Detection of Autism Spectrum Disorder using fMRI Functional Connectivity with Feature Selection and Deep Learning

Autism spectrum disorder (ASD) is notoriously difficult to diagnose despite having a high prevalence. Existing studies have shifted toward using neuroimaging data to enhance the clinical applicability and the effectiveness of the diagnostic results. However, the time and financial resources required to scan neuroimages restrict the scale of the datasets and further weaken the generalization ability of the statistical results. Furthermore, multi-site datasets collected by multiple worldwide institutions make it difficult to apply machine learning methods due to their heterogeneity. We propose a deep learning approach combined with the F-score feature selection method for ASD diagnosis using a functional magnetic resonance imaging (fMRI) dataset. The proposed method is evaluated on the worldwide fMRI dataset, known as ABIDE (Autism Brain Imaging Data Exchange). The fMRI functional connectivity features selected using our method can achieve an average accuracy of 64.53% on intra-site datasets and an accuracy of 70.9% on the whole ABIDE dataset. Moreover, based on the selected features, the network topology analysis showed a significant decrease in the path length and the cluster coefficient in ASD, indicating a loss of small-world architecture to a random network. The altered brain network may provide insight into the underlying pathology of ASD, and the functional connectivity features selected by our method may serve as biomarkers.