Thanks Bas. You are indeed correct. I haven't included anything very
specific to T1w scans in the segmentation, which is sometimes a good
thing, and sometimes not. On the positive side, the algorithm is very
flexible in terms of what kind of images it can deal with. On the
negative side, the algorithm is very flexible in terms of what kind of
images it can deal with.
A more constrained model of the image intensity distributions likely
to be encountered in T1w images would result in greater accuracy for
segmenting T1w images, but less flexibility in terms of what other
types of image it can deal with.
All the best,
-John
On 16 March 2012 15:49, Bas Neggers <[log in to unmask]> wrote:
> Hi Susanne,
>
> a good question, and something I contemplated about as well a couple of
> years ago.
>
> With the risc of calling on John's wrath upon me for telling you nonsense:
>
> Unified segmentation models tissue histograms based on a mixture of
> gaussians, and is therefore determined by the relative contrast in the image
> itself, and is hence not dependent on image modality. The tissue priors are
> probability maps from a template, and are as such modality independent
> itself.
>
> To see the logic and math behind this claim, check John's landmark paper:
>
> http://www.ncbi.nlm.nih.gov/pubmed/15955494
>
> It took me quite a while to fully grasp that paper (John's talk here in
> Utrecht helped a lot!), but it is quite an elegant approach IMHO. I am not
> sure whether there are modality dependent details in the implementation of
> unified segmentation, I didn't find them.
>
> At the very least, I was able to use unified segmentation on high resolution
> 2x2x2 mm³ EPI scans (from 3T scanner) directly, and normalize it without the
> use of a T1. As my EPI images had high resolution and pretty good contrast
> but were deformed a bit (the former causing the latter) this seemed to be
> the best way to reach good normalization. But some caution is warranted: I
> had to exclude 1 out of 13 subjects as the EPI of this individual wasnt good
> enough to segment. I have the feeling thatin general EPI contrast is a bit
> on the edge when trying this. After a functional session I ran 20 whole
> brain EPIs (with longer TR to cover the whole brain, my time series had
> smaller FOV) with the same angulation and hence deformation as the actual
> time series data. I then calculated the average of these 20 EPIs to improve
> the contrast. Therefore coregistration of the time series data with that
> average was near perfect, and I could subsequently use unified segmentation
> for normalization based on that average whole brain EPI. For more, read the
> methods in: http://www.ncbi.nlm.nih.gov/pubmed/22235303
>
> So use it with caution, you need very good EPIs for this to work I think,
> and check your results for all individual subjects carefully. I am sure
> unified segmentation was designed for use with higher contrast images such
> as T1 or T2, but it can work for other modalities, thus solving the
> distortion issue for high resolution EPI.
>
> Again, John might want to confirm this, or otherwise release his wrath upon
> me.
>
> Cheers,
>
> Bas
>
>
>
>
> On 03/16/2012 04:15 PM, Merz, Susanne wrote:
>>
>> Dear John, Bas, and list
>>
>> I'm sorry if this is a stupid question, but there is something I do not
>> understand about the suggestion to segment the EPI directly. I am assuming
>> that, as the alternative option is segmenting a T1 image, the EPI images are
>> not T1 weighted. However, the segmentation procedure (Segment button) seems
>> to be geared towards T1 images, with respect to what tissue shows up as what
>> intensity, and what the contrast in the image is. How would one handle
>> that? What templates would be used? Thanks for your help.
>>
>> Susa
>>
>> -----Original Message-----
>> From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On
>> Behalf Of John Ashburner
>> Sent: 13 March 2012 16:27
>> To: [log in to unmask]
>> Subject: Re: [SPM] Normalization of EPI using unified segmentation
>>
>> Thanks Bas,
>> I think you've just about covered all I had to say. You make the
>> important point that what works best will depend on the data. For the
>> regular spatial normalisation using mean squared difference and the
>> EPI template, the accuracy will depend on (among other things) how
>> similar the image intensities are to those used for constructing the
>> template.
