Early autism analysis and diagnosis system using task-based fMRI in a response to speech task

Abstract Autism spectrum disorder (ASD) is a neurodegenerative disorder characterized with lingual and social disabilities. Difficulties in language comprehension and delayed development in individuals with ASD follow a wide spectrum. Clinical judgment and the autism diagnostic observation schedule (ADOS) are the current gold standard for ASD diagnosis and severity grading. Developing objective computer-aided technologies for ASD diagnosis with the utilization of brain imaging modalities and machine learning is one of the main tracks in the current studies to understand autism. Task-based functional magnetic resonance imaging (fMRI) demonstrates the functional activation in the brain by measuring blood-oxygen-level-dependent variations in response to certain tasks. It is believed to hold discriminant features for autism. Computer-aided diagnosis systems have emerged to classify autistic subjects against typically developed peers with the deployment of task-based fMRI. A novel computer-aided grading framework is proposed to grade infants and toddlers guided by brain activation analysis. The proposed study includes 157 subjects, with ages ranging between 12 and 40 months, who underwent a response to speech task. The ADOS-calibrated severity score is selected to be the ground truth for dividing subjects into three groups: mild, moderate, and severe autistic containing 92, 32, and 33 subjects, respectively. Following data preprocessing and general linear modeling, a two-phase framework is designed. The first phase performs statistical brain activation analysis over each group and all combined. It is proven that an increased hypoactivation of the angular gyrus, superior temporal cortex, cingulate gyri, and primary auditory cortex is exhibited and increases with the increase of autism severity. Less lateralization is also present when activation of the left hemisphere regions is recorded. The second part performs machine learning autism grading system. Local and global feature extraction only from these region of interest brain areas are applied. We have developed a comprehensive two-stage classification pipeline. A comprehensive investigation of several feature vector and classifier pairs has been conducted to get the best performing classifier and discriminant features. The best performing classifier is the random forest. Fourfold cross-validation is applied for testing. The first stage classifies between moderate and the other two groups with an average accuracy of 83% (sensitivity = 73%, specificity = 83%). Following this stage the second stage classifies subjects as mild or severe ASD with an average accuracy of 81% (sensitivity 81% and specificity 76%).