Machine Learning Techniques for High-dimensional Neuroimaging Data Analysis

No Thumbnail Available
Yang, Xin
Journal Title
Journal ISSN
Volume Title
Middle Tennessee State University
In the past two decades, neuroimaging has become the most commonly used imaging technique for the study of human brain, which has given us insights about the complex neural characteristics of the human brain and also provided helpful information for the diagnosis of various diseases.
However, the analysis of neuroimaing data is extremely complex, requiring the use of sophisticated techniques from acquiring raw data to image processing and statistical analysis.The purpose of this dissertation is to provide accurate and efficient machine learning models for neuroimaging data analysis.
In this dissertation, we will focus on the study of two neuroimaging techniques: functional MRI data and MRI data.
Functional magnetic resonance imaging (fMRI) has become one of the most widely used techniques in investigating human brain function over the past two decades. However, the analysis of fMRI data is extremely complex due to its difficulties in big data processing. Hence, efficient and accurate machine learning models are necessary to interpret fMRI data by incorporating both spatial and temporal information. We will investigate a class of spatial multitask learning models which incorporates spatial information of each task's 2-dimensional neighborhood. Simulation and real application results show satisfactory performance from spatial multitask learning algorithms.
As Magnetic Resonance Imaging (MRI) has matured, a large number of researchers have studied Alzheimer's disease (AD) image data.
Many high-dimensional classification methods use structural MRI brain images for classification between AD and healthy individuals.
As computer computation power has improved, neural networks have been widely applied in Alzheimer's disease diagnosis. However, the first layer of this method is based on individual brain voxel, which means neural networks learn each voxel individually without considering the brain spatial information.
This method may lose some important information since the neighbor effect is ignored. Because the voxel of the brain is not isolated, in reality some brain area has an extremely close relationship.To overcome the shortcomings of the spatial correlation problem, we proposed a new technique called spatial regularization neural network (SRNN), which incorporates spatial information provided by each voxel's 3-dimensional neighbor voxels. It is successfully applied in real applications.
Machine Learning, Neuroimaging