>>
>> To be honest, I haven't tried normalising a wide variety of EPI scans
>> using segmentation, so I don't yet have a good idea about what sorts
>> of EPI scans can be normalised this way. Essentially though, the
>> segmentation is a bit more agnostic about the range of intensities
>> that different tissue types can have. This gives it additional
>> flexibility, which can sometimes help.
>>
>> Best regards,
>> -John
>>
>> On 13 March 2012 15:02, Neggers, S.F.W.<[log in to unmask]> wrote:
>>>
>>> Dear Xiaoying (cc to list as well),
>>>
>>> thats not bad per se, I guess. I never used your method, so please
>>> consult the list to get comments on that one.
>>>
>>> I think directly normalizing EPI to EPI template the classical way is
>>> just not very precise. It matches the course anatomical brain shape, and
>>> does hardly takeinto accound local morphology. Unified segmentation based
>>> normalization (http://www.ncbi.nlm.nih.gov/pubmed/15955494) seems to be more
>>> precise, but in general you need a good (mostly T1) scan for that. DARTEL
>>> performs even better, Ive heard (not using it yet).
>>>
>>> Check for example this paper:
>>>
>>> http://www.ncbi.nlm.nih.gov/pubmed/17616402
>>>
>>> I use unified segmentation mostly, sometimes on EPI. When your EPIs are
>>> good, you might consider using unified segmentation on that too, but see the
>>> caveats in my last email today.
>>>
>>> I'm sure John has to say a thing or 2 about this as well.
>>>
>>> Cheers,
>>>
>>> Bas
>>>
>>>
>>> --------------------------------------------------
>>> Dr. S.F.W. Neggers
>>> Division of Brain Research
>>> Rudolf Magnus Institute for Neuroscience
>>> Utrecht University Medical Center
>>>
>>> Visiting : Heidelberglaan 100, 3584 CX Utrecht
>>> Room B.01.1.03
>>> Mail : Huispost B01.206, P.O. Box 85500
>>> 3508 GA Utrecht, the Netherlands
>>> Tel : +31 (0)88 7559609
>>> Fax : +31 (0)88 7555443
>>> E-mail : [log in to unmask]
>>> Web : http://www.neuromri.nl/people/bas-neggers
>>> : http://www.brainsciencetools.com (CEO)
>>> --------------------------------------------------
>>>
>>> ________________________________________
>>> From: Xiaoying, Fan [[log in to unmask]]
>>> Sent: Tuesday, March 13, 2012 3:26 PM
>>> To: Neggers, S.F.W.
>>> Subject: RE: [SPM] Normalization of EPI using unified segmentation
>>>
>>> Dear Dr. Neggers,
>>>
>>> I read your message about normalization. It is very helpful. Could I
>>> ask one related question?
>>>
>>> I have some old data collected long time ago without high resolution
>>> T1WI. Can I directly normalize the mean EPI to the MNI template without
>>> segmentation? I am not sure why need to do EPI segmentation for
>>> normalization. You mentioned it is for good anatomical contrast in your
>>> EPIs. If idid not use segmentation, just use the mean EPI, would that be
>>> very bad? I am new on fMRi and it might be a silly question.
>>>
>>> thanks for your time,
>>>
>>> Best regards,
>>>
>>> Xiaoying
>>> ________________________________________
>>> From: SPM (Statistical Parametric Mapping) [[log in to unmask]] on behalf
>>> of Neggers, S.F.W. [[log in to unmask]]
>>> Sent: Tuesday, March 13, 2012 7:31 AM
>>> To: [log in to unmask]
>>> Subject: Re: [SPM] Normalization of EPI using unified segmentation
>>>
>>> Hi Thilo, all,
>>>
>>> I do not know of a particular paper investigating this, but can only
>>> report from my experience. Furthermore, I do not think it makes a lot of
>>> sense to study the differences between both strategies you mention, as it
>>> will all boil down to your specific EPI acquisition technique, and those are
>>> very different from site to site and study to study. So a general claim
>>> which strategy is superior isnt very generalizable. In other words, the MR
>>> physicists are right (aren't they always ;-) ).
