[MINC-users] Classify

[Jason Lerch]

On 26-Mar-09, at 8:39 AM, Jon Erik Ween wrote:

> Thanks Jason
>
> So, I tag 100 white matter, 100 grey [cortical and deep] voxels,
> fairly evenly distributed throughout the brain) using Display, with
> different label values for each tissue type (does the value matter?)
> and save as a .tag file?

That looks correct - and yes, the value does matter: you want it to be  
an integer which will give the corresponding value to the classified  
file. We usually use 0 for background, 1 for CSF, 2, for WM, 3 for GM.  
But that's merely convention - though a few scripts assume these values.

>
>
> You do this on the intensity normalized volume and not the native
> image, right?

Yup.

>
>
> Is this process sensitive to changes in scanner software? We upgrade
> the magnets from time to time.

If you pick tags each time then it is not sensitive to scanner  
software. If you train a classifier and reuse it then it is sensitive  
to scanner software (maybe - depends on whether the upgrade affected  
signal intensities with the sequence you are using). If you use  
spatial priors then it is once again not sensitive to scanner software.

>
>
> Where do I find an appropriate probability maps?

The probability maps are not distributed by the MNI as far as I can  
tell - but the spatial priors derived from those probability maps are  
part of the classify packages (see models/ntags_1000_bg.tag). These  
assume alignment to the MNI version of Talairach space - you ought to  
be able to use these without modification. They are based on 1000  
points per tissue class, each point having a 90% or greater likelihood  
of being of that tissue type.

If you want the probability maps themselves you'll have to contact  
somebody at the MNI (Alex Zijdenbos, I suppose). It's also not too  
hard to generate the probability maps yourself - classify a bunch of  
scans, separate out the individual tissue types, then use mincaverage  
on the binary maps to create your probability map. Once you have that  
you can use the extracttag command to get your tag point set.

Cheers,

Jason


> I'm supposing the MNI
> averages are not the ones. Do the autoreg models work for this?
>
> Thanks
>
> Jon
>
> Soli Deo Gloria
>
> Jon Erik Ween, MD, MS
> Scientist, Kunin-Lunenfeld Applied Research Unit
> Director, Stroke Clinic, Brain Health Clinic, Baycrest Centre
> Assistant Professor, Dept. of Medicine, Div. of Neurology
>    University of Toronto Faculty of Medicine
>
> Kimel Family Building, 6th Floor, Room 644
> Baycrest Centre
> 3560 Bathurst Street
> Toronto, Ontario M6A 2E1
> Canada
>
> Phone: 416-785-2500 x3648
> Fax: 416-785-2484
> Email: jween at klaru-baycrest.on.ca
>
>
> Confidential: This communication and any attachment(s) may contain
> confidential or privileged information and is intended solely for the
> address(es) or the entity representing the recipient(s). If you have
> received this information in error, you are hereby advised to destroy
> the document and any attachment(s), make no copies of same and inform
> the sender immediately of the error. Any unauthorized use or
> disclosure of this information is strictly prohibited.
>
>
>
> On 25-Mar-09, at 12:51 PM, Jason Lerch wrote:
>
>> Hi Jon,
>>
>> classify is based on using a training set to initialize the tissue
>> classification - either based on probability maps or on manually
>> selected tags. In other words, you could use Display to pick 100
>> points or so that you think are grey matter, 100 that are white
>> matter, etc., then use that to initialize the classification of every
>> voxel. Once you've trained the classifier once you can also save the
>> training data and apply it to new volumes without repicking  
>> individual
>> points (though this is quite scanner specific - i.e. it usually works
>> for one sequence from one site, not across sequences or sites - for
>> that probability maps are better).
>>
>> classify_clean uses classify but prunes those initial points first to
>> remove any outliers - this can be a benefit when you use spatial
>> priors based on probability maps.
>>
>> Hope this helps,
>>
>> Jason
>>
>>
>>
>>
>> On 25-Mar-09, at 12:18 PM, Jon Erik Ween wrote:
>>
>>> Dear minclist.
>>>
>>> Sorry for the incremental posts, but then, how is one to learn??
>>>
>>> Anyway, thanks to help from the list, I'm now in the possession of
>>> skull-stripped, talairach registered, lesion masked 3DT1 images of a
>>> group of patients and would like to tissue classify these volumes
>>> (and
>>> hopefully label major lobes etc) so I can calculate tissue and  
>>> lesion
>>> volumes (globally and in particular lobes, etc). Running
>>> "classify" (or classify_clean, I can't figure out the difference
>>> between these two, exactly) complains that I don't have a tag file
>>> (or
>>> in the case of classify_clean, that the tags are not in the
>>> volume). I
>>> was thinking that "classify" used the specified model to figure out
>>> which voxels were which tissue, or do I need to manually tag voxels
>>> in
>>> each of the target volumes?
>>>
>>> I, again, appreciate any guidance!
>>>
>>> Cheers
>>>
>>> Jon
>>>
>>> Soli Deo Gloria
>>>
>>> Jon Erik Ween, MD, MS
>>> Scientist, Kunin-Lunenfeld Applied Research Unit
>>> Director, Stroke Clinic, Brain Health Clinic, Baycrest Centre
>>> Assistant Professor, Dept. of Medicine, Div. of Neurology
>>>  University of Toronto Faculty of Medicine
>>>
>>> Kimel Family Building, 6th Floor, Room 644
>>> Baycrest Centre
>>> 3560 Bathurst Street
>>> Toronto, Ontario M6A 2E1
>>> Canada
>>>
>>> Phone: 416-785-2500 x3648
>>> Fax: 416-785-2484
>>> Email: jween at klaru-baycrest.on.ca
>>>
>>>
>>> Confidential: This communication and any attachment(s) may contain
>>> confidential or privileged information and is intended solely for  
>>> the
>>> address(es) or the entity representing the recipient(s). If you have
>>> received this information in error, you are hereby advised to  
>>> destroy
>>> the document and any attachment(s), make no copies of same and  
>>> inform
>>> the sender immediately of the error. Any unauthorized use or
>>> disclosure of this information is strictly prohibited.
>>> _______________________________________________
>>> MINC-users at bic.mni.mcgill.ca
>>> http://www2.bic.mni.mcgill.ca/mailman/listinfo/minc-users
>>
>> _______________________________________________
>> MINC-users at bic.mni.mcgill.ca
>> http://www2.bic.mni.mcgill.ca/mailman/listinfo/minc-users
>>
>
> _______________________________________________
> MINC-users at bic.mni.mcgill.ca
> http://www2.bic.mni.mcgill.ca/mailman/listinfo/minc-users