Date of Award


Degree Name

Doctor of Philosophy


Computer Science

First Advisor

Dr. Fahad Saeed

Second Advisor

Dr. Alvis C. Fong

Third Advisor

Dr. Kevin H. Lee

Fourth Advisor

Dr. Ajay Gupta


fMRI, machine learning, deep learning, high performance computing, Autism Spectrum Disorder, ADHD


Brain disorders are very difficult to diagnose for reasons such as overlapping nature of symptoms, individual differences in brain structure, lack of medical tests and unknown causes of some disorders. The current psychiatric diagnostic process is based on behavioral observation and may be prone to misdiagnosis.

Noninvasive brain imaging technologies such as Magnetic Resonance Imaging (MRI) and functional Magnetic Resonance Imaging (fMRI) make the process of understanding the structure and function of the brain easier. Quantitative analysis of brain imaging data using machine learning and data mining techniques can be advantageous not only to increase the accuracy of brain disorder diagnosis but also to unravel unknown facts about the complex function of the brain. Research studies have shown functional connectivities of brain contain discriminative patterns that are widely used in a variety of studies such as fMRI classification.

In this dissertation, we designed machine learning and deep learning models for diagnosing brain disorders such as ADHD and ASD using fMRI data. In order to reduce the risk of overfitting in deep learning methods, we proposed a data augmentation approach for generating artificial samples from available data. Our models are able to improve the accuracy of classifying healthy samples from patients up to 28%. comparing to state-of-the-art solutions.

Analysis of fMRI data considering a huge number of voxels (smallest addressable element of fMRI data) is very time-consuming. One example is computing pairwise functional connections between voxels using measures like Pearson’s correlation. To tackle this issue, we designed two GPU based frameworks based on matrix multiplication for computing pairwise correlations that deliver around 3 times speedup against state-of-the-art GPU based methods. We expanded these frameworks to compute dynamic functional connectivity which involves computing multiple sets of pairwise correlations each associated with specific time windows in original time series followed by designing two methodologies for reducing the space requirements of pairwise correlations.

Access Setting

Dissertation-Open Access

Included in

Data Science Commons