[Brains-users] brain mask - problems with automatic definition

[Christoph Christmann]

okay, now it seems to me that training the neural net could be a way 
to get better results. We have about 50 subjects with segmentations 
and correct brains masks. Is there a way to feed all this to the 
neural net? Or to make it learn these masks? We would of course 
provide all informations to others if requested.

christoph


At Monday 19.02.2007 17:16, Eugene Zeien wrote:
>I have, in fact, had very similar problems.  "standard workup"
>is not a part of my daily routine, but I needed to generate a
>brain mask for a project.  Rather than ask a more experienced
>person to create one, I decided "I can do this".  Then I
>proceeded to do just about everything wrong. :-)  One by one,
>I found the problems and fixed them.  Eventually (about 1 day)
>I had the neural net making pretty brains.
>
>The basic issue I see with your mask is:
>The neural net is not finding the expected anatomical boundaries.  So,
>the end result is merely the probability
>mask where more than 50% of the subjects used to train the
>neural net had a region identified as brain.
>
>Why?  Well, there can be several causes.
>   incorrect talairach bounds
>   over-inclusive blood traces which lead to
>   bad segmentation
>   incorrect image specified for the neural net's use
>
>There may be other ways to confuse the neural net, but these
>are the things I had done.  Of those problems, you've checked
>three.  The only one left is if the segmented image has CSF
>classified as blood.  So, load up the image, then
>   Viewers->Tracker
>and set the crosshairs in brains2 to somewhere in the
>ventricles.  The values (poke around several places) should
>be around 30.  If you see the values showing up as 0 or 1, then
>something has gone wrong in the segmentation.  The values
>outside the brain/head should be 0.  You should only find 1's
>where major blood vessels exist.
>
>On Mon, 2007-02-19 at 10:15 +0100, Christoph Christmann wrote:
> > hi gene,
> >
> > we always used proper talairach bounds. here, we just added the image
> > to some former work where we used the 'talairach box' lines for
> > different orientation purposes.
> >
> > are there any further hints regarding the brain mask? dont you have
> > similar problems?
> >
> > christoph
> >
> > At Friday 16.02.2007 16:52, you wrote:
> > >Christoph,
> > >    Are the red lines on the images you sent the Talairach bounds?
> > >If so, they are nowhere near the correct location.  Try re-defining
> > >the talairach boundaries, then run the neural net again.
> > >    I've attached a picture showing proper talairach boundaries.
> > >The 2 points near the center are supposed to be the anterior and
> > >posterior commissures.
> > >gene
> > >
> > >On Fri, 2007-02-16 at 12:35 +0100, Christoph Christmann wrote:
> > > > hi,
> > > >
> > > > the segmented image looks reasonable. I attach some pictures
> > > > elucidating the actual situation. I am especially surprised that the
> > > > cerebellum border is not found correctly although it seems to be
> > > > obvious as you can see in the brain_mask_cor_seg.jpg file. You can
> > > > see in the other two images, that T1 and T1 is realigned quite well
> > > > (cost function < 0.45).
> > > >
> > > > If you would have any idea to use other programs to get the brain
> > > > mask I would try this way, too. Actually, we started with BrainSuite
> > > > but found it to be less precise, too - although in other regions.
> > > > Maybe we end up with an SPM or FSL segmentation for the mask.
> > > >
> > > > Best regards
> > > >
> > > > Christoph
> > >
> > >
> >

-- 
Christoph Christmann                  christoph.christmann at zi-mannheim.de
Psychologist, MEE                     Fon/Fax +49 621 1703 63 18/05
Central Institute of Mental Health    http://www.zi-mannheim.de
Department of Clinical                D-68159 Mannheim J5
and Cognitive Psychology