ASD is a neurological and developmental disorder that affects how people interact with others, communicate, learn, and behave. According to the Centers for Disease Control and Prevention (CDC), 1 in 36 children and 1 in 45 adults are affected by autism in the U.S. in 2023. Currently, diagnosing ASD is difficult because ASD is diagnosed mainly by observing a patient's behavior and developmental history, as there is no automated or pathological solution established for diagnosis. There is a need to develop automated objective ASD diagnosis tools for quickly and accurately diagnosing ASD.

This study aims to develop an approach to automated autism spectrum disorder (ASD) detection using resting-state fMRI imaging with resting-state functional connectivity (RSFC) patterns learned with convolutional neural networks (CNN).

We proposed an approach to ASD detection using resting-state fMRI with RSFC patterns learned with CNN. RSFC measures the spontaneous low-frequency fluctuations between regions in the brain in fMRI data. We investigated the architecture of CNN for ASD detection based on RSFC patterns. The resting-state fMRI data from 895 individuals from the autism brain imaging data exchange (ABIDE) database was used in the study. The imaging data was acquired from 15 acquisition sites across the U.S. and consists of 556 healthy controls and 339 subjects diagnosed with ASD.

The proposed approach was tested on the data set by 10-fold cross validation. The CNN with a fixed architecture achieved high accuracy (M=0.999, SD=0.003), precision (M=1.0, SD=0.0), sensitivity (M=0.997, SD=0.009), and specificity (M=1.0, SD=0.0).

The results indicate that ASD detection models learned from RSFC data with CNN can detect ASD effectively in the resting-state fMRI imaging data. The proposed approach would be useful for developing objective ASD diagnosis tools and ASD research.