There are two complementary 3D+time "quality check" programs in the AFNI package.

3dTqual computes the median of each voxel time series; this is the "median brick". For each time point, it then computes 1 minuse the correlation (Spearman rank, or quadrant) correlation coefficient of the actual data brick at that time with the median brick. Values that are close to 0 mean that 3D volume is correlated highly with the median brick, which is good. Values much larger than 0 are not highly correlated with the median brick, which means they are probably corrupted by large head movements or ghosting artifacts. Essentially, 3dTqual looks for 3D volumes that are "weird".

3dToutcount is a more detailed look at the data. It computes the median and MAD of each voxel time series - the MAD provides an estimate of the variability of the time series. Then each time point is examined; if it is significantly away from the median of its time series (relative to the MAD of that time series), it is marked as an outlier. (The "-save" option lets you save a dataset that is marked with the "outlierness" of each data point.) Then, for each time point, the number of outliers in that 3D volume is counted. Volumes that have "too many" outliers are flagged as "weird". This is the same algorithm that to3d's quality check does, by the way.

Historically, 3dTqual was first. 3dToutcount was developed later, and is probablly better, since its "-save" option lets you actually flag the outliers.

The related program 3dDespike actually tries to repair large isolated spikes in 3D+time datasets. It was written to patch over a hardware problem we had with one of our scanners. I don't strongly recommend it, but it can be used when you have a dataset that has a few very corrupted data points that will throw the 3dDeconvolve statistics into a tizzy.

bob cox