Abstract:
|
Independent component analysis (ICA) is used to identify spatially independent resting-state networks (RSNs) from fMRI data. Estimation of subject-level RSNs depends on relating the results of group ICA (GICA) to subject-level fMRI data using methods such as back-construction (BR) (Calhoun et al., 2001) and dual-regression (DR) (Beckmann et al., 2009). A recently proposed hierarchical covariate ICA (hcICA) method (Shi and Guo, 2016) provides a model-based estimate of subject-level RSNs. BR and hcICA both require each subject to be included in a computationally demanding GICA, precluding cases where there are too many subjects to feasibly include or too few subjects to obtain reliable group RSNs. While DR avoids this requirement, it tends to result in noisy estimates. We propose using well-established RSNs based on GICA of large fMRI datasets as population "templates" for extracting RSNs in a new subject. Using an empirical Bayesian approach and EM algorithm, we combine population-level templates and unique subject characteristics to obtain reliable subject-level RSNs. The proposed methods are also computationally efficient and eliminate the need to label or match components.
|