Dear Jesper,
Thank you very much for further advices. We will try these options.
Riki
Jesper Andersson さんは書きました:
> Dear Riki,
>
>
>> We will try--biasres (from default 50,50,50 to 30,30,30: 20,20,20;
>> 10,10,10) and --in&refpointlist.
>> As for --pointlistlambda=1, how can we modify it?
>> Change by order of 1 (e.g., =2, 3, 4) or by 10?
>>
>
> A value of 1 should give you "fairly firmly anchored" points, whereas a
> value of e.g. 0.1 would allow more leeway. If you have some way of being
> certain about the points (e.g. if the represent some fiducial markers) you
> could use 10 or even 100 to make sure they are mapped exactly on top of
> each other.
>
> I am afraid that is as much advice as I can give on that. This weight
> (relative the sum-of-squared differences cost-function) is something I
> will need to determine before I "unhide" the option.
>
> Good Luck Jesper
>
>
>
>>
>>
>>> It sounds like you have done most of the things I would recommend. The
>>> critical thing is that fnirt assumes that everything that is present
>>> in one image is present also in the other (though it may have a
>>> different size and shape). So e.g. if it finds a ventricle in one
>>> brain it expects to find one in the other as well.
>>>
>>> This assumption is of course not fulfilled when you have a pathology
>>> (e.g. a tumor) or some foreign object (electrode) in one of the images
>>> and not in the other. Our recommendation then is to mask these areas
>>> out using some suitable combination of --inmask and --refmask. And it
>>> sounds like you have already done that.
>>>
>>> Another possibility (which does not address your electrodes, but might
>>> help those intensity hot-spots) is to use a higher resolution for the
>>> bias-field. It will mean longer execution time, and may not work if it
>>> means that intensity and shape gets too strongly correlated. But it
>>> would be worth a try.
>>>
>>> Another thing you might try is to complement your data with a list of
>>> point correspondences. This is a hidden feature, that is quite
>>> untested, that might potentially help you. If you are able to manually
>>> identify unique points in both volumes you can put these in two text
>>> files, and that means that these points will serve as an anchor
>>> between the two volumes (and might prevent fnirt from doing silly
>>> things in that region).
>>>
>>> Let us say e.g. that you have managed to identify two points on the
>>> cortical surface in both volumes. You should then create a file with
>>> the points identified in the --ref file, e.g.
>>>
>>> ref_pl.txt contains
>>> 15 45 67
>>> 18 52 63
>>>
>>> and one for the --in file
>>> in_pl.txt contains
>>> 20 42 56
>>> 23 50 50
>>>
>>> when running fnirt you then specify these with
>>>
>>> fnirt --ref=.... ..... --refpointlist=ref_pl.txt
>>> --inpointlist=in_pl.txt
>>>
>>> These points are not "completely anchored", but rather have some
>>> slack. The amount of slack is given by the
>>>
>>> --pointlistlambda
>>>
>>> parameter. The default is --pointlistlambda=1 , but you may wish to
>>> experiment with it.
>>>
>>> Also, please note that there is a reason I hid this option for the
>>> release. I need to do lots more testing of it, and there is a risk you
>>> may encounter problems with it.
>>>
>>> Good luck Jesper
>>>
>>>
>>>
>>>
>>
>
>
>
>
--
Riki Matsumoto M.D., Ph.D.
Department of Neurology
Kyoto University Graduate School of Medicine
Shogoin, Sakyo, Kyoto, JAPAN 606-8507
Tel: +81-(0)75-751-3772
Fax:+81-(0)75-751-9416 (for large transfer, use +81-(0)75-761-9780)
|