Bootstrap-based Ensemble Clustering of Resting-state fMRI Time Series

Ashtari, Pooya | 2016

619 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 49381 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Vosoughi Vahdat, Bijan
  7. Abstract:
  8. Studies in recent years have shown formation of strongly functionally linked sub-networks during rest, networks that are often referred to as resting-state networks. RSNs not only have basic information about the brain but also play a key role in detecting brain disorders, such as Alzheimer and Autism; Consequently, they have been remarkably noticed by neuroscientists. Numerous methods have been used in order to extract RSNs using resting-states fMRI time series. Independent component analysis (ICA) is the most common method, whi have been reported to show a high level of consistency neurophysiology; however, its results is unstable in subject-level. is weakness restricted the ICA applications to group-level, or as a preprocessing method in order to remove artifacts and dimension reduction. Several existing clustering algorithms for network estimation, like k-means and FCM, are sensitive to the initial condition; us, they fail in capturing clusters with complex paerns. Some other clustering algorithms, such as spectral clustering, Although are capable of adapt itself with complex data, does not show a consistent result in brain parcellation because of hierarical structure of the brain and existence of unbalanced clusters. In this research, using a novel Ensemble clustering method based on bootstrap we have estimated RSNs from rsfMRI time series. In this method, not only the nature of rsfMRI time series has captured by bootstrap, but also the strength of clustering has increased and the variance has decreased. e suggested clustering method is inspired by the Bootstrap Aggregating (Bagging) method and has been designed such that to be suitable for clustering time series data. Furthermore, in order to resampling rsfMRI time series, the residual bootstrap and block bootstrap methods have been modified so as to become applicable for fMRI data. The resulting network matrix from proposed clustering method is very sparse and need very low space compared to correlation and partial correlation matrices. In the end, we evaluated our method in two ways. First, we apply our method to rsfMRI data (HCP database) and evaluate its performance using a predefined cluter validity index. Second, we used the estimated network as a features for each subject in order to solve a gender classification problem, then we compared that to other networks based on classification accuracy. the simulation results have shown that our methods, despite the sparsity, has a precision equal to partial correlation matrix and with combining this network with correlation and partial correlation matrix,
    we reached an accuracy equal to 80 in gender prediction classification problem, which is higher than the reported values
  9. Keywords:
  10. Bootstrapping ; Functional Magnetic Resonance Imaging (FMRI) ; Clustering ; Time Series Bootstrap ; Time Series Clustering ; Resting-State Networks

 Digital Object List


...see more