Application of Deep Learning in Resting State Functional Magnetic Resonance Imaging Data Analysis

Abstract

Since the advent of deep learning, it has been used in a lot of fields like computer version, image recognition and speech recognition, etc. It has been made many achievements due to its outstanding ability on both classification and regression tasks. In this dissertation, we explore the use of both existing and newly designed deep learning techniques in resting state functional magnetic resonance imaging data analysis.

In the resting state, the human brain is still active and different regions are functionally connected to each other through intrinsic networks called Resting State Networks (RSNs). It is important to locate the RSNs and understand how these RSNs function between each other. Especially, in the clinical field, this will help provide better and more precise treatment plans for some mental disorder diseases. In this dissertation the Graph Sample and aggregate (graphSAGE), a typical graph neural network, is proposed for extracting RSNs. It treats each resting state functional magnetic imaging as a graph while analyzing the data. Compared with the classical methods such as seed-based method, independent component method, and dictionary learning method, the application of grahSAGE gives robust results. The classical methods need some user-defined prior values, such as thresholds of p values for both single subject analysis and group level analysis, and the number of components for independent component analysis (ICA) and dictionary learning analysis. The results depend on these prior assumptions and therefore subjective more or less. On contrast, graphSAGE does not need prior assumptions. The results are more objective and robust. Moreover, it can perform single subject analysis and group subject analysis simultaneously.

Attention Deficit Hyperactivity Disorder (ADHD) is a type of mental health disorder. It is a disease that can be seen from children to adults and affects the patient's normal life. Therefore, the accurate diagnosis as early as possible is very important for the treatment of the patient in clinical applications. Some traditional classification methods, although having been shown powerful in many other classification tasks, are not as successful in the application of ADHD classification. In this dissertation, we designed two novel deep learning approaches, called ICA-CNN method and correlation autoencoder method, respectively, for the ADHD classification task.

The ICA-CNN method first extracts independent components from each subject. These independent components are then fed into a convolutional neural network as input features to classify the ADHD patient from typical controls.

The correlation autoencoder method calculates the correlation between regions of interest of the brain, which is then the input of an autoencoder to learn the latent features. After the latent features are learned, they are used in the classification task by a new neural network. Both methods significantly outperform the classical methods such as logistic regression, support vector machines, and other methods used in previous studies.