Deep learning of human brain activations
Functional Magnetic Resonance Imaging (fMRI) is one of the leading methods for analyzing brain functions as it is non-invasive and has a high spatial resolution of the brain. FMRI recognizes activations in the brain through Blood Oxygen Level Dependent (BOLD) responses. Common methods of brain analy...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
2018
|
Subjects: | |
Online Access: | http://hdl.handle.net/10356/76151 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Functional Magnetic Resonance Imaging (fMRI) is one of the leading methods for analyzing brain functions as it is non-invasive and has a high spatial resolution of the brain. FMRI recognizes activations in the brain through Blood Oxygen Level Dependent (BOLD) responses. Common methods of brain analysis with fMRI use activations of individual regions of the brain to identify tasks. This is useful information if a human were to analyze the results visually. However, machine-learning models are able to learn complex relations between input features to classify multiple brain states. In our study, we consider 8 brain states – Emotion, Language, Gambling, Motor, Relational, Social, Working Memory and the Resting state.
As opposed to considering brain activation in individual regions to recognize a state, we propose an alternative method, which considers the correlation of functional activation between a pair of regions as the features to recognize the brain state. When applied to a Deep Neural Network (DNN) based encoder with fMRI data gathered from large-scale fMRI database, we achieved a far better accuracy than previous works. Using a technique called Principal Sensitivity Analysis on the trained DNN, we identified subject-independent features or brain regions that were crucial to the differentiation of various tasks. Furthermore, it highlighted the potential of our decoder and method of data representation as it was also able to illustrate the attenuation of the default mode network in resting state versus task-focused states. |
---|