Development of a new computational model for mapping, learning and mining of 3D spatio-temporal fMRI data
| aut.embargo | No | en_NZ |
| aut.thirdpc.contains | No | en_NZ |
| aut.thirdpc.permission | No | en_NZ |
| aut.thirdpc.removed | No | en_NZ |
| dc.contributor.advisor | Kasabov, Nikola | |
| dc.contributor.advisor | Pears, Russel | |
| dc.contributor.author | Murli, Norhanifah | |
| dc.date.accessioned | 2016-04-06T02:41:33Z | |
| dc.date.available | 2016-04-06T02:41:33Z | |
| dc.date.copyright | 2015 | |
| dc.date.created | 2016 | |
| dc.date.issued | 2015 | |
| dc.date.updated | 2016-04-06T02:22:24Z | |
| dc.description.abstract | The application of data mining techniques, particularly classification of spatio-temporal 3D functional magnetic resonance images has received growing attention in the literature. Spatio or spatial component as well as temporal component are factors of high importance in determining and recognizing brain state in response to external stimuli. Structural and functional brain data have been hugely collected, in an attempt to improve brain cognitive abilities and processing capabilities, as well as advancement in medicine, health, education, Brain Computer Interface and games. A particular spatio- and spectro-temporal brain data (STBD), functional Magnetic Resonance Imaging (fMRI), provides a comprehensive detail of brain activation when a certain stimulus is presented to the subject resulting from the changes of oxygen level in the blood vessel of the brain. This oxygen difference between active neurons and inactive neurons is captured in sequence, and the images generated from this (the fMRI) are composed of tens of thousands of individual voxels. These massive voxels are the features to this thesis, which became one of the challenges that had to be faced, in addition to the complex format of the data itself. To some extent, conventional machine learning techniques has successfully process and classify fMRI data. However, these techniques are only best at dealing spatial data, which completely neglect the temporal information that this data has. Thus, this study proposes and presents a novel computational model that specifically process spatial and temporal information of fMRI data, which make use of the newly proposed NeuCube model as its foundation. The derived model, denoted as NeuCube<sup>B</sup> utilized the 3D evolving SNN architecture of NeuCube in mapping and learning the data. The model learns from the data; then creates and updates connections between the neurons based on their weights. These connections represent chains of neuronal activities which could be reproduced even when only part of the stimuli data is presented, therefore making the NeuCube connections as an associative memory. The model can be used not only to classify brain activation patterns, but also to determine functional trail from the data i.e. to identify brain areas that receive the most activation from the stimulus. There are two case studies presented in the thesis involving different set of fMRI data which are in different format. The dataset is used and experimented by many researchers, which utilized different types of conventional machine learning techniques. In NeuCube<sup>B</sup>, the fMRI features (voxels) are modelled and studied as both spatial and temporal information involving phases of data reading, mapping, and encoding, before they are transferred to initialization and unsupervised learning stage. Connectivity of neurons in the network could be visualized and studied. The visualization can reveal crucial spatio-temporal relationship unseen from the data that are completely ignored by the standard classifiers. For both experiments involving two different sets of fMRI data, NeuCube<sup>B</sup> model results in better classification accuracy as compared to the standard classifiers. From this result, it can be concluded that NeuCube<sup>B</sup> model is not vulnerable to noise, that normally reside in fMRI data. In addition the result can be further interpreted to better understand the brain activation under which the brain data is collected. However, these results and interpretations could still be improved, and further exploration on the subject matter is indeed a huge research prospect. | en_NZ |
| dc.identifier.uri | https://hdl.handle.net/10292/9659 | |
| dc.language.iso | en | en_NZ |
| dc.publisher | Auckland University of Technology | |
| dc.rights.accessrights | OpenAccess | |
| dc.subject | Functional Magnetic Resonance Imaging (fMRI) | en_NZ |
| dc.subject | Evolving spiking neural networks | en_NZ |
| dc.subject | NeuCube | en_NZ |
| dc.subject | Spatio- spectro- temporal data | en_NZ |
| dc.title | Development of a new computational model for mapping, learning and mining of 3D spatio-temporal fMRI data | en_NZ |
| dc.type | Thesis | |
| thesis.degree.discipline | ||
| thesis.degree.grantor | Auckland University of Technology | |
| thesis.degree.grantor | Auckland University of Technology | |
| thesis.degree.level | Doctoral Theses | |
| thesis.degree.name | Doctor of Philosophy | en_NZ |