[Brains-users] Some questions regarding segmentation
Eugene Zeien
eugene-zeien at uiowa.edu
Wed Jan 25 06:17:41 PST 2006
http://www.psychiatry.uiowa.edu/pipermail/brains-users/
Christoph Christmann wrote:
> Dear Ronald, dear Eugene,
>
> thanx a lot for the very detailed answers. I will definitely go all the
> recommended steps and report
> any problems that cannot be solved by your answers or the manual. Is
> there a kind of periodical FAQ for the mailing list?
>
> Best regards
>
> Christoph
>
> At Tuesday 24.01.2006 19:50, Ronald Pierson wrote:
>
>> Okay, sounds like you are on the right track. The brain box is
>> intended to
>> be the size that you show, and there is a reason why. In our
>> segmentation
>> routine we let BRAINS pick plugs throughout the brain as a means of
>> sampling
>> the tissue types. We chose this set of boxes because we want to only
>> pick
>> those areas that we are 100% sure are brain. This is undersized, so
>> we know
>> that what is in here will be brain or CSF, not bone, fat, etc. That
>> makes
>> sure that the training classes used for classification are the tissues we
>> want. Next, a set of discriminant functions are created to classify the
>> image, based on those training sets. Then, the entire image is
>> classified.
>>
>> Sounds like what you have done to generate the brain will do the job.
>> The
>> brain mask that we use is generated with an artificial neural network,
>> that
>> was trained on manually traced brains. The accessed through >workup >run
>> neural >MR5/MR6 > Brain.
>>
>> We added a second segmentation step a few years ago. This takes the ANN
>> brain mask, cleans it a bit with several routines, then reruns the
>> segmentation using the brain mask as the region for picking plugs. This
>> helps get a more representative set of tissue plugs and helps account for
>> intensity fall-off in the posterior and inferior regions. This is
>> invoked
>> with the little button "bootstrap" on the last screen of the segmentation
>> GUI.
>>
>> I would guess that if you use both the T1 and T2 for segmentation you
>> should
>> end up with a good image, but if you don't, here are a few suggestions:
>>
>> The T1 and T2 must be well registered. This is the cause of many
>> problems
>> in segmentation. Check spots on the arterior regions, placing the
>> crosshairs at some clear borders on the T1 and T2 image. Toggle
>> between the
>> two (using the spacebar or the Selector) and make sure the borders are in
>> the same place on both images. Also check the cerebellum.
>>
>> The other main problems with the segmentation routine itself are with the
>> sampling of the tissues, or as we call it, plug picking. If you load the
>> csf_plugs.mask, grey_plugs.mask and white_plugs.mask, you will see
>> what has
>> been picked. The main things that affect this are set in the first
>> window
>> of the BRAINS2 segmentation GUI. The ones I could suggest changing are
>> permissiveness and trim mean outlier. Permissiveness refers to the first
>> step in purifying the plugs (training classes). For each plug of 8
>> voxels
>> the variance is found. If that is above a certain level, the plug is
>> considered to NOT be pure tissue, and it is discarded. Permissiveness
>> refers to how much variance we allow, the larger the number the more
>> variance. If your images are speckly or poorly registered, you may
>> need to
>> increase it.
>>
>> After the purity of the plugs is tested, they are grouped by k-means
>> clustering into the tissue classes. Next, the outliers are removed from
>> each set of plugs (now called tissue training classes). Trim Mean
>> Outlier
>> (TMO) refers to how many SD from the mean (in each image type) at
>> which the
>> plugs are removed. If you remove too much (a TMO that is too low) there
>> will be unclassified areas (black voxels) on the borders between CSF,
>> Grey
>> and white - it won't know what to do with partial-volumed voxels. If you
>> remove too little (a TMO that is too high) you may end up with fuzzy,
>> indistinct borders between the tissue types.
>>
>> So, I suggest
>> 1. Use both T1 and T2
>> 2. Check registration
>> 3. Look at the plugs and see if they really represent the tissues.
>> If not,
>> try changed in the permissiveness and the trim mean outlier.
>>
>> Long answers to a short question, but I think it is helpful if one
>> looks at
>> the process - you can sometimes pick up on why there are problems, though
>> sometimes it is difficult.
>>
>> Feel free to email screen shots of your results, they are quite
>> helpful to
>> my understanding the problem.
>>
>> Ron
>>
>>
>> -----Original Message-----
>> From: brains-users-bounces at psychiatry.uiowa.edu
>> [mailto:brains-users-bounces at psychiatry.uiowa.edu] On Behalf Of Christoph
>> Christmann
>> Sent: Tuesday, January 24, 2006 11:10 AM
>> To: brains-users at psychiatry.uiowa.edu
>> Subject: RE: [Brains-users] Some questions regarding segmentation
>>
>> Dear Ronald,
>>
>> Thanks a lot for your suggestions!
>>
>> We obtain only T1 and T2 data but at least this is mandatory for all
>> of our anatomic data. So we easily can include T2 data, too. I wasnt
>> quite sure if the data was used for the segmentation, too, even
>> though I was able to specify the T2 data in the segmentation dialog.
>> I used the T2 data for the blood vessel definiton, too. I'll go
>> trough all steps in the manual again and try the segmentation without
>> prior skull stripping.
