 |
 |
Kamila Szulc wrote: > Perhaps someone could comment on the ways they use to normalize image > intensities between subjects in case of voxelwise analysis Hi Kamila, I rescaled a bunch of images to have the same median intensity within a certain region, and this seemed to give quite good results for PCA etc on T1 weighted images. It sounds rather basic compared to histogram equalisation, but I found that more complicated methods didn't seem to give as sensible looking results (but your mileage may vary, as they say...) More specifically, based on some advice that intensities in the Pons region of the human brain would be relatively unaffected by Alzheimer's Disease (compared to other regions, which apparently may change in T1 intensity due to the disease process), I decided to rescale the images so that all subjects/timepoints had a fixed median pons intensity. In practice, what I did was to compute registration matrices from all images to avgT1 with flirt (-dof 12), then apply the inverse transforms to a manually created mask that I made in the avgT1 space (using fslview), re-binarized this mask, and then divided all image intensities by the median of the voxels within this mask (I think I did this with Matlab/SPM, but you should be able to do the same with a few avwmaths commands). Also, since I couldn't guarantee that the registered region would fall entirely inside the pons (and because I was a bit lazy to check on the large number of images that I had!), I plotted a graph of coefficient of variation (std/mean within region) against scan, and this flagged up a few scans where the region included some non-WM tissue. Not a brilliant method, but reasonable, and quick to do... I also considered affine (or higher-order polynomial) rescaling of intensities, e.g. instead of simply making one regional value fixed, a scaling and offset could be computed to fix two regional values, or n regional values fixed with an (n-1)th order polynomial. E.g. I looked briefly at fixing the medians over (FAST/SPM) GM/WM/CSF segmentations. But it seemed to me that I was getting better results from a simple rescaling... Hope that helps, Ged.