Tutorial 10 - Independent Component Analysis (ICA) and Resting State fMRI

We know that fMRI task data is comprised of known signals plus various sources of noise (such as drift, breathing, motion artifacts). Our approach thus far has been to predict the known sources of activation changes while either preprocessing data to reduce noise (i.e., residuals) or to model out known sources of noise (for example, when we include motion parameters as predictors of no interest). This approach is theoretically driven data analysis.

However, we may not always know how to model certain patterns of activation (e.g., activation related to a seizure or hallucination) and there may be sources of noise we haven't accounted for. Data-driven analyses can provide a valuable tool in such cases.

The goal of Spatial ICA is to conduct a data-driven decomposition of fMRI signals into a collection of independent sources. As we should expect with fMRI, many of these sources will be driven by noise while others will be related to experimental conditions.

In spatial ICA, we are computing a decomposition that will provide a collection of independent time courses that can be used to model observed signal over space – either all voxels, or some subset of voxels. Then, for each voxel, ICA will tell us to what degree (positive or negative) each component contributes to the observed signal.

In other words, ICA will give us a collection of components with which we can model each voxel’s observed time course as a weighted sum, which can be visualized as follows.

This basic model should seem familiar – as it’s quite similar to how we model observed signal as a weighted sum of predictor functions in the GLM. Here though, ICA will automatically compute both the independent components and the model weights such that certain statistical properties are satisfied. In particular, the independent components are generated such that they collectively share an approximately minimal amount of mutual information. A nicely balanced explanation of how ICA works, including for fMRI analysis, is available at

http://users.ics.aalto.fi/whyj/publications/thesis/thesis_node8.html

Why ICA?

ICA can be a useful complement to traditional GLM analyses.

Common uses of ICA in fMRI are:

1. Removing noise

Artifacts often account for a lot of variance. The approaches we have seen so far include filtering out artifacts (e.g., low-frequency drift) and modelling out artifacts with predictors of no interest (especially motion parameters). However, these approaches may still leave considerable variance accounted for (thus increasing residuals). For example, head motion may cause artifacts that do not match expectations (e.g., an abrupt motion can yield spike artifacts). Thus another approach is to extract artifacts from the data itself.

ICA can be used to flag noise components and their beta weights. These components can then be subtracted from the data to leave cleaner data. In this approach, the removal of noise will reduce the residuals and improve statistical significance (Thomas et al., 2002). Alternatively, the time courses from noise components could be extracted and included in a GLM as predictors of no interest.

2. Identifying Networks in Resting State Data

Fluctuations in brain networks during resting state scans are spontaneous and can't be modelled with a GLM. One common approach to analyzing resting state data is to use ICA to identify networks with correlated activation (and/or deactivation). For example, ICA can identify foundational brain networks (e.g., default-mode network, DMN; task-positive network; sensory and motor networks).

3. Identifying Activation that is Not Easily Modelled

Some paradigms expect activation that is correlated with events but not in a straightforward way. For example, a participant may have a seizure or halluciantion that evokes neural processes that are not well modelled with convolved box-car predictors.

ICA in BrainVoyager

BrainVoyager offers a simple implementation of ICA for a single run that makes it easy to visualize components. Though beyond the scope of the course, BrainVoyager also offers a more complex implementation, the Group ICA Plug-in, that can investigate ICA components across participants.

In this part of the tutorial, we will interpret the results of applying ICA to three different scans – one experimental scan and two resting state scans. Our goal is to identify independent components and their associated statistical maps that correspond to BOLD signal as opposed to various sources of noise. We will use 3 techniques to help determine whether an IC is signal or noise:
1) IC time course examination,
2) IC map examination, and
3) fingerprinting.

The following instructions will guide you through interpretation of ICA applied to an task-based scan from our localizer. Then you will apply the same techniques to two resting state scans in an effort to find the components most likely to be driven by BOLD signal.

Examining One IC

When running ICA, the experimenter must decide how many components to search for. In this example, we used Brain Voyager's default to decompose the data into 30 independent components.

1) Open the anatomical file P02_Anat-S1_BRAIN_IIHC_MNI.vmr and then, from the Analysis menu, select Overlay ICA…. For this tutorial, we are using pre-computed Independent Component Map (.ica) files. To run ICA yourself, consult the Brain Voyager documentation at

BrainVoyager Documentation

1.1. Go back and Click the Load .ICA… button under the Analysis tab then select and open Loc1-30Comps-DefaultParams.ica

1.2. Check the Show component time courses box. The Time Course Plot window will then open for you to visualize IC time courses.

