 |
 |
> Thanks for your help. Unfortunately, reslicing does not fix the > problem. I am able to get DBM to run though. It still gives me that > strange output of Xs, however, and the jy_image that it outputs is > not correct. The jy_image looks like a plaid pattern with values > ranging from .999962.... to 1.0000439... There is no brain shape. > When I run the same images using SPM2, the output seems to look as it > should, however the inhomogeneity in the images seems to influence > the deformations. I believe that SPM5 has additional features for > DBM including a homogeneity correction, so I would like to be able to > use the SPM5 version if possible. I don't think the bias correction is the problem. > I have run several of my images > with both SPM2 and SPM5 and get this same pattern of results. I'm confused. I thought it was working OK in SPM2. > I have > also tried different inhomogeneity parameters in SPM5 and still get > the same results. How do the images appear if you do a check reg? Also, check the voxel sizes via the Display button. If one gives a +ve x voxel size, and the other gives one that is negative, then this could cause the sort of problem you're experiencing. Checking the scalefactors and offsets may also be helpful. Are the voxels in the background (air) around an intensity of zero (or slightly higher), and are intensities within the brain larger and positive. > > Also, regarding inhomogeneity correction, I am wondering as to the > validity or advantage of doing iterative inhomogeneity corrections. > I came across a posting on the listserv saying that bias correction > would be most accurate if done iteratively during segmentation, but > that the current version of SPM does not do this. Bias correction is done iteratively within the segmentation algorithm. > I thought to try > this: > > 1) Spitting out a bias corrected (m* image) from the segmentation/ > normalization procedure. > 2) Rerunning segmentation/normalization with the bias corrected image > and spitting out a bias corrected image > 3) Rerunning segmentation/normalization with the newest bias > corrected image and spitting out a bias corrected image, etc. This would give slightly different answers because of the regularisation that is used. There is a penalty on the estimated bias that penalises its deviation from zero. If you repeatedly bias correct the last bias corrected image, then this panalty has less and less effect. > > I ran this iterative procedure on one of my images pre-treatment and > post-treatment and the difference image between the two realigned > images was much improved (i.e. the magnitude of the differences were > smaller and more consistent spatially) when compared with the > realigned images after using just one bias correction. You could try reducing the regularisation used by the bias correction. This will give the model more flexibility, which may be a good thing for your data. > From this > empirical observation, I would recommend performing an iterative bias > correction for images with lots of inhomogeneity, however, wanted to > make sure this would not introduce other errors or bias into the > data. Any thoughts? The default regularisation was set to a value that worked fairly well for images from the FIL. In other centres, there may be more or less bias artifact. If there is zero bias artifact, then it would be better not to correct. If there is only a tiny amount, then a more heavily regularised bias correction would do a better job. For lots of bias, you would decrease the regularisation. Best regards, -John