Dear Will,
On Mon, 28 Mar 2016 15:05:50 +0000, William McGeown <[log in to unmask]> wrote:
>Hi, sorry to repost. I didn't receive a reply before and was hoping someone could provide some advice.
>
>After doing segmentation using CAT12 (on patients
>with Alzheimer's Disease) and examining the GM segments afterwards I have
>noticed that some WM appears to have been misclassified as GM in a small number
>of participants (but in enough participants that I wouldn't want to discard
>them readily).
>This appears to be a problem that is not unique to CAT12, and that is sometimes encountered using a variety of segmentation methods.
>
>1) I was wondering whether better segmentation of GM in these cases might be achieved via alternative settings in the CAT12 segmentation?
>Might altering the denoising or local adaptive segmentation defaults help?
>If so, what values would be recommended for this issue?
>
Are you referring to WM that is affected by white matter hyperintensities (wmh)? WMH are often occurring in brains of AD patients (but also in brains of old healthy control) and are characterized by lower intensitiy of WM in a T1 image. If this is the case the lower image intensity in WM that is rather that of GM will mostly lead to misclassification. If you have an additional FLAIR image you can try to use the lesion segmentation toolbox:
http://www.applied-statistics.de/lst.html
However, you can also try the experimental function in CAT12 to correct for WMH. For this purpose change the option for WMHC in cat_defaults.m to "2":
cat.extopts.WMHC = 2;
This will fill the lesions in WM.
>2) If I adjust the settings - I take it that I should segment the entire dataset again with those settings to be consistent?
If you change the WMHC option this should be done for the entire dataset to prevent bias. However, as I mentioned this option is experimental and you should check your data carefully.
>
>3) Also I wondered how much of a concern misclassifications of GM regions would be practically? The misclassification mostly relates to periventricular regions, but is also present in some deeper WM.
>In the model estimation stage a GM mask will be created - which could potentially remove the contribution of those misclassified regions in the statistical analysis(?). I could check the mask to see whether it includes those regions. Opinions would be appreciated.
If these misclassifications only occur in some but not all subjects this will lead to very large local variance in these regions (some subjects have GM, most not) which will largely minimize sensitivity to find something. Furthermore, if you use an absolute threshold for masking (which is recommended) values of 0.1 or above will usually exclude those regions from analysis. My personal experience with the ADNI dataset is that even if WMH occur frequently this will not affect effects in other regions because of the large local variance and the masking. However, there might be worst case scenarios where WMH more often occur in one group and can then result in false positive effects.
Best,
Christian
>
>Thanks,
>
>Will
|