 |
 |
Hi Moo, Historically, smoothing the statistic images has been ill-advised since it will alter the null distribution of the statistic. E.g. a t(10) image will certainly not be a t(10) after smoothing, but it's not clear what it should be. I guess the other motivation is thinking about the best model and the units for that model. If you think of the need for smoothing as saying that 'We really wish we had a model for our data that enforced contiguity', you can ask in what units that contiguity is best expressed. I think the standard approach is that the contiguity is on the original data in original units, and so the original data should be smoothed. If you can instead argue that contiguity is more naturally expressed on the parameter estimated (e.g. correlation coefficient) and you can deal with the distributional headaches, then I guess you could make a case for smoothing the statistic map. At any rate, it's worth noting that in some special cases (spatially pooled variance) the statistic image is a linear function of the data, and then it won't matter when you smooth. Take care! -Tom On Dec 3, 2007 4:18 PM, Moo K. Chung <[log in to unmask]> wrote: > I would like to know if some of you had a similar problem or dilemma. > > I am computing a correlation map out of a linear model at each voxel. > I always thought smoothing images first and getting correlation map > is a more or less correct procedure. However, my student computed > the correlation out of unsmoothed images and smoothed correlation map > later. > The smoothed correlation map looks better ! > I thought about this a while and I have to conclude that smoothing > correlation > map is not a bad idea actually. > > Smoothing is a process of degrading information. So the question is > where to degrade information. Do we want to degrade information in the > original image > or correlation. If we compute correlation out of smoothed images, it > is like computing > correlation out of something that is already degraded. So smoothing > correlation seems > to make more sense. But then you worry about random field theory > assumptions > if you are a random field guy. :) > > ____________________________________________ Thomas Nichols, PhD Director, Modelling & Genetics GlaxoSmithKline Clinical Imaging Centre Senior Research Fellow Oxford University FMRIB Centre