Chapter 4 Example box: Group-ICA networks from different datasets

Introduction

The aim of this short example is to show you a set of resting state networks to help you get familiar with recognising these common networks across a variety of datasets.

This example is going to show you the result of a group-level ICA that was performed using the melodic tool in FSL. However, please note that other packages are available for running ICA.

Please download the dataset for this example here:

Data download

The dataset you downloaded contains several examples of the melodic_IC.nii.gz file that is created when running a group-level ICA. In the example below, you will take a look at the group-level ICA components and compare the results that can be obtained in different conditions. In the first section, you will be able to look at the influence of data quality and quantity on the ICA decomposition. The second part is aimed to provide an understanding of splitting of ICA components as the dimensionality of the decomposition is increased.


Different acquisitions

As a general rule of thumb, ICA works better the more data it is given. As a result, having good quality data with high temporal and spatial resolution will result in better definition and separation of components. Additionally, having more subjects also improves the ability of a group-level ICA to accurately define components. Lastly, the preprocessing performed on the data, in particular the amount of spatial smoothing, plays a very important role in how ICA components will end up looking. In this section, you are going to have a look at three different group-level ICA decompositions that are all at the same dimensionality (25 components). These were each obtained from different datasets and have different data quality, subject numbers and preprocessing (as summarised in the table below).

Dataset TR (in seconds) Number of volumes per subject Voxel size (in mm) Number of subjects Smoothing FWHM (in mm)
Standard EPI 2 180 3 x 3 x 3.5 20 5
Multiband a 1.14 790 2 x 2 x 2 64 0
Multiband HCP 0.73 4800 2 x 2 x 2 820 0

Open a command line terminal and change directory into the file you have download called 'data_4.2' (using cd). We are now going to use FSLeyes to visualise these three different sets of group ICA components using the following command:

fsleyes -std melodic_IC_25_s0_n64_MB6.nii.gz -n multiband_1 -un -cm red-yellow -nc blue-lightblue -dr 5 15 \
 melodic_IC_25_s0_n820_MB8_HCP.nii.gz -n multiband_HCP -un -cm red-yellow -nc blue-lightblue -dr 30 100 \
 melodic_IC_25_s5_n20_EPI.nii.gz -n standard_EPI -un -cm red-yellow -nc blue-lightblue -dr 5 15 &

Click on the locks in the bottom-left window to make sure the volumes change with each other. You can now use the eye toggle button in the same window to switch maps on and off in order to compare the components across the three different datasets.

These maps are all re-ordered to so that each volume contains the best match across the three ICA results. The first 8 maps (volumes 0-7) show relatively good correspondence between the datasets. Towards the later volumes numbers you can see that the components are more likely to be noise (at least in some datasets). We would not expect noise components to look the same across different datasets. Have a look at several of the components and compare how they appear visually across the different datasets to get a feel for the influence of different variables on the results of a group ICA.


Different dimensionality

The aim of the first section was to get a feel for how ICA components may look different as a result of the quality and quantity of the data. In this next section you are going to look at the effects of the dimensionality of the ICA decomposition. For this, we have two examples of group-level ICA decompositions that were performed on identical data. The only difference is the number of components that was extracted (which was 25 in one analysis and 50 in the second). Please run the following command to open both of the ICA results up in FSLeyes:

fsleyes -std melodic_IC_25_s0_n820_MB8_HCP.nii.gz -un -cm red-yellow -nc blue-lightblue -dr 30 100 -n 25\
  melodic_IC_50_s0_n820_MB8_HCP.nii.gz -un -cm red-yellow -nc blue-lightblue -dr 30 100 -n 50 &

Now go to the 'View' menu at the top and switch on ortho view to get the maps lined up next to each other (this will give you two windows that are both showing overlapping results for both 25 and 50). Next, make sure that the window on the left is showing the 25 results (by clicking the 'eye' toggle button next to 50), and make sure that the window on the right is showing the 50 results (by clicking the 'eye' toggle button next to 25).

You can now have a look at a few examples of how components might split if the dimensionality increases. In the textbox next to 'Volume' on the bottom of the left panel, put one of the volumes shown below in the column for 25 (i.e., either 2, 5, or 9). Now have a look at the numbers in the same rom for the 50 column to show how the main component was split across multiple smaller new components.

25 50
2 5, 9, 11, 14
5 32, 33, 35
9 10, 20

You can see, for example, in the second example (volume 5 in d=25), that the original network is split into left and right lateralized regions (32 & 33) and the medial region (35). In the last example (volume 9 in d=25), the original network is split into anterior (20) vs posterior (10) subregions.

At this point, it is worth noting that the way specific networks are split under increased dimensionality is likely to vary for different datasets and different dimensionalities. Nevertheless, we hope that these examples are useful to give you an idea of the type of splitting we might see when increasing the dimensionality of an ICA.


Data credits

This example includes data from the Human Connectome project (https://www.humanconnectome.org/).

Note that group ICA for the HCP examples was run on the cortical surface. The volumetric results displayed in this example were obtained by performing dual regression (multiple regression of stage 1 timeseries onto the volumetric data).