Checkpoint: Preprocessing

Now is a good time to review what we have done so far:

  1. We downloaded an fMRI dataset that has anatomical and functional images;

  2. We looked at the data in FSLeyes, FSL’s data viewer;

  3. We preprocessed the data using FEAT, FSL’s preprocessing tool.

Along the way you learned how to check the images before and after each preprocessing step. As you apply the same quality checks to different datasets, you will come across images that are difficult to judge - they may seem to be on the border of being acceptable or unacceptable.

It may be confusing at first. But over time you will develop your judgment about what images are clearly good, which ones are obviously bad, and which ones you will need to think carefully about either keeping or discarding. The more you think about why the results of a preprocessing step look good or bad, the easier it will become to make your judgments quicker and more accurately.

Try the following exercises in order to increase your fluency with FSL and to improve your judgment about the output from each step.

Exercises

  1. Run BET on the anatomical image sub-08_T1w.nii.gz with two separate fractional intensity thresholds: 0.1 and 0.9. Take a snapshot of each output image with FSLeyes using the camera button (it is located in the upper middle part of the viewer). Note the differences between the two. Is the output what you expected? If you had to use one image or the other, which one would you choose?

  2. Preprocess run 2 of the functional data using the FEAT GUI. To do this, select sub-08_task-flanker_run2.nii.gz from the func directory, change the output directory to run2, and make sure Preprocessing is selected from the dropdown menu. Keep the other settings the same as when you analyzed run 1. Do the same quality checks that you did for run 1.

  3. Preprocess run 1 using a 3mm smoothing kernel, keeping the other preprocessing options the same. (Make sure, however, to change the output directory to a new name in order to keep the output organized.) Before you look at the output, run another analysis with a 12mm smoothing kernel. Think about what you would expect the preprocessed functional data to look like, and then load the filtered_func_data.nii.gz images from each analysis into FSLeyes. How do they compare to your predictions?

  4. Preprocess run 1 using 3DOF for registration and normalization. How is the output different from what you saw when you ran the preprocessing with 12DOF? Why? (Hint: Review the Registration and Normalization page for possible reasons.)

  5. Rerun registration for run 1 using BBR instead of 12DOF. What difference does it make? How would you make a case to someone that you should use one instead of the other?

When you have finished doing the exercises, click on the Next button to begin the module on Statistics and Modeling.

Warning

Make sure you have preprocessed both run1 and run2 before proceeding; you will need both of them in order to run a higher-level analysis and use the scripts that are provided on Github.