Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion
Jonathan Power, Kelly Barnes, Avi Snyder, Brad Schlaggar, Steve Petersen
Neuroimage 2012 Feb 1; 59(3):2142-54
Most visitors to this site will be familiar with this paper already, it's by far my most cited paper. A lot of work has been done on motion artifact since this paper emerged and there's little need to flesh out here what is fleshed out elsewhere. So I'll just tell the story of how this paper emerged.
My doctoral thesis was to be the study, using resting state fMRI, of how the pediatric brain changes organizationally as it becomes a young adult brain. The Petersen/Schlaggar lab had collected an impressive amount of resting state data for the time, most of it in young adults, some in teenagers, and some in school-age children. For the first 1-2 years in lab I studied young adults, doing the analyses that led to my 2011 Neuron paper. Most of my time was spent learning Matlab and wrestling with how to visualize data effectively. I spent zero time worrying about processing the data, since it had already been processed when I arrived in lab.
Once I felt I knew the young adult data well enough I began to study children in the summer of 2010. I had no inkling that motion was going to be a confound. The lore was that someone had already looked at motion one summer a few years ago and it hadn't made much difference in functional connectivity. I remember being worried about brain sizes mainly. Nico Dosenbach was in lab that year for a one-year fellowship during his residency, and If Nico hadn't been so anal about forming his datasets from the lowest-motion runs in each subject, I probably would have never checked for motion effects myself. One day Nico handed me a text file listing low-motion child datasets. Late that night I ran them through a modularity analysis and the pictures that emerged looked like networks from young adults, not children. I double-checked the links and reran the code on adults, children, and these "good" child datasets. Nothing changed. I remember actually feeling sick.
Over the next few months I began to pry into the processing stream at WashU. There, we typically use "4dfp" image processing tools, which are written and maintained by Avi Snyder. These are great, fast, flexible tools, but they're idiosyncratic to our institution and I wasn't sure I could use them to do what I wanted. So I began to pull preprocessed images into Matlab and learned to do functional connectivity processing on my own. This took months to figure out satisfactorily, but eventually I could control the data well and quickly.
By SFN 2010 I was sufficiently alarmed that I put the motion-related findings, including scrubbing and distance-dependence, on my poster, though that wasn't part of the abstract. For the next few months I actually did little work on motion artifact. I was preoccupied with revising the Neuron paper. And I didn't want to rush out something that was wrong. I wasn't a methods person by training or inclination and I was paranoid that I might have done something stupid, myself, to produce the artifact (this is why, at one point in the paper, I show that the artifact is still there when you just bandpass filter and spatially blur the data without any further processing). It took me until the mid-spring to write the paper and the paper underwent 3 rounds of revision (one reviewer really did not like the paper), so it took a while to come out, almost a year after I first noticed the problem.