>>>
>>> And the best answer about what to do, is, as always: it depends.
>>>
>>> Indeed, EPI deformations can be so substantial in the entire image and
>>> especially in certain areas (temporal, orbitofrontal) that the
>>> coregistration with T1 will be poor, and hence normalization of your EPIs
>>> will be way off when using T1 unified segementation parameters. When you
>>> have sufficient anatomical contrast in your mean EPI, for example, you would
>>> then be better of using unified segmentation on your mean EPI directly
>>> (apply norm.parameters to all EPI dynamics). When doing this it is important
>>> to check unified segmentation results for every subject though, you need
>>> high resolution EPIs for this to separate GM and WM. Unified segmentation,
>>> in my experience, often wont work for low resolution EPI (say anything at or
>>> above 3x3x3 mm). I recently used this strategy of direct EPI segmentation
>>> normalization in a 3T 2x2x2 mm^3 2DEPI study, see:
>>> http://www.ncbi.nlm.nih.gov/pubmed/22235303
>>>
>>> When accelerating your EPI readout (for example by using parallel imaging
>>> schemes such as SENSE, and/or by using a high acquisition bandwith) you can
>>> substantially reduce spatial distortions, for low res fMRI they would be
>>> hardly noticable on standard 8 channel headcoils with acceleration factors
>>> of 2 to 3. In that case, the 1st strategy where EPI is normalized based on
>>> unified segmentation parameters from (coregistered) T1 would work best.
>>>
>>> In short: check your spatial distortion in your EPIs with respect to your
>>> T1. When they are large, use EPI segmentation directly for normalization
>>> (provided you have good anatomical contrast in your EPIs), check
>>> segmentation results manually for every subject. When EPI spatial
>>> distrortion it is neglegible, use T1 unified segmentation parameters for
>>> your EPIs, be sure to check coreg between T1 and mean EPI thoroughly for
>>> every subject. When spatial distortion is high AND EPI anatomical contrast
>>> is poor, you are lost and you will have to design better EPI acquisition (or
>>> live with classical normalization based on course anatomical shape, use
>>> enormous amounts of smoothing).
>>>
>>> As said, one EPI isnt the other...
>>>
>>> Cheers,
>>>
>>> Bas
>>>
>>> --------------------------------------------------
>>> Dr. S.F.W. Neggers
>>> Division of Brain Research
>>> Rudolf Magnus Institute for Neuroscience
>>> Utrecht University Medical Center
>>>
>>> Visiting : Heidelberglaan 100, 3584 CX Utrecht
>>> Room B.01.1.03
>>> Mail : Huispost B01.206, P.O. Box 85500
>>> 3508 GA Utrecht, the Netherlands
>>> Tel : +31 (0)88 7559609
>>> Fax : +31 (0)88 7555443
>>> E-mail : [log in to unmask]
>>> Web : http://www.neuromri.nl/people/bas-neggers
>>> : http://www.brainsciencetools.com (CEO)
>>> --------------------------------------------------
>>>
>>> ________________________________________
>>> From: SPM (Statistical Parametric Mapping) [[log in to unmask]] on behalf
>>> of Thilo Kellermann [[log in to unmask]]
>>> Sent: Tuesday, March 13, 2012 11:59 AM
>>> To: [log in to unmask]
>>> Subject: [SPM] Normalization of EPI using unified segmentation
>>>
>>> Dear SPM users,
>>>
>>> the seemingly simple question is if a normalization of functional images
>>> (EPI)
>>> will be more accurate using the following steps:
>>> 1) the structural image (usually T1) is co-registered to the mean EPI
>>> 2) normalization parameters are determined for the T1 (via unified
>>> segmentation) and
>>> 3) normalization parameters are applied to EPI time series
>>>
>>> The alternative would be to apply the unified segmentation procedure
>>> directly
>>> to the mean EPI which provides normalization parameters for the time
>>> series.
