A Benchmark of Preprocessing Strategies for Autism Classification from Resting-State Functional Magnetic Resonance Imaging

Currently, the diagnosis of autism spectrum disorder (ASD) is a subjective behaviorally-based process, which requires one or more professionals. In recent years, strategies based on the automated analysis of resting-state functional magnetic resonance imaging (rs-fMRI) have become a promising candidate that could reduce subjectivity in ASD diagnosis. Unfortunately, the outcome of computerized rs-fMRI analysis is highly dependent on preprocessing stages. To date, there is no consensus regarding which preprocessing methods are more suitable according to the techniques used in the analysis of the images. For this reason, in this work we analyze the performance of different machine learning algorithms when classifying ASD patients vs typically developing patients using the four preprocessing strategies provided by Configurable Pipeline for the Analysis of Connectomes. Specifically, this study is based on rs-fMRI data from the Autism Brain Imaging Data Exchange I dataset. The objective is to evaluate the impact in the performance on a given machine learning model when using different preprocessing strategies on rs-fMRI sequences. Experiments with a dataset of 175 rs-fMRI sequences show that the best results are obtained using the strategy of global signal regression with support vector machines, achieving an accuracy and a $F_{1}$ -score of 69.7% and 60.5%, respectively.