Solution: It depends upon what you are doing. Typically the approach used is to preprocess your data the same way you would for, say, an fMRI analysis of another type (e.g. SPM). In this case, the functional data are spatially realigned, normalized (if desired) and smoothed in order to run ICA. However, if your goal is to examine potential artifacts in your data, then you may not want to smooth, since the artifacts can be "hidden" by the smoothing process.
Solution: Image intensity need not be scaled and images should be in 3D analyze format. All the images should have the same voxel size but the time points or number of images can be different for subjects.
Question 3: How do I change the defaults?
Solution: The defaults are specified in icatb_defaults.m file. Please see the appendix section of the GIFT walk-through.
Question 4: How do I handle out of memory errors?
Solution: Out of memory errors are common for multivariate approaches like ICA. Thus you want to minimize, wherever possible, the amount of needed memory. For example, don't reslice your data into voxels smaller than the acquired voxel size (since this doesn't add information and just increases memory). If you have very high resolution data, you may even in some extreme cases need to resample your data or work on a subset of the data. Finally, you will benefit greatly by closing all existing MATLAB applications and running MATLAB with out its Java Virtual Machine (JVM) . MATLAB with out JVM can be run by entering command matlab -nojvm at the DOS or UNIX prompt (or change your windows shortcut to include the -nojvm flag so it reads, e.g. "..\matlab.exe -nojvm". This will reduce the amount of memory MATLAB needs to run.
Solution: The ICA process is not interrupted with these errors. The explanation for each error is given below:
Question 6: Are there any guidelines for choosing the number of components?
Solution: The number of components can be estimated by selecting the estimate dimensionality step in setup ICA. If the number of components are estimated to be very high (like 100 components for 200 images) then around 20 to 30 components should give a reasonable answer.
Solution: Probably not, unless the algorithm fails to converge or you are interested in algorithmic comparisons.
Question 8: What is done during the "Group ICA" step for a single subject single session analysis?
Solution: For single subject single session analysis, there is only a single data reduction stage, and you cannot compute the group stats (e.g. T-maps).
Question 9: Are there any benefits to running subjects in a group vs. running them individually?
Solution: It is possible to run them individually, however it can be difficult to know how to compare components (since they do not come out in any particular order and there may be different components revealed in different subject). It is not as straightforward to do group analyses with ICA as it is for, say, a model-based approach. For more details please refer to the 2001 Group ICA paper in Human Brain Mapping. A related question is whether to run one or two ICA analyses on different groups. If the groups are patient vs control it may make sense to run two ICA analyses. We will soon be adding some tools to GIFT to allow multi-group comparison. In summary, the advantages of running subjects in a group over running them individually are as follows:
Solution: During the data-reduction step time points are not averaged but calculated using Principal Component Analysis (PCA). PCA finds the orthogonal components and the number of orthogonal components corresponds to the number 50 in the example. This typically corresponds to more than 99% of the variance in the data.
Question 11: What happens if the calculate ICA step doesn't converge during the analysis?
Solution: You can select the calculate ICA step instead of running all the group ICA steps. After the ICA step converges, you have to select one at a time back reconstruction, calibrate components and group stats steps.
Question 12: How does the SPM2/SPM5/SPM8 design matrix affect ICA?
Solution: SPM design matrix doesn't effect ICA process or run analysis step in GIFT (unless you are using the semi-blind ICA method). The SPM design matrix is used to sort components temporally. A detailed explanation of how SPM design matrix is used is given in appendix section of the Walk-through or in the HTML help manual (When you click Help button in the toolbox, the HTML help manual will open).
Question 14: How do I view all subject time courses at once?
Solution: All subject time courses can viewed at once by selecting kurtosis as sorting criteria and selecting all data sets in component set to sort. You can view all the time courses by clicking on SplitTimecourses button on the expanded view of the time course figure.
Solution: You can do different analyses with the same functional data by copying the parameter file and the Subject file to a different directory. After copying the files, use the setup ICA GUI to change the parameters for the new analysis by specifying the same prefix for the output files you have selected before (If the parameter file name is test_ica_parameter.mat then test is the prefix). You need not to select the subjects again.