>>>
>>> Some people (usually neuroscientists) argue that the first procedure is
>>> more
>>> accurate since the T1 provides much more spatial details (good contrast
>>> between GM, WM and CSF) which improve the normalization.
>>>
>>> Other people (usually MR physicists) argue that the distortions of the
>>> two
>>> modalities (T1 weighted images vs. T2* weighted images) are quite
>>> different.
>>> Therefore, good normalization parameters for the structural image are not
>>> necessarily helpful in order to transform EPI images into standard MNI
>>> space
>>> in the best possible way.
>>>
>>> Is this debate still a "matter of taste" or is there any study that
>>> already
>>> addressed this issue (which I am not able to find...)? I am aware of the
>>> study by Crinion and colleagues (Neuroimage 2007) which addressed a
>>> related
>>> but still different question. Are there any estimates of displacement
>>> using
>>> the procedures described above?
>>>
>>> Thanks a lot for your comments,
>>> Thilo
>>>
>>> --
>>> Thilo Kellermann
>>> RWTH Aachen University
>>> Department of Psychiatry, Psychotherapy and Psychosomatics
>>> JARA Translational Brain Medicine
>>> Pauwelsstr. 30
>>> 52074 Aachen
>>> Germany
>>> Tel.: +49 (0)241 / 8089977
>>> Fax.: +49 (0)241 / 8082401
>>> E-Mail: [log in to unmask]
>>>
>>> ------------------------------------------------------------------------------
>>>
>>> De informatie opgenomen in dit bericht kan vertrouwelijk zijn en is
>>> uitsluitend bestemd voor de geadresseerde. Indien u dit bericht onterecht
>>> ontvangt, wordt u verzocht de inhoud niet te gebruiken en de afzender
>>> direct
>>> te informeren door het bericht te retourneren. Het Universitair Medisch
>>> Centrum Utrecht is een publiekrechtelijke rechtspersoon in de zin van de
>>> W.H.W.
>>> (Wet Hoger Onderwijs en Wetenschappelijk Onderzoek) en staat
>>> geregistreerd bij
>>> de Kamer van Koophandel voor Midden-Nederland onder nr. 30244197.
>>>
>>> Denk s.v.p aan het milieu voor u deze e-mail afdrukt.
>>>
>>>
>>> ------------------------------------------------------------------------------
>>>
>>> This message may contain confidential information and is intended
>>> exclusively
>>> for the addressee. If you receive this message unintentionally, please do
>>> not
>>> use the contents but notify the sender immediately by return e-mail.
>>> University
>>> Medical Center Utrecht is a legal person by public law and is registered
>>> at
>>> the Chamber of Commerce for Midden-Nederland under no. 30244197.
>>>
>>> Please consider the environment before printing this e-mail.
>>>
>>>
>>> The information in this e-mail is intended only for the person to whom it
>>> is
>>> addressed. If you believe this e-mail was sent to you in error and the
>>> e-mail
>>> contains patient information, please contact the Partners Compliance
>>> HelpLine at
>>> http://www.partners.org/complianceline . If the e-mail was sent to you in
>>> error
>>> but does not contain patient information, please contact the sender and
>>> properly
>>> dispose of the e-mail.
>>
>>
>> The University of Aberdeen is a charity registered in Scotland, No
>> SC013683.
>
>
>
> --
> --------------------------------------------------
> Dr. S.F.W. Neggers
> Division of Brain Research
> Rudolf Magnus Institute for Neuroscience
> Utrecht University Medical Center
>
> Visiting : Heidelberglaan 100, 3584 CX Utrecht
> Room B.01.1.03
> Mail : Huispost B01.206, P.O. Box 85500
> 3508 GA Utrecht, the Netherlands
> Tel : +31 (0)88 7559609
> Fax : +31 (0)88 7555443
> E-mail : [log in to unmask]
> Web : http://www.neuromri.nl/people/bas-neggers
> : http://www.brainsciencetools.com (CEO)
> --------------------------------------------------
|