If every sub-brick outlier count is 100, then the median is 100 and the MAD is 0, so the threshold would be 100. In to3d, none of the counts are strictly larger than the threshold, so none of the sub-bricks would be considered to have "a lot" of outliers.

If each outlier count is 10, except for one value of 20, then the median count is 10 and the MAD is 0 (MAD = median absolute deviation; in this example, the absolute deviations are all 0 except for one value of 10, so the median of these is 0 -- if there are at least 3 sub-bricks, that is). So the threshold is 10, and that one sub-brick is considered to have "a lot" of outliers.

The presumption is that most sub-bricks are "good", but that some are "bad".
If you want more precise control over outlier detection, you can use the 3dToutcount program, which can label each voxel in the dataset with a number that indicates its "outlierness".

The original purpose of this was to detect some spiking problems that we have intermittently had with our 3 Tesla scanners here. Later, the 3dDespike program was written to remove spikes.

bob cox