Linear Temporal Filtering

Disclosure: this MEDx module was one of the two projects I was first assigned to in 1998 when I joined Sensor Systems for the first time. (My other project was the clinical brain mapping project, a much longer-term project, on which I served as Lead Engineer. Ask Medical Numerics (nee Sensor Systems) about it! :-) )

(1) Select Toolbox --> Functional --> Filtering..., and within the Filtering dialog box click on the Temporal Filtering tab.

(2) Set Input Group to New Group. Either type it in, or use the Select... button, whichever you like better.

(3) Set Filter Type to High Pass.

(4) Under Parameters, set High pass (sec) to 240. This is the cut-off frequency threshold for the high-pass filter, expressed in seconds rather than in Hertz.

A rule of thumb for how to select a "reasonable" value for this parameter is twice the cycle length of your on-off cycle of interest. For example, in the Finger Tapping motor paradigms, 24 seconds of the Fixation condition alternates with 24 seconds of the Tapping condition. This means that one cycle length of the effect of interest is 48 seconds. Twice 48 seconds is 96 seconds. If you look at the Protocol File for the MotorL paradigm in /export/w/apps/tcl/motorL_ProtocolFile.txt, you'll see that the parameter TemporalHighPassFilterCutoff has indeed been set to 96.

It's a little trickier in the Implicit Reading paradigm, since there are not two but three conditions, Fix, RealWord, and FalseFont. In this paradigm, 42 seconds of RealWord are followed by 18 seconds of Fix, followed by 42 seconds of FalseFont, followed by another 18 seconds of Fix. If you consider RealWord and FalseFont as both being activation conditions, then the cycle of interest is only 60 seconds long, in which case the cut-off frequency could be set to 120 seconds. However, whoever set up this paradigm evidently considered RealWord and FalseFont to be distinct conditions, so their cycle of interest was 120 seconds long, making their cut-off frequency 240 seconds.

Note that we are using different cut-off frequencies between the Finger Tapping motor paradigms and the Implicit Reading paradigm. Some might argue that for consistency, the same cut-off frequency should be used. What do you think?

(5) Set TR (sec) to 3.0, and leave Family set to Butterworth.

(6) Click on the Preview button. This displays the filter you have set up in both the frequency and temporal domains.

(7) Click on the Apply button to attempt to filter the data. You'll get a vague error message about "Duplicate Sequence Numbers". This indicates that even though the images are loaded into MEDx, MEDx still somehow doesn't know the proper temporal order of the images! This is one of the clunky quirks of MEDx. Dismiss the error message.

(8) To inform MEDx that the images are indeed in the proper temporal order, first make sure you are looking at the Group page of 84 scans (use the Page Manager if necessary). Then select Toolbox --> Functional --> Time Series.... In the Time Series dialog box, click on the OK button. Voila! Problem fixed. Yes, I know, very unintuitive.

(9) NOW back in the Filtering dialog box, click on the Apply button again. This will now apply the temporal filter to your fMRI data. This process does not create a new Group -- the old data is overwritten by the new temporally filtered data.

The next step is Spatial Smoothing. This is also filtering, but in the spatial rather than temporal domain.

For more on this module, see this stand-alone MEDx tutorial. (I wrote this tutorial while I was at Sensor Systems.) See also Section 31.7.3 of Chapter 31 of the MEDx User's Guide.

Return to tutorial main page