2) Make sure IC 1 is selected then close or move aside the Independent Component Maps window.

2.1. Without changing any thresholds, explore the IC map overlaying the anatomical scan.

Question 1: Where do you notice high degrees of variance explained by this IC? Looking at the time course for this component, does it look like what you would expect real activation to look like? Why or why not? Based on the spatial distribution and time course, do you think this is BOLD signal or noise? What might explain this component?

3) Reopen (or go back to if it's still open) the Independent Components Map window (Analysis > Overlay ICA…)

3.1. With IC 1 selected, click the Fingerprint button and the IC Fingerprint window will appear.

There are certain properties that we expect for "real" activation vs. artifacts.

1) Real activation should have a high degree of clustering because adjacent voxels fall in the same functional region; whereas, some sources of noise may be more "higgledy piggledy" (one voxel here and there.

2) Real activation should have a temporal autocorrelation (as indicated by the one-lag autocorrelation) because hemodynamics do not change dramatically from volume to volume; whereas, some artifacts will not have a strong autocorrelation.

3) Real activation should have power in intermediate temporal frequencies; whereas, artifacts often show up in low frequencies (e.g., drift) and high frequencies (e.g., breathing, cardiac).

4) The histogram of voxel intensities should have different properties for real activation (high skewness, low kurtosis vs. artifacts.

Since this is a lot of information to evaluate, we can make the task easier by rendering a "fingerprint" to visualize these attributes in a polar plot. In the white fingerprint below (or the grey-blue one next to it), the 11 spokes represent different attributes (e.g., 12 o'clock = degree of clustering) and the length of each spoke indicates the strength of that attribute (e.g., the plot below shows a high degree of clustering but a low kurtosis). This is the expected pattern for real BOLD activation. Note that it looks a bit like a Mercedes symbol -- three prongs at 12 o'clock, ~5 o'clock, and ~8 o'clock.

Question 2: Based on its fingerprint alone, does IC1 appear to be real activation or an artifact?

Question 3: Using the techniques described above, quickly look through each of the 30 ICs and identify two or three that you think are driven by BOLD signal, and two or three that are driven by different types of noise.

Take advantage of the fact that we have a protocol for this scan – and so we might expect activation in certain areas and at certain times.

Question 4: ICA can be used for task data to determine BOLD activation and sources of noise. Give one example of when ICA could be useful for understanding BOLD activation beyond the conventional model-driven approach. How could ICA be used to clean up noisy data?

Resting State ICA

ICA can also be applied as a data-driven approach to find networks in resting-state scans.

Question 5: Based on the lecture, what is a model-driven approach to analyze resting state data? When might the model-driven approach vs. the data-driven approach be appropriate?

One additional anatomical file P99-Anat_IIHC_MNI.vmr and two more ICA files (P-99_RestEyesClosed1-S1R1_3DMCTS_LTR_THPGLMF2c_MNI.ica and P-99_RestEyesClosed2-S1R1_3DMCTS_LTR_THPGLMF2c_MNI.ica) are provided.

For P-99_RestEyesClosed1-S1R1_3DMCTS_LTR_THPGLMF2c_MNI.ica, use the same steps as above to quickly examine each of the 30 IC time courses, IC maps, and fingerprints.

(Note: for resting state ignore the error message when clicking show timecourse - there is no protocol because it is resting state),

Question 6: For P-99_RestEyesClosed1-S1R1_3DMCTS_LTR_THPGLMF2c_MNI.ica, list two or three ICs that are likely to be driven by BOLD response, and two or three that are likely to be noise. Support each decision using all three interpretation techniques: IC time course examination IC map examination, and fingerprinting. Briefly explain what evidence each of these interpretations provides for your argument.

Question 7: Which component in P-99_RestEyesClosed1-S1R1_3DMCTS_LTR_THPGLMF2c_MNI.ica that looks like task-positive vs. resting-state networks? What do the colors maps mean? Why would BV have a button that allows you to flip the color mapping?

Question 8: Based on your experiences in this tutorial, even though ICA is cool, what factors can make it challenging to interpret?