Artificial Intelligence Based Methods for Autism Spectrum Disorder Diagnosis from fMRI Data

Abstract

Brain and cognition diseases such as Autism are common types of diseases that are caused by brain disorder or dysfunctions. Earlier detection of brain diseases can call for proper medical intervention and treatment and prevent irreversible brain tissue damages. Medical imaging is one of the tools to assist clinicians in brain disorder diagnosis. In modern neuroscience and clinical study, neuroscientists and clinicians often use non-invasive imaging techniques to validate theories and computational models, observe brain activities and diagnose brain disorders. Among many medical imaging modalities, magnetic resonance imaging (MRI) becomes the mainstream diagnosis mechanisms for brain diseases due to its noninvasive nature. MRI is increasingly being used in the diagnosis of brain disorders such as autism, and is becoming the imaging choice in evaluation of brain and cognition disease. With more MRI datasets becoming available and the advent of deep learning enabled artificial intelligence, it is becoming widespread and efficient to apply artificial intelligence techniques for brain disorder diagnosis. However it still remains a major challenge to utilize deep learning models to classify fMRI data accurately in order to provide useful assistance for medical practitioners. This thesis attempts to address Autism disease diagnoses by applying cutting-edge deep learning algorithms to functional MRI data. After introducing relevant background information, objectives of the study, and the structure of the thesis, the thesis starts with a comprehensive review of recent developments of the prevalent methods for brain disorder diagnosis using fMRI data. It categorizes the existing fMRI interpretation for classification and the representative corresponding methods, analyzed the major challenges of the problem and discussed the advantages and disadvantages of these methods, and points out some future research directions. To follow the summarized research directions and address the thesis objectives, the thesis next presents an autoencoder based two-staged Autism diagnosis algorithm by incorporating the graph-theoretic enabled features. The proposed deep neural networks based model utilizes graph centralities derived features in order to improve the classification accuracy. Then, an autoencoder based semi-supervised framework is presented for brain disorder diagnosis using fMRI derived functional connectivity as input features. The proposed semi-supervised model jointly optimizes the autoencoder construction error and the supervised classification loss simultaneously, and tuned the autoencoder to learn the hidden representations towards the ultimate classification goal. The proposed semi-supervised framework can also integrate unlabelled data into the training process and achieve improved classification performance. Inspired by the success of geometric deep learning for handling data with underlying structure being non-Euclidean, a graph attention network model is presented to utilize the inherent network characteristics of the brain functional networks and temporal patterns of fMRI data to learn the non-Euclidean topological information. Brain functional networks derived from fMRI based functional connectivity are examined from a graph theory point view in order to incorporate important topological information to the brain disorder diagnosis. Besides, temporal features of the fMRI data are combined with the graph centralities for improved performance. To summarize the work in the thesis, a series of deep learning enabled artificial intelligence methods for brain disorder diagnosis are proposed from fMRI data. These methods target different aspects of fMRI data, and utilize different facets of fMRI data and/or its combinations to devise different learning algorithms for brain disorder diagnosis. Experimental results show the proposed methods are either better than the existing competing methods, or fill the gaps that currently available methods have not yet addressed.