Pseudo-real fMRI data generation and its utility toward quantitative evaluation of analytical methods

Functional magnetic resonance imaging (fMRI) modality has been widely employed to measure neuronal activations of the human brain using such as a model-based general linear model (GLM) and data-driven independent component analysis (ICA) approaches. In this study, we were motivated to investigate the performance of two popular methods with a hypothesis that these methods would have advantages and disadvantages depending on the variability of the fMRI data across subjects in both temporal and spatial domain. To quantitatively evaluate two methods, the pseudo-real fMRI data were generated by combining the decomposed non-neuronal components estimated from real resting-state fMRI data and artificially generated neuronal components with varying degree of temporal and spatial pattern variability of task related activation patterns in an individual level. Using the pseudo-real fMRI data, the assessment of each method was conducted by comparing the estimated activations to reference neuronal activations. Our results indicated that the degree of spatial overlap size across subjects and degree of temporal pattern variability would be important factor to choose a proper analytical method.