There was an error in my previous post.
I used 4-4-4mm FWHM filter for
pre-smoothing, not 6-6-6mm. (I was confused with my other FLB study that is
done in the PET-CT).
So, the estimated smoothness by SPM (before
doing additional smoothing with 10mm FWHM) actually made sense…
Now, everything is clear!
Thanks!
Ji Hyun
From: SPM (Statistical
Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Ji Hyun Ko
Sent: Tuesday, May 25, 2010 4:14
PM
To: [log in to unmask]
Subject: Re: partial correlation
for ligand PET study in SPM5
Dear Alexander,
That explains a lot.
Thank you so much for your quick feedback!
Best,
Ji Hyun
From: Alexander
Hammers [mailto:[log in to unmask]]
Sent: Tuesday, May 25, 2010 4:09
PM
To: Ji Hyun Ko
Cc: [log in to unmask]
Subject: Re: [SPM] partial
correlation for ligand PET study in SPM5
Dear Ji Hyun,
Ah, you hadn't mentioned the HRRT before - I didn't know we were
talking Formula One PET scanning ;-)!
In that case, if initially you didn't apply any smoothing after normalisation
but prior to analysis, the FWHM does look as if it could conceivably be derived
from HRRT images. By pre-smoothing the 4D images, you would simply have removed
some of the noise, but the BPND images would still be as rough as normal PET
images. As I'm sure you know, for PET parametric image correlation analyses as
for any other SPM analysis, smoothing is necessary for statistical reasons as
Tom already said, as well as neuroanatomical ones (to get similar regions in
register) - so not only justifiable as per your question, but necessary.
In the event of being interested in really small areas, you could also
try to preserve the HRRT advantages by doing an ROI analysis, too, and
correlate the mean BPNDs from those.
Hope this helps,
Alexander
On 25 May 2010, at 20:43, Ji Hyun Ko wrote:
Dear Alexander,
Thanks for your comment.
Actually, our raw PET images (4D) look
somewhat similar to the figure that you referred, because we used HRRT scanner
(it’s really noisy).
However, smoothing makes the image much
better, and the BP image actually looks fine without additional smoothing.
Anyway, that’s one of the reasons
why we smoothed before making BP image.
I thought that pre-smoothing may help with
motion correction, coregistration with MRI and SRTM fitting because there is
really a lot of noise in HRRT image.
There are several modeling papers tested
SRTM for FLB, and most of them suggest that it is good enough although some
suggest some degree of bias.
The reason why we used SRTM for FLB was to
use the residual t-test (Aston et al., Human Brain Mapping, 2000) as we
previously did in Ko et al. (NeuroImage, 2009, vol 46(2), pp 516-521). We used
PET-CT in the 2009 paper, so the image was much less noisy. In the paired t-test
using residuals (Aston et al., 2000), there was no problem with t-values, and
we verified it by post-hoc VOI analysis.
Here, since there is no validated way to
do the voxel-based correlation analysis using residuals of fitting (Aston et
al., 2000; and we are currently working on it!), we used SPM for the
correlation analysis.
So, the possible reason that I can think
of is pre-smoothing of raw PET images did not end up with smooth BP image which
I cannot explain why...
If we assume that we didn’t have any
other quantification errors, is it justifiable to do smoothing on the BP image
before we do SPM correlation analysis?
Best,
Ji Hyun
From: Alexander Hammers [mailto:[log in to unmask]]
Sent: Tuesday, May 25, 2010 1:32 PM
To: Ji Hyun Ko
Cc: [log in to unmask]
Subject: Re: [SPM] partial correlation for
ligand PET study in SPM5
Dear Ji Hyun,
Ok, there's the likely solution - your PET
model probably doesn't quite work! The images didn't "loose
smoothness" but can become very "rough" by aberrant high and low
values in neighbouring pixels. What do your images look like - something like
the top left image in Fig 3A in Neuroimage 38 (2007) 82 – 94?
I've never worked with FLB. As a D2 ligand
it's possible that the SRTM would work - but here it presumably produced
aberrant values. This could have a large variety of reasons:
- subject motion
- unsuitable model
- quantification errors (e.g. different
scalefactors from frame to frame)
- reconstruction errors (e.g. emission -
transmission mismatch)
- and, quite possibly, mixing different
kinetic classes by smoothing your 4D prior to modelling.
I'd start by looking for subject motion
and leaving out the pre-modelling smoothing (plus of course prior modelling
work - do others use SRTM for this ligand, and, if so, using which model?). I
definitely wouldn't accept the images (and results) as they are now - you
likely had a major quantification error in there which you've now simply
smoothed into oblivion.
Hope this helps!
