Autism Spectrum Disorder Diagnosis From fMRI Images Using Deep Learning Techniques

Abstract

Heterogeneous mental disorders such as Autism Spectrum Disorder (ASD) are notoriously difficult to diagnose, especially in children. The current psychiatric diagnosis process is based solely on the observation of behavioral symptoms (DSM-5 / ICD-10). These qualitative methods can lead to mistakes in diagnosis. To have a slightly better and more accurate diagnosis, we need to find advanced and scalable methods of machine learning that allow us to more accurately distinguish healthy people from people with mental disorders. To do this, in this thesis we used resting-state fMRI data from ABIDE1 database, which includes 17 imaging sites. Most of the previous works done on this database have focused on one or a limited number of sites. But in this research, all sites have been used to increase the generalizability of the work. After performing some preprocessings such as slice time correction, motion correction, noise signal removal according to the CPAC method.Then, using the Atlases, the average time series are extracted from different ROIs of the brain and we defined a correlation matrix for each data. Then, by using Chi-Square feature selection method, the most important features are identified and only these features are used for classification. A novel architecture of convolutional neural networks with two-dimensional convolutional layers is used to analyze and classify data. To evaluate the model, three sets of experiments were designed based on different data sets and atlases used. Using this method, the highest accuracy obtained in the experiments was 73/53%, which was higher than previous works for classifying autistic people and healthy people.