A Comparative Assessment of Statistical Approaches for fMRI Data to Obtain Activation Maps
Functional Magnetic Resonance Imaging (fMRI) lets us peek into the human mind and try to identify which brain areas are associated with certain tasks without the need for an invasive procedure. However, the data collected during fMRI sessions are complex; this time series of 3D volumes as images of the brain does not allow for straightforward inference.
Multiple models have been developed to analyze this data and each comes with its intricacies and problems. Two of the most commonly used models are 2-step General Linear Model (GLM) and Independent Component Analysis (ICA). Meanwhile both are trying to answer the same question, GLM is driven by Frequentist Statistics and is a parametric approach, whereas ICA is driven by Computational Optimization and is a nonparametric approach. In practice, these models are often perceived as black boxes, but it is important to understand the underlying assumptions and modeling choices to infer how well the model fits the data. We compared these two approaches by first discussing their theoretical framework, and then fitting the models to real and simulated fMRI data by using packages developed and readily available in R. The real data are obtained from an open source database named BOLD5000, a large-scale, slow event-related fMRI dataset. In the BOLD5000 study images were presented for 1 second, with 9 seconds of fixation between trials. The goal is to compare the activation maps from analysing real data and to understand under which circumstances the models do not fit well by using simulations. We compared two tasks: viewing an image vs fixating a cross for one subject only (as we decided to remain in the native subject space). The outputted activation maps for both the GLM and ICA indicated increased activity in the occipital lobe during the image viewing task. We expect the maps from GLM to be more contaminated by noise than those produced by ICA and for this to be revealed by the comparison of the activation maps. We also expect the ICA model to capture activation better for datasets simulated under conditions that deviate from the GLM assumptions when simulating data.