All the best,
Alexander
---------------------------------
Alexander Hammers, MD PhD
<image002.jpg>
NEW Affiliation:
Chair in Functional Neuroimaging
Fondation Neurodis
http://www.fondation-neurodis.org/
NEW Postal Address:
CERMEP – Imagerie du Vivant
Hôpital Neurologique Pierre Wertheimer
59 Boulevard Pinel
69003
Telephone
+33-(0)4-72 68 86 34
Fax
+33-(0)4-72
68 86 10
Email
[log in to unmask]">[log in to unmask];[log in to unmask]">[log in to unmask]
Web (
http://www1.imperial.ac.uk/medicine/people/alexander.hammers.html
http://www.csc.mrc.ac.uk/Research/Groups/ECN/Epilepsy/
---------------------------------
Other affiliations:
Visiting Reader; Honorary Consultant Neurologist
Division of Neuroscience and Mental Health,
Faculty of Medicine
Room 243,
and
Honorary Reader in Neurology; Honorary Consultant
Neurologist
Department of Clinical and Experimental Epilepsy
UCL
On 25 May 2010, at 18:12, Ji Hyun Ko wrote:
Dear Tom,
I think I found a solution.
Originally, I smoothed the dynamic images
before producing binding potential (BP) images with 6mm-6mm-6mm FWHM Gaussian
filter. When I produced the binding potential images (which is 4D -> 3D), I
think it somehow lost its smoothness…<- Does this make sense to you?
So, I smoothed the BP images (3D) with 10-10-10mm
FWHM filter, and then run the SPM.
Then, it gave me estimated smoothness of
about 20mm.
Now it gives me realistic t-values!
So, I guess my problem is
“solved,” but I am still curious why the images lost the smoothness
when I produced BP images (4D->3D).
I used Roger Gunn’s simplified
reference tissue model.
Best Regards,
Ji Hyun
To: [log in to unmask]
CC: [log in to unmask]
From: [log in to unmask]
Subject: Re: [SPM] partial correlation for ligand PET study in SPM5
Date: Sun, 23 May 2010 23:29:17 +0100
Dear Ji,
Hmm... ~5mm FWHM
smoothness... does that surprise you? What sort of reconstruction filter
was used? I would have thought the smoothness would be much greater (the
estimated smoothness accounts for both intrinsic and applied smoothness, and
should at least exceed the size of any applied Gaussian kernel smoothing.
-Tom
On Fri, May 21, 2010
at 2:56 PM, Ji Hyun Ko <[log in to unmask]>
wrote:
Dear Tom,
Thank you so much for your advice. I will
take a look at the SnPM.
Anyway, the SPM gave me the following:
Degree of freedom = [1.0, 4.0]
FWHM = 4.3 4.7 5.3 mm mm mm; 2.1 2.4
2.6 (voxels);
Volume: 1156096 = 144512 voxels = 8360.6
resels
Voxel size: 2.0 2.0 2.0 mm mm mm; (resel =
13.25 voxels)
Best,
Ji Hyun
From: [log in to unmask] [mailto:[log in to unmask]] On
Behalf Of Thomas
Nichols
Sent: Friday, May 21, 2010 4:16 AM
To: Ji Hyun Ko
Cc: [log in to unmask]
Subject: Re: [SPM] partial correlation for
ligand PET study in SPM5
Dear Ji,
You are right to be skeptical of amazing
results from just 7 subjects. Can you also report your estimated
smoothness (FWHM in voxels and mm, and Resel count)? That will help to
see if you have a chance of getting reasonable accuracy out of the RFT
corrected P-values.
With such a tiny amount of data, you also
might want to try a nonparametric approach, with variance smoothing in
particular. See SnPM toolbox under SPM extensions for more details.
-Tom
On Thu, May 20, 2010 at 9:14 PM, Ji Hyun
Ko <[log in to unmask]> wrote:
Dear SPM users,
I have two questions regarding how to do the partial correlation for a
ligand PET study.
I have 7 patients and each of them had one FLB-PET scan. I produced binding
potential maps (3D image) using some in-house software.
I have two covariates. One of the covariates is the behavioural score and
the other is age which should be controlled for (nuisance covariate).
So, I have design matrix of X = [behaviour, age, constant].
Of course, Those are column vectors.
And, I used C = [1 0 0].
My first question is...is this the right way to do partial correlation?
My second question is about extent threshold.
I put 0.001 for uncorrected height threshold, and put 10 for extent
threshold.
And, I think the SPM gives me some unrealistic values.
All of the p_corrected in cluster-level was less than 0.001.
And, in the left bottom of the result window, it gives:
Height threshold: T=7.17, p=0.001 (1.000)
Extent threshold: k = 10 voxels, p=0.000 (0.000)
Expected voxels per cluster, <k>=0.156
Expected number of clusters, <c>=0.00
Expected false discovery rate, <=0.15
I think there is something wrong.
How can I have such "good" p_corrected values with only 7 scans?
Please help...
Best,
Ji Hyun
____________________________________
Ji Hyun Ko, PhD
Post-doctoral Fellow,
Centre for Addiction and Mental Health &
Toronto Western Research Institute,
Physical Address:
Mailing Address:
PET Centre, CAMH
Phone: +1-416-535-8501 ext. 7396
Fax: +1-416-979-3855
--
____________________________________________
Thomas Nichols, PhD
Principal Research Fellow, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
Email: [log in to unmask]
Phone, Stats: +44 24761 51086, WMG: +44 24761 50752
Fax: +44 24 7652 4532
--
____________________________________________
Thomas Nichols, PhD
Principal Research Fellow, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
Email: [log in to unmask]
Phone, Stats: +44 24761 51086, WMG: +44 24761 50752
Fax: +44 24 7652 4532
MSN Dating: Find
someone special. Start now. Start now!