Bayesian modeling and inference for functional magnetic resonance imaging of the visual cortex

Abstract

For the most e ective use of functional magnetic resonance imaging (fMRI), mapping the brain signals to a statistically valid map is crucial. The common approach to create a statistic at each voxel is applying the frequentist or the classical statistics. However, there are many challenges raised by the use of classical statistics to test the functional data such as the multiple comparison problem, and the limitation in the interpretation of the parameters. As an alternative, a Bayesian approach can be used to assess the data based on the posterior probability distributions of the parameters. In this study, the power of Bayesian inference was compared against classical inference in random e ect analyses: A group data collected from visually stimulated volunteers was assessed following a simulation study. In order to assess the results of the statistical inference for the group level, the variation of the e ect sizes with respect to stimulus frequency was used. A comparison was performed between the change in the e ect sizes of lateral geniculate nuclei (LGN) and primary visual area (V1) during graded visual stimulation by using the posteror probability maps (PPMs) with an e ect size threshold of zero. This comparison became possible with the fact that once we had the posterior probabilities the activity in LGN was able to be visualized by changing the e ect size threshold and without decreasing the signi cancy threshold, which is not possible to achieve with classical inference where the data is tested against the null hypothesis. Despite of the small magnitude of activation in LGN we could show the connectivity between V1 and LGN and the di erences in response characteristics during graded visual stimulation.|Keywords : fMRI, Bayesian inference, visual cortex, LGN, PPMs.