sorry, I didn't mean to say it is half broken. I meant that we haven't had time to go through the set of regression and system tests that we want to in order to make sure that everything works as well as we can make it. We've been going through that process and finding little things that we can improve, and have decided to make the improvements and retest rather than release things. Mostly the current dev version works extremely well.
sorry for the confusion.
Bruce
On Sat, 22 Oct 2005, Fornito, Alexander wrote:
Hi Bruce, Just wanted to clarify: When you say the dev releases are "half broken" does that mean there may be some minor niggles but results generated should be (on the whole) trustworthy, or does it mean that we should definitely avoid doing a complete study (eg., from surface generation through to morphometric or fMR analysis) using a dev version? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
-----Original Message----- From: freesurfer-bounces@nmr.mgh.harvard.edu on behalf of Bruce Fischl Sent: Sat 22/10/2005 2:57 AM To: freesurfer@surfer.nmr.mgh.harvard.edu Subject: [Freesurfer] good news and bad news
Hi All,
I'll start with the good:
thanks to Anders Dale (UCSD), Howard Pinsky (CorTechs) and Christian Euseman (MGH), we have finally signed an open source agreement for FreeSurfer. We will be posting a "read-only" tarball on the website in the next couple of weeks, and anyone is free to download it and look at it. We will *not* be supporting a make/configure type env for people to download and build, as we simply aren't ready to do so. We plan to implement such a procedure, but it will be a while. Instead of waiting until that was ready, I thought people would like to have access to the source to at least see what things are doing, and be able to see file formats and such.
Now the bad news. I know we've been promising a new release, but it is being delayed for another month or two. Partially this is due to the opensourcing, which has taken our time and resources, but it is also because we *really* don't want to release something that is half broken. There are also a number of features that we want to get into the release, and not have updates to it right away. We have continued to post current versions on the website as prereleases, but *please* be aware that we can't guarantee that these versions will be compatible with the one we ultimately release. I think they will be, but can't say for sure. Our hope is that datasets processed with the prerelease versions can be rerun automatically with recon-all --rerun or something like that, but again, no guarantees.
cheers, Bruce
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi, We're about to start a longitudinal study and are keen to get accurate measures of structural change. Currently we are planning to acquire a T1 and T2, and have some spare time for another acuqisition. I've noticed in a recent paper (Fischl, et al (2004) Neuroimage,23, S69-S84) you discuss the benefits of combining T1 and PD images to aid segmentation. Just wandering if there are any plans to incorporate multichannel segmentation into freesurfer, and if so, would you recommed acquiring two T1s and averaging, or a single T1 and PD? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
Hi Alex,
unfortunately the current answer depends on what part of the brain you are interested in. For cortex, it appears that the raw CNR of mp-rage is significantly better than the multi-channel (T1/PD/T2*) CNR. On the other hand, subcortically for things like pallidum/putament, the multi-channel segmentation has much higher CNR.
We do have multi-channel aseg segmentation in place, but haven't used it a ton (since we have so much mp-rage data around).
cheers, Bruce
On Mon, 24 Oct 2005, Fornito, Alexander wrote:
Hi, We're about to start a longitudinal study and are keen to get accurate measures of structural change. Currently we are planning to acquire a T1 and T2, and have some spare time for another acuqisition. I've noticed in a recent paper (Fischl, et al (2004) Neuroimage,23, S69-S84) you discuss the benefits of combining T1 and PD images to aid segmentation. Just wandering if there are any plans to incorporate multichannel segmentation into freesurfer, and if so, would you recommed acquiring two T1s and averaging, or a single T1 and PD? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi, I was just wanting to know more about the .curv and .sulc files. Specifically, I had three questions: 1 - What exactly does the value assigned to each vertex represent (eg., is it the curvature of the vertex in the .cuv file?) 2 - How are they derived/calculated? 3 - What do they mean neuroanatomically (eg., does .sulc correspond to the depth of a sulcus?). Many thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
1. This is the smoothed mean curvature - the average of the two principal curvatures smoothed spatially.
2. sulc is the integrated dot product of the movement vector with the surface normal over the inflation process. Things that move consistently outwards thus get a positive sign (sulci) and inwards a negative sign (gyri).
3. curv highlights secondary and tertiary folds as much as primary ones, whereas sulc highlights the primary, deep folds.
