Dear Gabriel,
> taking the mean of the voxels included in the voxels would act as a 2nd smoothing step ?
To reduce Gaussian noise in digital images the (arithmetic) mean filtering would be a simple filter, with the pixel (or voxel) value being replaced by the arithmetic mean of its own value and that of neighbouring pixels (e.g. the "8-connected" pixels). There's also the simple moving average for time courses. Now in your case, taking the mean (or, depending on purpose, the sum) of the cluster values is somewhat different, as there's no moving "window" or "grid" across which you average, but it should be alright.
Note that smoothing does not just eliminate noise, but also introduce blurring. E.g. if you have a very precise mask for say, the thalamus, and a very good segmentation with every voxel within that mask having the GM value of 1, then the GM volume (assuming these values represent volume and not probabilities) is just no. of voxels x 1. If you applied the mask to smoothed data, the GM volume would be a different (and incorrect) one due to the blurring around the thalamus boundaries. Working with a smoothed mask should help then. Which is also interesting with regard to labeling activations and such, as data might be smoothed more strongly than the labels. Thus people might conclude that peak 123 corresponds to thalamic nucleus xyz, overseeing that their data is much more imprecise.
> that using smoothed values would be scientifically better because these are the data I had significant cluetsres. Am I correct
It's more consistent. If you ran a sort of localizer analysis on smoothed data, then you should also probably rely on what you get.
However, there's the danger of introducing biases. The voxels you obtained are not just true-positive ones, but also false-positive ones (noise "lifting" the voxels significant); and you also didn't detect every voxel with a true effect (due to noise making some voxels non-sig.). Besides, while your approach sounds reasonable at first (and also finds application frequently) you might miss interesting regions. Assume there's a group difference in region 1, but also large variance due to variable X. Due to the latter that region might not show up significantly in your initial model. Or another example, quite frequently seen when it comes to intersubject variability / single group: Start with a first group model to find regions with a sig. difference "task A vs. task B", then test whether these regions correlate with a behavioral measure or a personality/questionnaire score. Now there might be regions which are activated by a subset of subjects only (e.g. those with a high score), and accordingly, they might not show up in the initial contrast then. In fact, a region might be activated more strongly in every single subject, but if the variance is too high corresponding voxels wouldn't reach significance.
Best
Helmut
|