Machine Learning to Analyze fMRI Brain Activity Recordings

Date

2019-05

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

In recent years, data-driven machine learning (ML) algorithms have seen increased use to study brain activity fMRI time series. Giving mathematical solutions by constructing quantitative models has proved to be a reliable approach to overcome clinical obstacles. These algorithms provide non-invasive tools to analyze brain activity and pave the way to discover biomedical results that can guide physicians to make more precise clinical predictions. One goal of this dissertation was to facilitate noninvasive localization of the epileptic focus. We designed a deep learning algorithm to classify the patients by using some of the interactivity coefficients as the input of a neural network (NN) and introduced parsimony constraints to build a very restricted MLP to avoid overfitting. Intra/Inter-region connections obtained from resting-state fMRI showed a strong association to clinical diagnosis of seizure onset zone at the lobe level. The results may improve the surgical outcomes as successful seizure surgery is predicated upon the ability to localize the seizure onset zone. We also narrowed down and detected a more accurate focus within each lobe. Another goal of the dissertation was to combine spectral analysis of graph Laplacians with simulated annealing to automatically generate optimized clustering of time series by minimizing the clustering cost function. We explained and implemented this method on cortex parcels. Additionally, Graph Mining algorithms have been investigated with the achievement of binary classification. We applied a Frequent Subgraph Mining algorithm to mine the frequent subgraphs in a graph data set of brain activity and then achieved feature vectors to represent the graphs. The method was successful in some binary classification tasks. It also provided useful information about the functional connectivity of patients in the same class and helped to gain a deeper insight into the topological differences between different patients with different focuses.

Description

Keywords

Time Series Classification, Deep learning, Machine learning, Mutual Information, Functional magnetic resonance imaging (fMRI), Epilepsy Seizure Focus, Graph Mining

Citation

Portions of this document appear in: Azencott, Robert, Viktoria Muravina, Rasoul Hekmati, Wei Zhang, and Michael Paldino. "Automatic clustering in large sets of time series." In Contributions to Partial Differential Equations and Applications, pp. 65-75. Springer, Cham, 2019. And in: Hekmati, Rasoul, Robert Azencott, Wei Zhang, Zili D. Chu, and Michael J. Paldino. "Localization of Epileptic Seizure Focus by Computerized Analysis of fMRI Recordings." arXiv preprint arXiv:1812.04533 (2018). And in: Hekmati, R., R. Azencott, W. Zhang, and M. J. Paldino. "Machine Learning to Evaluate fMRI Recordings of Brain Activity in Epileptic Patients." In Q-Bio Conference. 2018.