Bruce
On Tue, 25 Oct 2005, Fornito, Alexander wrote:
Hi, I was just wanting to know more about the .curv and .sulc files. Specifically, I had three questions: 1 - What exactly does the value assigned to each vertex represent (eg., is it the curvature of the vertex in the .cuv file?) 2 - How are they derived/calculated? 3 - What do they mean neuroanatomically (eg., does .sulc correspond to the depth of a sulcus?). Many thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
So if I wanted to derive some kind of measure of the complexity of cortical folding, would .curv be the way to go? Would it be safe to say regions of higher curvature have greater cortical folding?
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
________________________________
From: Bruce Fischl [mailto:fischl@nmr.mgh.harvard.edu] Sent: Tue 25/10/2005 10:57 PM To: Fornito, Alexander Cc: freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] .sulc and .curv
1. This is the smoothed mean curvature - the average of the two principal curvatures smoothed spatially.
2. sulc is the integrated dot product of the movement vector with the surface normal over the inflation process. Things that move consistently outwards thus get a positive sign (sulci) and inwards a negative sign (gyri).
3. curv highlights secondary and tertiary folds as much as primary ones, whereas sulc highlights the primary, deep folds.
Bruce
On Tue, 25 Oct 2005, Fornito, Alexander wrote:
Hi, I was just wanting to know more about the .curv and .sulc files. Specifically, I had three questions: 1 - What exactly does the value assigned to each vertex represent (eg., is it the curvature of the vertex in the .cuv file?) 2 - How are they derived/calculated? 3 - What do they mean neuroanatomically (eg., does .sulc correspond to the depth of a sulcus?). Many thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi,
I am running mris_glm to compare the cortical thickness between two groups. I have a question, which method is used for correcting multiple comparison in the analysis? Thanks,
Antao
mris_glm does not correct for multiple comparisons itself. However, you can use fdr inside of tksurfer, or ...
Steve Smith and I just worked out how to use the FSL randomise program to compute the vertex-wise threshold. Randomise implements permutation testing, which is much less conservative than FDR or GRF.
When you run mris_glm, make sure to specify the --y output (something like --y y-lh.mgh). then run mri_surf2surf to convert it to nifit, something like:
mri_surf2surf --srcsubject average7 --trgsubject average7 \ --srcsurfval y-lh.mgh --src_type mgh \ --trg_type nii --trgsurfval y-lh.nii --hemi lh
You will also need to convert the design matrix produced by mris_glm (something like y.X.mat) into ascii. This can be done in matlab with something like:
load y.X.mat save('X.asc','X','-ascii')
Then run:
randomise -i y-lh -o y-lh \ -d X.asc -t design.con -n 5000 -V
Where design.con has your contrasts
The output will be something like: y-lh_max_tstat1.mgh, which you can view with tksurfer with something like:
tksurfer average7 lh inflated -overlay y-lh_max_tstat1.mgh
We're still working out the details on this (obviously:), you may have to play with this a little to get the command lines exactly correct.
Note that randomise program cannot do cluster-based thresholding because it is not aware that these values are really on the surface (not in a volume), but the max stat will work.
doug
Antao Du wrote:
Hi,
I am running mris_glm to compare the cortical thickness between two groups. I have a question, which method is used for correcting multiple comparison in the analysis? Thanks,
Antao
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
FYI, FSL has a nice site documenting the randomise program :)
http://www.fmrib.ox.ac.uk/fsl/randomise/index.html
Doug Greve wrote:
mris_glm does not correct for multiple comparisons itself. However, you can use fdr inside of tksurfer, or ...
Steve Smith and I just worked out how to use the FSL randomise program to compute the vertex-wise threshold. Randomise implements permutation testing, which is much less conservative than FDR or GRF.
When you run mris_glm, make sure to specify the --y output (something like --y y-lh.mgh). then run mri_surf2surf to convert it to nifit, something like:
mri_surf2surf --srcsubject average7 --trgsubject average7 \ --srcsurfval y-lh.mgh --src_type mgh \ --trg_type nii --trgsurfval y-lh.nii --hemi lh
You will also need to convert the design matrix produced by mris_glm (something like y.X.mat) into ascii. This can be done in matlab with something like:
load y.X.mat save('X.asc','X','-ascii')
Then run:
randomise -i y-lh -o y-lh \ -d X.asc -t design.con -n 5000 -V
Where design.con has your contrasts
The output will be something like: y-lh_max_tstat1.mgh, which you can view with tksurfer with something like:
tksurfer average7 lh inflated -overlay y-lh_max_tstat1.mgh
We're still working out the details on this (obviously:), you may have to play with this a little to get the command lines exactly correct.