>>
>> One question, however, remains: In the segmentation dialog there is a
>> specification of a standard box (talairach) which is too small for
>> all our brains (I couldnt figure out the reason for that - I append a
>> picture). Alternatively I used a mask (concurrent option) that I
>> created with a skull stripped image (by means of the image tools-
>> threshold option). Is there an alternative algorithm? What could be
>> wrong with the box?
>>
>> Thanks again!
>>
>> Sincerely
>>
>> Christoph
>>
>>
>> At Tuesday 24.01.2006 15:29, Ronald Pierson wrote:
>> >I suggest you take some time to go through the Standard Workup manual
>> >that can be found on the WIKI page of our web site.
>> >
>> >http://www.psychiatry.uiowa.edu/wiki/index.php?title=BRAINS_GETTING_STAR
>> >TED
>> >
>> >BRAINS does a full workup, including creation of a fantastic mask of the
>> >brain. We don't call it skull stripping, it really is more focused on
>> >careful definition of brain and CSF.
>> >
>> >Standard workup is normally done with a T1 and T2, with addition of a PD
>> >if it is available. A T1-only classified image is really not the best,
>> >no matter what package is used to create it.
>> >
>> >A quick summary of our workup -
>> >
>> >1. The first step is to realign the T1 (AC-PC), pick Talairach
>> >parameters, then coregister the T2 to the T1 (using alignlinear or
>> >mutual information coregistration from the toolbar).
>> >
>> >2. Then, blood is traced. Tissue classification (we also call this step
>> >segmentation) is completed using both a T1 and T2 so that there is a
>> >very clear definition of brain vs nonbrain. If you us only a T1-like
>> >image, you will need to use a different designation for blood, maybe the
>> >250 voxels from the top of the image (outside of the image space). That
>> >is because blood is very similar in intensity to gray matter on a T1,
>> >and you can't make any distinction to discriminate between blood and
>> >gray matter without more information from the T2.
>> >
>> >3. At that point, and actually as part of tissue classification, a mask
>> >is generated that is analogous to what you refer to as skull stripping.
>> >We go the extra step to manually edit this brain mask before measuring
>> >brain volumes. We really feel that this brain mask is far superior to
>> >anything else currently available.
>> >
>> >FYI - we actually don't like the use of basal gray and basal white any
>> >more. There really is too much variability between traces in what gets
>> >picked.
>> >
>> >Ronald Pierson
>> >
>> >
>> >-----Original Message-----
>> >From: brains-users-bounces at psychiatry.uiowa.edu
>> >[mailto:brains-users-bounces at psychiatry.uiowa.edu] On Behalf Of
>> >Christoph Christmann
>> >Sent: Tuesday, January 24, 2006 3:02 AM
>> >To: brains-users at psychiatry.uiowa.edu
>> >Subject: [Brains-users] Some questions regarding segmentation
>> >
>> >Hi all,
>> >
>> >At first I am facing a serious problem during the segmentation. I did
>> >skullstrip, import, resample, crop and lighten up a FLASH image. Then
>> >I created a mask of the brain (with image tools - threshold) the
>> >Talairach parameters, and sampled ROIs for venous blood and basal
>> >white and grey matter. Images, ROIs and masks show up in the
>> >segmentation dialogs (i.e. no appearance of a '-1'). However, after
>> >running segmentation (without BOOTSTRAP) the tissue qualification
>> >seems shifted, for example white matter will completely be qualified
>> >as grey matter. Has anyone an idea what's inducing this behaviour and
>> >how this can be fixed?
>> >
>> >A second question is concerning the skull stripping. Is there any
>> >tool apart from MRIcro that could be used effectivly to separate the
>> >brain?
>> >
>> >The last question applies for the use of T2 images. While it seems
>> >quite reasonable to integrate the T2 images for the ROI analysis I
>> >wonder if the T2 image is used for the segmentation of the T1 image.
>> >
>> >Thanks in advance for any hint!
>> >
>> >Christoph
>> >
>> >
>> >
>> >
>> >
>> >--
>> >Christoph Christmann christma at as200.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 Neuropsychology D-68159 Mannheim J5
>> >and Clinical Psychology
>> >
>> >************************************************************************
>> >**********************
>> >IMPORTANT: The contents of this email and any attachments are
>> >confidential. They are intended for the
>> >named recipient(s) only.
>> >If you have received this email in error, please notify the system
>> >manager or the sender immediately and do
>> >not disclose the contents to anyone or make copies thereof.
>> >*** eSafe scanned this email for viruses, vandals, and malicious
>> >content. ***
>> >************************************************************************
>> >**********************
>> >
>> >_______________________________________________
>> >Brains-users mailing list
>> >Brains-users at www.psychiatry.uiowa.edu
>> >http://www.psychiatry.uiowa.edu/mailman/listinfo/brains-users
>>
>> --
>> Christoph Christmann christma at as200.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 Neuropsychology D-68159 Mannheim J5
>> and Clinical Psychology
>> *** eSafe scanned this email for malicious content ***
>> *** IMPORTANT: Do not open attachments from unrecognized senders ***
>> ****************************************************************************
>>
>> ******************
>> IMPORTANT: The contents of this email and any attachments are
>> confidential.
>> They are intended for the
>> named recipient(s) only.
>> If you have received this email in error, please notify the system
>> manager
>> or the sender immediately and do
>> not disclose the contents to anyone or make copies thereof.
>> *** eSafe scanned this email for viruses, vandals, and malicious content.
>> ***
>> ****************************************************************************
>>
>> ******************
>
>
More information about the Brains-users
mailing list