Dear John and all other experts
In the meantime I test two ways of segmentation using the data of a stroke
patient. Attached you can find some images.
1. version: I masked the lesion using the anatomy of the first measurement.
I enter this mask during the segmentation of the anatomy for the first and
for the second session (sessions are 2 months apart). The lesion was
properly masked out during the first measure, so that there are not gray
matter in this area (image: P04_FM_2mess_SegMasked). However when seeing the
gray matter results of the second session (image: P04_FM_3mess_SegMasked)
you can see gray matter in the region of the lesion. How can this be
possible?
2. version: as Christine propose I used the segmentation procedure of the
VBM8-toolbox. The gray matter results shows still some gray matter in the
region of the lesion (image:P04_FM_2mess_SegVBM8), while the results in the
second session seems to be much better (image:P04_FM_3mess_SegVBM8). Could
the strange results during the first session be due to the patient
movement? When you look at the anatomy of the first measurement
(image:rP04_FM_2mess_Anatomy), this image is not so nice compare to the
anatomy in the second measurement (image:rP04_FM_3mess_Anatomy).
I read the paper of Seghier you recommend me but I couldn't see that there
was a link to get this additional tissue class or some kind of toolbox, so
that I can test this segmentation procedure. Is there anything?
For your help I thank you very much.
All the best,
Natalia
-----Ursprüngliche Nachricht-----
Von: Natalia Estévez [mailto:[log in to unmask]]
Gesendet: Mittwoch, 21. September 2011 17:23
An: [log in to unmask]
Betreff: WG: [SPM] Dartel template for normalizing function image comment
Dear John
Thank you very much for your answer. I will have a look at the paper and see
if it helps.
It is probably a stupid question, but in case that I have to mask the lesion
out, can I still use Dartel to process my data (functional and anatomical)?
Or is it only possible to use Dartel when you have the whole brain?
I thank you very much again for your help.
All the best,
Natalia
-----Ursprüngliche Nachricht-----
Von: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] Im
Auftrag von John Ashburner
Gesendet: Mittwoch, 21. September 2011 16:49
An: [log in to unmask]
Betreff: Re: [SPM] Dartel template for normalizing function image comment
Adding to some of Jonathan's comments....
> For structural image analysis, you would just normalize all of the
> rc1* (or rc2*, for white matter) files to MNI space, and conduct your
> statistics on those.
Results should be marginally better if the rc1 and rc2 images are used for
the estimation stage, and the deformations applied to the c1 and c2. This
saves an additional resampling of the data.
>> 2. I’m analyzing data of stroke patients and when I use the “New
>> Segment” option the lesion is also shown as gray matter and some
>> parts of the lesion as white matter. But this happens with the
segmentation too.
>> Attached you can find a picture, the first two images are the c1 and
>> c2 after “Segmentation”. The next two images are c1 and c2 after “New
Segment”
>> and the last image is the anatomy of the patient, where you can see
>> the lesion. Can these files be used or is it maybe better to mask out
>> the lesion?
>
> I did not see the attachment, so I can't comment on your particular
> case. Also, I have not worked with stroke patient data before, so
> hopefully someone else can comment more knowledgeably. However, it
> does not surprise me that segmentations might fail (or be a bit odd)
> on these images, because they do not conform to what the segmentation
> algorithm "expects" (based on tissue probability maps). You may want
> to have a look at Seghier et al. (2008), who found that including an
> additional tissue class to model the lesion was successful.
> Otherwise, I would think that masking out the lesion is the best
> approach (in the case where the standard approach is not working).
>
> Seghier ML, Ramlackhansingh A, Crinion JT, Leff AP, Price CJ (2008)
> Lesion identification using unified segmentation-normalisation models
> and fuzzy clustering. NeuroImage 41:1253-1266.
The SPM segmentation model is pretty simple and assumes that brains consist
only of GM and WM. There is no part of the model for dealing with regions
affected by stroke. The behaviour in these regions will depend on how the
stroke changes the signal intensity. Typically, affected WM will have an
intensity closer to that of GM, so is more likely to be segmented as GM.
To achieve a really good automatic segmentation of stroke lesions etc, would
need a useful dataset for training a segmentation algorithm.
Basically, lots of scans and some manually defined lesion masks. Even then,
it may be a bit tricky to make the approach generalise to MR scans that
differ from the training data.
Best regards,
-John
|