Contributive sources analysis: a measure of neural networks' contribution to brain activations

Abstract

General linear model (GLM) is a standard and widely used fMRI analysis tool. It enables the detection of hypothesis-driven brain activations. In contrast, Independent Component Analysis (ICA) is a powerful technique, which enables the detection of data-driven spatially independent networks. Hybrid approaches that combine and take advantage of GLM and ICA have been proposed. Yet the choice of the best method is still a challenge, considering that the techniques may yield slightly different results regarding the number of brain regions involved in a task. A poor statistical power or the deviance from the predicted hemodynamic response functions is possible cause for GLM failures in extracting some activations picked by ICA. However, there might be another explanation for different results obtained with GLM and ICA approaches, such as networks cancelation. In this paper, we propose a new supplementary method that can give more insight into the functional data as well as help to clarify inconsistencies between the results of studies using GLM and ICA. We introduce a contributive sources analysis (CSA), which provides a measure of the number and the strength of the neural networks that significantly contribute to brain activation. CSA, applied to fMRI data of anti-saccades, enabled us to verify whether the brain regions involved in the task are dominated by a single network or serve as key nodes for particular networks interaction. Moreover, when applying CSA to the atlas-defined regions-of-interest, results indicated that activity of the parieto-medial temporal network was suppressed by the eye field network and the default mode network. Thus, this effect of networks cancelation explains the absence of parieto-medial temporal activation within the GLM results. Together, those findings indicate that brain activations are a result of complex network interactions. Applying CSA appears to be a useful tool to reveal additional findings outside the scope of the "fixed-model" GLM and data-driven ICA approaches.