Note that randomise program cannot do cluster-based thresholding because it is not aware that these values are really on the surface (not in a volume), but the max stat will work.
doug
Antao Du wrote:
Hi,
I am running mris_glm to compare the cortical thickness between two groups. I have a question, which method is used for correcting multiple comparison in the analysis? Thanks,
Antao
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi Doug, we are eager to try the FSL randomise program on a FS data set we have. However, we have one (stupid, maybe) question: is the file design.con something that we make manually, or is it generated by some of the previous processes?
If we are going to make it ourselves, how should it be? In our GLM, we have two classes (males, females) and two variables (memory score, hippocampal volume), and we want to assess the relationship between the memory score and thickness when gender and hippocampal volume is regressed out), e.g. DOSS: 0 0 1 0
Thanks, - Anders
FYI, FSL has a nice site documenting the randomise program :)
http://www.fmrib.ox.ac.uk/fsl/randomise/index.html
Doug Greve wrote:
mris_glm does not correct for multiple comparisons itself. However, you can use fdr inside of tksurfer, or ...
Steve Smith and I just worked out how to use the FSL randomise program to compute the vertex-wise threshold. Randomise implements permutation testing, which is much less conservative than FDR or GRF.
When you run mris_glm, make sure to specify the --y output (something like --y y-lh.mgh). then run mri_surf2surf to convert it to nifit, something like:
mri_surf2surf --srcsubject average7 --trgsubject average7 \ --srcsurfval y-lh.mgh --src_type mgh \ --trg_type nii --trgsurfval y-lh.nii --hemi lh
You will also need to convert the design matrix produced by mris_glm (something like y.X.mat) into ascii. This can be done in matlab with something like:
load y.X.mat save('X.asc','X','-ascii')
Then run:
randomise -i y-lh -o y-lh \ -d X.asc -t design.con -n 5000 -V
Where design.con has your contrasts
The output will be something like: y-lh_max_tstat1.mgh, which you can view with tksurfer with something like:
tksurfer average7 lh inflated -overlay y-lh_max_tstat1.mgh
We're still working out the details on this (obviously:), you may have to play with this a little to get the command lines exactly correct.
Note that randomise program cannot do cluster-based thresholding because it is not aware that these values are really on the surface (not in a volume), but the max stat will work.
doug
Antao Du wrote:
Hi,
I am running mris_glm to compare the cortical thickness between two groups. I have a question, which method is used for correcting multiple comparison in the analysis? Thanks,
Antao
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
-- Douglas N. Greve, Ph.D. MGH-NMR Center greve@nmr.mgh.harvard.edu Phone Number: 617-724-2358 Fax: 617-726-7422
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
The dsign.con looks something like this:
--------------cut here ------------------- %! VEST-Waveform File /NumWaves 2 /NumContrasts 2 /PPheights 1.000000e+00 1.000000e+00
/Matrix 1.000000e+00 -1.000000e+00 -1.000000e+00 1.000000e+00 ------------- cut here -------------------
This file can be created automatically by the Feat GUI when running the group (gfeat) mode. Alternatively, you can create it by hand. Each row in the matrix just corresponds to a contrast vector.
Be aware that permutation/randomization cannot be used on all designs -- they must be orthogonal, which basically means that you cannot use nuisance regressors. I think this means that your design is not appropriate for this procedure :). Most of the time, permutation is used to test for a difference between classes (with no nuisance variables).
doug
a.m.fjell@psykologi.uio.no wrote:
Hi Doug, we are eager to try the FSL randomise program on a FS data set we have. However, we have one (stupid, maybe) question: is the file design.con something that we make manually, or is it generated by some of the previous processes?
If we are going to make it ourselves, how should it be? In our GLM, we have two classes (males, females) and two variables (memory score, hippocampal volume), and we want to assess the relationship between the memory score and thickness when gender and hippocampal volume is regressed out), e.g. DOSS: 0 0 1 0
Thanks,
- Anders
FYI, FSL has a nice site documenting the randomise program :)
http://www.fmrib.ox.ac.uk/fsl/randomise/index.html
Doug Greve wrote:
mris_glm does not correct for multiple comparisons itself. However, you can use fdr inside of tksurfer, or ...
