The slice repair function was written to address a particular kind of intermittent electronic noise that affected the fMRI data from pediatric subjects who could not be rescanned. The noise caused certain slices to have many intensity spikes beyond 10% variation from the mean of the voxels. The spikes were both positive and negative, consequently, the artifact did not necessarily show up in the global mean (art_global function). The artifact (if present) is easy to see using the Contrast Movie (e.g. see the HBM 2005 poster in the documentation), but not as easy to see when viewing images of the raw data.
To answer your questions, try running the Contrast Movie before and after slice correction, and see whether the data after slice correction looks clean enough. To detect fewer bad slices, raise the threshold. Alternatively, if the first few scans are very noisy, try running the filter starting a few scans later than the first scan.
The acceptability of noise filtering will depend on the reviewers. Large spike transients may cause large estimation errors from the GLM. The noise filter replaces the large spike noise events with small unbiased "interpolation algorithm noise" on some slices within some volumes. The interpolation introduces a small bias and degrees of freedom error in the activation calculation. My opinion is that the noise filtering process generally reduces large errors to small errors, and thus is a net gain for the analysis of non-replaceable fMRI data. But it is hard to quantify how much noise filtering is OK, and other researchers may have different suggestions.
Best regards,
Paul
----- Original Message -----
From: "Ben Xu (NIH/NINDS) [E]" <[log in to unmask]>
To: [log in to unmask]
Sent: Monday, September 27, 2010 9:32:06 AM
Subject: [SPM] ArtRepair of bad slices
Hi,
I'm using the Artifact Repair Toolbox (v4) to correct bad slices in some of the EPI runs. For most of the runs, it clearly corrected/repaired the bad slices. However, for some of the runs which had no apparent bad slices (visually inspected and ran through art_global), the "Bad slices..." function detected "bad slices" in many volumes and the "bad slice" numbers across volumes are more or less the same. (Subject's head motion in these runs was minimal, i.e., less than .5 mm in all directions.) Here are the questions;
1. Could this be due to over correction by the "Bad slice.." function? Or, are those slices likely bad, but just too subtle to be visually obvious?
2. If it's due to over correction, what parameters can be adjusted to avoid that?
3. How many/what proportion of bad slices in a scan run could be considered acceptable for a scan data to be included in analysis.
Here are the options I used for the slice repair:
1. Bad slices: detect and repair
2. Which repair methods to use?: Repair Bad Slices and Write BadSliceLog
3. Which mask image to use?: Automatic
4. Select threshold (default): 5
Thank you very much for your help!
Ben
|