rs-fMRI analysis using spatio-temporal sparse convolutional neural networks

Abstract

Neuropsychiatric diseases such as Autism Spectrum Disorder (ASD) and Schizophrenia cause various behavioral and communication dysfunctions in human life. Resting state functional magnetic resonance imaging (rs-fMRI) is used to detect and characterize functional changes in the brain associated with these disorders. Machine learning methods are known to perform well in classifying fMRI images and have proven to have great potential in the field of computer aided diagnosis. In most of the previous studies, hand-crafted features have been used in fMRI analyzes and classifications to date. This prevents the system from being end-to-end and causes spatial or temporal information to be lost due to dimension reduction. The method presented in this study works end-to-end as well as being fed with an entire 4-dimensional fMRI sequence. It is faster than traditional convolutions and recurrent neural networks of the same size, thanks to the sparse convolutional layers that are the building blocks of the network. Experiments with schizophrenia and ASD fMRIs have shown similar performance to those in the literature, despite limited resources.