Steve Smith and I just worked out how to use the FSL randomise program to compute the vertex-wise threshold. Randomise implements permutation testing, which is much less conservative than FDR or GRF.
When you run mris_glm, make sure to specify the --y output (something like --y y-lh.mgh). then run mri_surf2surf to convert it to nifit, something like:
mri_surf2surf --srcsubject average7 --trgsubject average7 \ --srcsurfval y-lh.mgh --src_type mgh \ --trg_type nii --trgsurfval y-lh.nii --hemi lh
You will also need to convert the design matrix produced by mris_glm (something like y.X.mat) into ascii. This can be done in matlab with something like:
load y.X.mat save('X.asc','X','-ascii')
Then run:
randomise -i y-lh -o y-lh \ -d X.asc -t design.con -n 5000 -V
Where design.con has your contrasts
The output will be something like: y-lh_max_tstat1.mgh, which you can view with tksurfer with something like:
tksurfer average7 lh inflated -overlay y-lh_max_tstat1.mgh
We're still working out the details on this (obviously:), you may have to play with this a little to get the command lines exactly correct.
Note that randomise program cannot do cluster-based thresholding because it is not aware that these values are really on the surface (not in a volume), but the max stat will work.
doug
Antao Du wrote:
Hi,
I am running mris_glm to compare the cortical thickness between two groups. I have a question, which method is used for correcting multiple comparison in the analysis? Thanks,
Antao
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
-- Douglas N. Greve, Ph.D. MGH-NMR Center greve@nmr.mgh.harvard.edu Phone Number: 617-724-2358 Fax: 617-726-7422
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi Bruce, I've managed to dig up some images where we have both a T1 and PD for the same person. Is the multi-channel segmentation implemeted in one of the dev versions? Is it simply a mater of averaging the T1 and PD instead of multiple T1s? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
________________________________
From: Bruce Fischl [mailto:fischl@nmr.mgh.harvard.edu] Sent: Mon 24/10/2005 10:19 PM To: Fornito, Alexander Cc: Freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] PDs?
Hi Alex,
unfortunately the current answer depends on what part of the brain you are interested in. For cortex, it appears that the raw CNR of mp-rage is significantly better than the multi-channel (T1/PD/T2*) CNR. On the other hand, subcortically for things like pallidum/putament, the multi-channel segmentation has much higher CNR.
We do have multi-channel aseg segmentation in place, but haven't used it a ton (since we have so much mp-rage data around).
cheers, Bruce
On Mon, 24 Oct 2005, Fornito, Alexander wrote:
Hi, We're about to start a longitudinal study and are keen to get accurate measures of structural change. Currently we are planning to acquire a T1 and T2, and have some spare time for another acuqisition. I've noticed in a recent paper (Fischl, et al (2004) Neuroimage,23, S69-S84) you discuss the benefits of combining T1 and PD images to aid segmentation. Just wandering if there are any plans to incorporate multichannel segmentation into freesurfer, and if so, would you recommed acquiring two T1s and averaging, or a single T1 and PD? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi Alex,
I don't think we distribute an atlas for T1/PD yet, so not really. And definitely don't average the T1s and PDs - that will remove almost all contrast.
Bruce On Sat, 29 Oct 2005, Fornito, Alexander wrote:
Hi Bruce, I've managed to dig up some images where we have both a T1 and PD for the same person. Is the multi-channel segmentation implemeted in one of the dev versions? Is it simply a mater of averaging the T1 and PD instead of multiple T1s? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
From: Bruce Fischl [mailto:fischl@nmr.mgh.harvard.edu] Sent: Mon 24/10/2005 10:19 PM To: Fornito, Alexander Cc: Freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] PDs?
Hi Alex,
unfortunately the current answer depends on what part of the brain you are interested in. For cortex, it appears that the raw CNR of mp-rage is significantly better than the multi-channel (T1/PD/T2*) CNR. On the other hand, subcortically for things like pallidum/putament, the multi-channel segmentation has much higher CNR.
We do have multi-channel aseg segmentation in place, but haven't used it a ton (since we have so much mp-rage data around).
cheers, Bruce
On Mon, 24 Oct 2005, Fornito, Alexander wrote:
Hi, We're about to start a longitudinal study and are keen to get accurate measures of structural change. Currently we are planning to acquire a T1 and T2, and have some spare time for another acuqisition. I've noticed in a recent paper (Fischl, et al (2004) Neuroimage,23, S69-S84) you discuss the benefits of combining T1 and PD images to aid segmentation. Just wandering if there are any plans to incorporate multichannel segmentation into freesurfer, and if so, would you recommed acquiring two T1s and averaging, or a single T1 and PD? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi Bruce and Alex.
