploration of Autism Spectrum Disorder Classification from Structural and Functional MRI Images

Abstract

There are strong indications that structural and functional magnetic resonance imaging (MRI) may help identify biologically relevant phenotypes of neurodevelopmental disorders such as Autism spectrum disorder (ASD). Extracting patterns from MRI data is challenging due to the high dimensionality, limited cardinality and data heterogeneity. In this paper, we explore structural and resting state functional MRI (rs-fMRI) for ASD classification using available ABIDE II dataset, using several standard machine learning (ML) models and convolutional neural networks (CNNs). To overcome the high dimensionality problem, we propose a simple data transformation method based on histograms calculation for the standard ML models and a simple 3D-to-2D and 4D-to-3D data transformation method for the CNNs in ASD classification. Numerous research has been done for ASD classification using the online available ABIDE I dataset, and several with the ABIDE II dataset, the latter mostly working with single-site classification studies. Here, we take the whole ABIDE II dataset using all structural and functional raw data. Our results show that the proposed methods achive state-of-the art results of high 71.4% accuracy (functional) and 73.4% AUC (structural) compared to the best performing results in literature of 68% accuracy (functional) for ASD classification on all ABIDE dataset.