Fuzzy Clustering of fMRI Data Sets

Authors: Federico de Martino, Alain Smolders, Rainer Goebel

Purpose. Clustering of fMRI data provides a data-driven technique to find clusters of voxels that share similar time-courses. Such a technique is particularly useful when used for slow event related design in which the experimenter is interested in finding different brain areas exhibiting different event-related brain dynamics (e.g. mental clock task).

Requirements. The plugin requires version 1.8 or higher of BrainVoyager QX. In addition, a VMR file must be loaded before starting the plugin. For optional event-related averaging (see below), the selected protocol must have "volume resolution".

Running the plugin. To start the plugin, click the "Clustering of time series" entry in the Plugins menu and follow the instructions. The Log window will report about the progress of analysis. The plugin will first ask for a VTC file to be used for applying the clustering method. After the .VTC file has been selected, the plugin will ask wether you want to perform the clustering on masked data. It is advised to provide a cortex-based masked file to reduce the number of voxels for further calculations.

If masking is selected, the program will ask for a respective .MSK file. After reading the VTC data, the plugin will ask wether you want to perform clustering on event-related averaged time series for a condition of interest.

If event-related averaging is selected, a window for averaging is defined by asking for the number of TRs before (pre) and after (post) the onset of the condition are asked in order to define the window of averaging. After this information has been provided, the plugin will ask for the protocol file to be used for creating the averaging. The plugin will then ask for a condition in the protocol to be used for averaging (only one condition can be selected for averging at present).

The next optional step allows data reduction using Principal Component Analysis (PCA). If selected, the plugin asks for one of two strategies to reduce the dimensionality of the data. Strategy 1 asks for the number of components to retain. Strategy 2 asks for the minimum variance explained in the data (default option). The plugin will inform how many components need to be included to retain the specified variance proportion.

The plugin now asks for the number of clusters for classifying the voxels. Since this classification is "fuzzy", the plugin will finally ask for a "fuzziness" coefficient. It is recommended to simply accept the suggested default value.

The plugin now iteratively performs the clustering operation. The Log informs about the progress with regard to a objective function. When finished, the plugin asks for a name for the resulting clustering file, which will be the saved in the format of a .ICA file. To visualize the obtained cluster assignement maps, open the "Overlay Independent Components" dialog by selecting the "Overlay ICA" menu item in the "Analysis" menu. You should now see a screen similar to the one below. Click at the cluster maps to inspect how the voxels have been clustered.

The overlayed maps reflect the computed fuzzy membership values with respect to the selected cluster. Membership values vary between [0, 1] but have been rescaled for visualization purposes between [0, 10]. The "Min" threshold of a map can be increased (decreased) to show voxels with a more (less) clear assignment to the current cluster. The time course of a cluster can be inspected by checking the "Show component time course" option. The time course plot shows the time course of the centre of the selected cluster. If the data has been averaged, the "Use protocol" option in the time course plot should be deselected since the displayed protocol snippet does not match the onset of the selected event-related averaging condition.