We currently think we get more MR information by collecting T1/T2 multimodal scans for the sake of tissue-type segmentation. We have a lot of data with T1/T2 and 0.5 mm voxels, and even more data with T1/T2/PD and 1.015625 mm voxels. And we know how to massively run these studies through a T1-only FreeSurfer run.
We would be >very< interested in helping with any >multi-modal< atlas definition, both the 'aseg' and 'aparc' cases. Any guidance you could provide for running the multi-modal tissue-classifier included in Freesurfer would be a big help.
Alex, you can't simply average T1 and PD since in T1, white matter looks lighter than gray matter, which looks lighter than csf, while with PD/T2, white matter looks darkest, and csf is bright white. Combining different kinds of spectral scales permits the tissue classification to be more specific and sensitive.
Greg Harris University of Iowa Psychiatry Brain Imaging Lab
Hans Johnson wrote:
Please contact the freesurfer team (Bruce Fischl may be a starting point) and see if we can help generate a multi-modal atlas tuned the T1/PD data that we collect with our definitions of structures. This is really the only way we can make fair comparisons between the methods.
Hans
Fornito, Alexander wrote:
Hi Bruce, I've managed to dig up some images where we have both a T1 and PD for the same person. Is the multi-channel segmentation implemeted in one of the dev versions? Is it simply a mater of averaging the T1 and PD instead of multiple T1s? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
From: Bruce Fischl [mailto:fischl@nmr.mgh.harvard.edu] Sent: Mon 24/10/2005 10:19 PM To: Fornito, Alexander Cc: Freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] PDs?
Hi Alex,
unfortunately the current answer depends on what part of the brain you are interested in. For cortex, it appears that the raw CNR of mp-rage is significantly better than the multi-channel (T1/PD/T2*) CNR. On the other hand, subcortically for things like pallidum/putament, the multi-channel segmentation has much higher CNR.
We do have multi-channel aseg segmentation in place, but haven't used it a ton (since we have so much mp-rage data around).
cheers, Bruce
On Mon, 24 Oct 2005, Fornito, Alexander wrote:
Hi, We're about to start a longitudinal study and are keen to get accurate measures of structural change. Currently we are planning to acquire a T1 and T2, and have some spare time for another acuqisition. I've noticed in a recent paper (Fischl, et al (2004) Neuroimage,23, S69-S84) you discuss the benefits of combining T1 and PD images to aid segmentation. Just wandering if there are any plans to incorporate multichannel segmentation into freesurfer, and if so, would you recommed acquiring two T1s and averaging, or a single T1 and PD? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi Greg,
we have done a lot of work (mainly Xiao Han) looking at multi-spectral segmentation, and a couple of things are clear:
1. The raw CNR of mp-rage is better than T1/PD/T2* from SPGR sequences. 2. T2 data are hard to get with identical bandwidth and read-out direction to mp-rage, meaning that they will be distorted differently, significantly lowering their value. 3. T1/PD/T2* has better CNR than mp-rage for many subcortical structures (e.g. pallidum).
points #2 and #3 are things we are actively working on. Also, note that the cortical labeling is based on folds, not image intensities, so "multimodal" has no real meaning in that context. We have played with adding image intensities (and other things) to the labeling, but it hasn't been much help yet. Also note that T2 images would only be useful if they were manually labeled.
cheers, Bruce
On Fri, 28 Oct 2005, Greg Harris wrote:
Hi Bruce and Alex.
We currently think we get more MR information by collecting T1/T2 multimodal scans for the sake of tissue-type segmentation. We have a lot of data with T1/T2 and 0.5 mm voxels, and even more data with T1/T2/PD and 1.015625 mm voxels. And we know how to massively run these studies through a T1-only FreeSurfer run.
We would be >very< interested in helping with any >multi-modal< atlas definition, both the 'aseg' and 'aparc' cases. Any guidance you could provide for running the multi-modal tissue-classifier included in Freesurfer would be a big help.
