New approaches to statistical analysis of fMRI data

Functional magnetic resonance imaging (fMRI) enables recording of changes in blood oxygenation level and thereby provides insight into human brain function. The collected fMRI data have to be analyzed statistically, and group analysis is crucial for generalization of the results. The selection of the statistical test depends on properties of the data, as well as on the research question. Since the brain-activity-related fMRI signal is small compared with noise in the data, removing various kinds of artifacts enables more effective and reliable detection of the true stimulus-related activations. To find out how the human brain works in everyday-life situations, fMRI experiments have started to move from traditional, highly controlled setups to using more naturalistic stimuli and tasks, such as viewing movies. Such rich and continuous stimuli bring challenges to data analysis, and call for data-driven approaches such as independent component analysis (ICA). In contrast to parametric statistical tests (e.g. the t-test) commonly used in fMRI analysis, nonparametric methods make only minimal assumptions about the data and may thereby provide more robust results. However, the widely-used nonparametric permutation test is computationally heavy. This thesis introduces SumLog, a new sensitive and computationally efficient nonparametric method for group-fMRI analysis. Some artifacts such as linear drifts are easy to remove from the data by e.g. filtering, whereas physiological noise is more difficult to eliminate. This thesis thus introduces a new maxCorr method that allows extraction from group-fMRI data subject-specific components that are connected to respiration, heartbeat, and movement. In contrast to many existing methods, maxCorr does not need any reference signals or information about the stimulus, and it is thereby well-suited for cleaning data collected during naturalistic stimulation. Data-driven methods such as ICA are widely used for analyzing fMRI data from naturalistic experiments. ICA can unravel functional brain networks but both the spatial patterns and time courses of the networks depend on the number of the estimated components. In this thesis, group-ICA was applied at four dimensionalities to fMRI data collected during movie viewing. The analysis implied subdivision of three cortical networks into functionally feasible subnetworks.This thesis contributes to more effective and reliable statistical analysis of fMRI data from conventional and naturalistic experiments. It also adds to the understanding of human brain function in naturalistic conditions.