I have investigated the issue a bit longer and figured out that this happened because of the way featquery does its weighting. As it says in the manual: the mask (weights) is (are) scaled such that the mean (over non-zero voxels) is 1. So in fact, the mask is divided by its average value before multiplying with the PE values. This behavior can result in something that is eerily close or even identical to a binary mask when using a thresholded mask. For example, when using a mask containing anatomical probability values that is thresholded at 50% (values below 50% thrown out), all weighting is in fact abolished. The mean of the range [50-100] is about 75, and dividing the values 50-100 by 75 leads to a range of .66 to 1.33. Considering that integer values in these masks are retained, you get back a binary mask. This may not be what people expect to happen when not having thoroughly thought about the meaning of the above phrase in the manual. Maybe the division operation should be made optional, such that in cases like these one can still have differential weighting.
The same applies to the use of thresholded stats images for masks, for example using a range of Z = 1.6 running up to about Z = 3 leaves weights in the range of about .7 to 1.3 which may not be what people expect when their lowest value is twice as small as their highest value.
Thanks,
Johannes
|