Autism Spectrum Disorder Detection Based on Wavelet Transform of BOLD fMRI Signals Using Pre-trained Convolution Neural Network

Autism spectrum disorder (ASD) is a mental and the main problem in ASD treatment has no definite cure, one possible option to control its symptoms. Conventional assessment using questionnaires may not be accurate required evaluation of trained experts. Several attempts use resting-state functional magnetic resonance imaging (fMRI) as an assisting tool combined with classifier have been reported for detection. Still, researchers barely reach accuracy 70% replicated models independent datasets. Most studies used connectivity structural measurements ignored temporal dynamics features fMRI data analysis. This study aims present several convolutional neural networks tools detection based on dynamic classification improve prediction results. The sample size 82 subjects (41 41 normal cases) collected from three different sites Brain Imaging Data Exchange (ABIDE). default mode network (DMN) regions are selected blood-oxygen-level-dependent (BOLD) signals extraction. extracted BOLD signals' time-frequency components converted scalogram images input pre-trained feature extraction such GoogLenet, DenseNet201, ResNet18, ResNet101. two classifiers: support vector machine (SVM) K-nearest neighbours (KNN). best results 85.9% achieved by DenseNet201 classified these KNN classifier. Comparison previous studies, indicated good potential proposed model diagnosis cases. From another perspective, presented method can applied analysis rs-fMRI other type brain disorders.