There are no set rules about the best values to use for the parameters in ArtRepair, and it may be helpful to understand the background of the question. Rapid scan-to-scan motions (e.g. one mm/TR) can cause spin history artifacts in the data. These artifacts occur when protons excited in one slice position are moved (by head movement) to another slice position within a TR, and they can generate a very large false signal. Whereas usual BOLD signals (except in visual cortex) are less than 1% signal change, the spin history artifacts can be 10% to 50% signal change. If these artifacts correlate with a task, the parameter estimates for the beta's will be poor. Even worse, spin history artifacts usually affect most of the voxels on one or more slices, leading to spatially widespread artifacts that will tend to pass the cluster size thresholds for significance.
To check if the scan-to-scan motion threshold is OK, use art_movie to view the scans AFTER the art_global step. In the art_movie display, real BOLD signals are nearly invisible while artifacts show up as bright yellow (positive) or bright blue (negative). If there are many big artifacts remaining after artifact filtering, the threshold was probably too lenient.
But using a strict threshold is balanced against the fact that more volumes are marked for repair. When a large fraction of volumes need to be repaired, completely deleting them would be the best, but this option was difficult to code. Instead, we used repair by interpolation. This approach introduces a small bias into the estimates of the beta parameters, and the bias gets worse as the repaired segment becomes longer. From my informal tests, there could be up to a 10% bias (e.g. 0.9% instead of 1%) for 20% repaired volumes, so my personal rule-of-thumb is a maximum of 20% repairs. For these cases, the goal was to estimate more accurate answers by excluding large artifacts at the cost of including a small bias. One caveat is that this limit was set for group studies which only use the beta estimates from each subject. For a study of activations in a single subject, long repairs inflate the possibility of a false positive activation, and the limit on fraction of repaired volumes should be tighter.
- Paul
Question:
I'm using the bad volumes function in art repair to correct for subject motion. I had originally defined a specific motion threshold to be used on all subjects (1mm/TR) and to repair all volumes above that threshold. However, with some subjects setting to that threshold would mean correcting for more than 10% of the volumes. I'm planning to increase the threshold for these subjects up to the point where 10% or less of volumes are corrected. I would not set the threshold higher than 2mm/TR. I'm wondering if this approach would be reasonable or if there is a better way to correct for motion.
|