Anthony Ang wrote:
> Hi Steve,
>
> I had obtained GM & WM volumes of the various brain lobes, cerebellum
> and GM volumes for caudate, putamen, insula & thalamus of subjects by
> registering MNI152-maxprob-thr0 to the subject's native space and
> using the subject's fast-segmented GM and WM binary images to mask out
> the unwanted regions.
>
> As I continue to face problems of low reliabilities (could either be
> due to poor fast-segmentation into GM/WM or less precise fnirt
> atlas-to-subject registration), I have to come back to this mailing
> list to seek your advice.
>
If you believe the unregistered atlas - reading on it sounds like you
don't :( - then you must double check your registrations. Wouldn't it be
better to transform data to standard space for classification and
comparison?
> I notice that the 3 MNI atlases do not cover all the WM for the
> different brain lobes. The MNI152-maxprob-thr50 atlas covers the least
> amount of WM for the different brain lobes, followed by the
> MNI-maxprob-thr25 and then MNI-maxprob-thr0 atlas.
>
The probabilistic atlases are determined by the results of expert
segmentations of one kind or another, e.g., the MNI atlas is a hand
segmentation of a single subject which was then non-linearly registered
to 50 example structurals, the 50 samples were linearly transformed into
MIN152 space and averaged - no really! See
http://www.fmrib.ox.ac.uk/fsl/data/atlas-descriptions.html for more
details.
In the FSL bundled atlases each voxel value may be interpreted as the
percentage of segmentations which agreed that the voxel had a particular
label.
> Even the MNI-maxprob-thr0 atlas has missing WM for the brain lobes
> towards the centre of the brain (around the ventricles and caudate).
> Hence, I have concerns that the WM volumes that I obtained for the
> different brain lobes are not accurate and are less than the actual WM
> volumes, thereby contributing to the variation in the results that I
> had obtained. Is my concern valid and can you suggestion any way to
> overcome this problem?
>
If you're convinced that the unthresholded atlas has insufficient
coverage then you will need a different atlas. Have you looked at the
Harvard-Oxford one?
> Also, I don't really understand what is meant by the atlas thresholded
> at 0, 25 and 50%. Can you explain this to me and what are the possible
> situations/applications you may need to use either the 0, 25 and 50%
> thresholded atlas? In other words, what circumstances would determine
> the choice of the threshold you should use?
>
The thr0 atlas contains all the labels even if evidenced by only a
single segmentation, thr25 only those labels where more than 25% of the
segmentations agreed, and thr50 more than 50% of them agreed. Choice of
atlas would be based on how little evidence you're willing to accept for
a given classification.
> Thank you very much.
>
> Best regards
> Anthony
>
>
> On Tue, Oct 28, 2008 at 6:22 PM, Steve Smith <[log in to unmask]> wrote:
>
>> Hi,
>>
>> NO! the atlas is already in the correct space of the MNI152 - it's just dilated because you're looking at the thr0 version. You can get exactly what you want (more 'average' masks by either using one of the more sensible thresholding images (thr25 or thr50) or just by taking the full unthresholded 4D probability images and thresholding them yourself - but that shouldn't be necessary. In general however I might recommend using the Harvard-Oxford atlases instead as these have more detail in them.
>>
>> Cheers.
>>
>>
>
>
|