Independent Component Analysis (ICA) is most commonly used in EEG. It’s application in there is relatively straight forward. I’m a MRI person by trade and it’s use has been slowly creeping in from FSL’s ICA denoiser to using ICA to understand functional network connectivity in FSL or GIFT (one day I’ll get to a tutorial on GIFT). In this post I’ll cover the basics of what an ICA is and how it’s used in fMRI. This is by no means extensive and should supply you with enough knowledge to give you familiarity and resources to dive deeper.
I’m a huge fan of Ye’s music and my internet is out so I can’t stream it. Thankfully, my upstairs neighbor is blasting Graduation. Unfortunately, my downstairs neighbor has poorer taste in music and is playing some Imagine Dragons song. My goal is to only listen to Ye but I’m having a hard time listening to it over a band that might be worse than Nickelback. Is there a way to unmix these sounds and output the only one I want so my ears don’t bleed? Yes, ICA is here to save the day! Science really does save lives.
How can ICA do this? Like every tool we use, there are some assumptions we must make about our data:
Data is not normally distributed, meaning not Gaussian (statistically independent).
Data sources are linear, meaning they can be added to each other in a linear fashion.
You might be thinking, that doesn’t make sense - if you’re combing many independent variables, then according to the central limit theorem they should be gaussian in nature. Which is true but through linear transformation of matrices, it’s possible to maximize nongaussianity and find separate components. If you would like a relatively simple explanation of the maths, read through this paper that I found helpful in understanding ICAs.
In the brain (fMRI) the BOLD signal is mostly from oxygenated blood and some other sources (movement, other weird noise in the MRI) that generate the BOLD signal. To unmix the signals that we don’t care about from the one we do, we can use ICA to do this. fMRI data is extremely noisy. There are wiggly participants, weirdly shaped brains, a really strong magnet doing physicists know what, a cardiovascular system that allows us to get a signal and at the same time makes it hard to measure what we’re interested… The list goes on but you get the point. Despite all of these pesky problems scientists have persevered and continue to create solutions, like ICA. FSL’s MELODIC is a great tool to accomplish denoising. While it still has qualitative parts to selecting the components to keep, the results are fairly robust among trained researchers.
The basics of how fMRI analysis worked (sometimes still done this way) are poking around the brain and looking at which voxels are active in a condition compared to another condition with tightly controlled experiments and models. You couldn’t look at the time course of the data, how it fluctuates over time, or one brain area’s relationship to another area. ICA is one solution to investigate how BOLD fluctuates over time and how it relates to other areas.
If you want some more details on how ICA is used in fMRI, read this paper, and then this one. I strongly suggest brushing up on your matrix multiplication. One of my collaborators, Jeanette Mumford has a great video on this and many other great videos on statistics in fMRI!