Dear Erick,
> In your opinion a computer (Linux) with 8 cores and 16 GB (or 24 GB)
> of ram memory is sufficient to perform the registration using for
> example a warp resolution of 5mm?
>
> The registration errors in the cortex, expected from a warp
> resolution of 5mm, are in the range 1-2.5 mm?
I am afraid it is not that simple. Firstly it is very hard to
establish actual registration errors for non-linear registration since
this is something that differ from voxel to voxel and for which one
would need to know the ground truth for each voxel.
What I have been able to do is to use a data set with manually
delineated cortical regions for a number of subjects. I have then
registered these subjects scans to each other and looked at how well
the manually defined regions have aligned/overlapped. What I find from
that is that there are large improvements in overlap (~between 50 to a
100% increases) when going from an affine transform to 10mm fnirt.
When comparing 2mm fnirt to 10mm fnirt I find only marginal
improvements in this population of normal subjects (this is the NIREP
data set).
I I was to venture a guess as to what that might mean in reduction of
voxel-wise registration error I would guess maybe 0.5-1mm.
I.e. you will not obtain a halving of the registration errors by
halving the knot-spacing, but rather you will achieve a modest
improvement.
>
> Thanks again for your comments and your excellent tool!
Cheers!
Jesper
>
> Erick
>
>
> ________________________________________
> De: FSL - FMRIB's Software Library [[log in to unmask]] En nombre de
> Jesper Andersson [[log in to unmask]]
> Enviado el: lunes, 13 de diciembre de 2010 11:36
> Para: [log in to unmask]
> Asunto: Re: [FSL] Advanced normalization options with FNIRT
>
> Dear Erick,
>
> The limitations to the resolution obtainable with the current version
> of fnirt is not so much execution time as it is memory demands. I
> think realistically that you cannot obtain a higher warp resolution
> than 5mm, and even then you will need a powerful computer with a lot
> of memory.
>
> That does not mean you cannot register your images to a template with
> higher resolution than that. You can use the warp fields you obtain
> from the "standard" fnirt registration to resample your images into
> MNI 1mm or 0.5mm space by specifying one of those as the --ref
> argument to applywarp. It should also be noted that a warp resolution
> (knot spacing) of 10mm (the standard) does not imply a 10mm
> registration error. My guesstimate is that registration errors in the
> cortex are in the range 2-5mm, and less for central structures.
>
> The next version of fnirt will have the ability use a warp resolution
> down to 1mm with reasonable memory requirements. When testing this
> version I have found marginal advantages with 1mm over and above 10mm
> for normal healthy subjects. The big difference, and benefit, is
> mainly evident for elderly atrophied brains.
>
> Jesper
>
> On 13 Dec 2010, at 10:21, Erick Canales Rodriguez (Benito Menni CASM)
> wrote:
>
>> Dear FSL experts,
>>
>> By default FNIRT was designed to be a "medium resolution” method
>> that provides an excellent tradeoff between the accuracy of the non-
>> linear registration and execution time (and working memory).
>> However, my main interest is to apply the non-linear registration as
>> accurate as possible – the execution time per subject is not a
>> problem for my application. I am willing to wait as long as needed
>> (within an interval of 24 hours) in order to obtain the most
>> accurate result.
>>
>> This particular mode of operation is not aimed to be used in
>> standard methods like VBM. In my project I need to obtain accurate
>> high resolution “displacement field”, which will be used to
>> transforms particular points from the native space to the MNI space
>> (for example, the points belonging to a particular tract {xi,yi,zi,
>> i=1..n} reconstructed by DTI and deterministic tractography
>> algorithms).
>> There are any “optimal” configuration files to perform this type of
>> high resolution registration from the native space to the templates
>> MNI152_T1_1mm or MNI152_T1_0.5mm?
>>
>> Any recommendations would be appreciated,
>>
>> Many thanks,
>>
>> Erick J. Canales-Rodríguez
>>
>> (+34) 93 652 99 99 ext. 255
>> Unitat de Recerca, Benito Menni CASM
>> Sant Boi del Llobregat (Barcelona), Spain
>
> Este mensaje se dirige exclusivamente a su destinatario y puede
> contener información privilegiada o confidencial. Si no es vd. el
> destinatario indicado, queda notificado de que la utilización,
> divulgación y/o copia sin autorización está prohibida en virtud de
> la legislación vigente. Si ha recibido este mensaje por error, le
> rogamos que nos lo comunique inmediatamente por esta misma vía y
> proceda a su destrucción.
>
> Aquest missatge va dirigit exclusivament al seu destinatari i pot
> contenir informació privilegiada o confidencial. Si vosté no és el
> destinatari indicat, li notifiquem que la utilització, divulgació i/
> o còpia sense autorització està prohibida en virtud de la legislació
> vigent. Si ha rebut aquest missatge per error, li preguem que ens ho
> comuniqui el més aviat possible.
>
|