Preprocessing for fMRI (and a little bit for diffusion MRI)

This chapter is about the processing of fMRI data that is performed after the acquisition of the data, and before using a statistical model to try to infer what parts of the brain were involved in the task. It is aimed primarily at people who are about to undertake their first fMRI project, or who have already completed one or two and who want a greater understanding of the analysis steps they have been taught to perform. It will focus on explaining the different steps that constitute what is traditionally referred to as “preprocessing.” But it will also touch upon some of the MR-physics relevant to the acquisition, as well as on the statistical modeling that is used for the inference, as we think this makes it easier to understand why we do some of the preprocessing. The aim of preprocessing is twofold: 1. To improve location accuracy, i.e. to ensure that we are able to accurately assign an observed activation to the right part of the brain anatomy. 2. To increase statistical power, i.e. to try to detect and remove as much variance unrelated to the experimental task as possible, thereby making it more likely that any activation is statistically significant. The division is not necessarily as clear cut as that. For example, correcting for subject movement over time will primarily be aimed at increasing statistical power, but it is easily realized that large, uncorrected, movement would also impair localization. This chapter will give an overview of the following preprocessing steps. Distortion correction : fMRI images are distorted. It will be explained why this is, and how it can be corrected. Movement correction: Subject movement is the greatest source of unwanted variance in fMRI. The different ways in which movement can affect the data will be discussed, along with methods for how it can be corrected. Slice timing correction: The different slices of an fMRI volume will be acquired at different times, while the statistical modeling often assumes a single time point. This will be explained, along with ways to correct it. Physiological noise correction: It will be discussed how breathing and cardiac pulsation introduce unwanted variance to the data, and how it can be corrected. Removal of unwanted variance: Independent Component Analysis (ICA) and “Scrubbing” are methods for removal of unwanted variance that may remain even after applying the corrections above. Co-registration of functional and structural data: fMRI images often have poor resolution and tissue contrast, which makes anatomical orientation difficult. It is therefore useful to align them to a high resolution structural (e.g., T1-weighted) image.