Alex, you can't simply average T1 and PD since in T1, white matter looks lighter than gray matter, which looks lighter than csf, while with PD/T2, white matter looks darkest, and csf is bright white. Combining different kinds of spectral scales permits the tissue classification to be more specific and sensitive.
Greg Harris University of Iowa Psychiatry Brain Imaging Lab
Hans Johnson wrote:
Please contact the freesurfer team (Bruce Fischl may be a starting point) and see if we can help generate a multi-modal atlas tuned the T1/PD data that we collect with our definitions of structures. This is really the only way we can make fair comparisons between the methods.
Hans
Fornito, Alexander wrote:
Hi Bruce, I've managed to dig up some images where we have both a T1 and PD for the same person. Is the multi-channel segmentation implemeted in one of the dev versions? Is it simply a mater of averaging the T1 and PD instead of multiple T1s? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
From: Bruce Fischl [mailto:fischl@nmr.mgh.harvard.edu] Sent: Mon 24/10/2005 10:19 PM To: Fornito, Alexander Cc: Freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] PDs?
Hi Alex,
unfortunately the current answer depends on what part of the brain you are interested in. For cortex, it appears that the raw CNR of mp-rage is significantly better than the multi-channel (T1/PD/T2*) CNR. On the other hand, subcortically for things like pallidum/putament, the multi-channel segmentation has much higher CNR.
We do have multi-channel aseg segmentation in place, but haven't used it a ton (since we have so much mp-rage data around).
cheers, Bruce
On Mon, 24 Oct 2005, Fornito, Alexander wrote:
Hi, We're about to start a longitudinal study and are keen to get accurate measures of structural change. Currently we are planning to acquire a T1 and T2, and have some spare time for another acuqisition. I've noticed in a recent paper (Fischl, et al (2004) Neuroimage,23, S69-S84) you discuss the benefits of combining T1 and PD images to aid segmentation. Just wandering if there are any plans to incorporate multichannel segmentation into freesurfer, and if so, would you recommed acquiring two T1s and averaging, or a single T1 and PD? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi Greg and Bruce, I figure averaging the T1s and PDs would eliminate the contrast, just (incorrectly) figured that might be how the info would be combined in freesurfer if the multi-spectral segmentation was already included in the current dev versions.
From experience (and in line with Bruce's comments), 2 or more T1s seems to give more than enough contrast for the cortex. I was primarily interested in using the PDs for the subcortical aseg, as some of my labels seem too big. However, having followed your discussion on another thread, it seems that this may be due to the atlas labelling (eg., thalamus touching caudate) rather than an error in the algorithm for my images. Given that, I'm not sure the PD would add that much more ...
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
-----Original Message----- From: Greg Harris [mailto:Gregory-Harris@uiowa.edu] Sent: Sat 29/10/2005 9:32 AM To: Fornito, Alexander; Bruce Fischl Cc: Freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] PDs?
Hi Bruce and Alex.
We currently think we get more MR information by collecting T1/T2 multimodal scans for the sake of tissue-type segmentation. We have a lot of data with T1/T2 and 0.5 mm voxels, and even more data with T1/T2/PD and 1.015625 mm voxels. And we know how to massively run these studies through a T1-only FreeSurfer run.
We would be >very< interested in helping with any >multi-modal< atlas definition, both the 'aseg' and 'aparc' cases. Any guidance you could provide for running the multi-modal tissue-classifier included in Freesurfer would be a big help.
Alex, you can't simply average T1 and PD since in T1, white matter looks lighter than gray matter, which looks lighter than csf, while with PD/T2, white matter looks darkest, and csf is bright white. Combining different kinds of spectral scales permits the tissue classification to be more specific and sensitive.
Greg Harris University of Iowa Psychiatry Brain Imaging Lab
Hans Johnson wrote:
Please contact the freesurfer team (Bruce Fischl may be a starting point) and see if we can help generate a multi-modal atlas tuned the T1/PD data that we collect with our definitions of structures. This is really the only way we can make fair comparisons between the methods.
Hans
Fornito, Alexander wrote:
Hi Bruce, I've managed to dig up some images where we have both a T1 and PD for the same person. Is the multi-channel segmentation implemeted in one of the dev versions? Is it simply a mater of averaging the T1 and PD instead of multiple T1s? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
From: Bruce Fischl [mailto:fischl@nmr.mgh.harvard.edu] Sent: Mon 24/10/2005 10:19 PM To: Fornito, Alexander Cc: Freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] PDs?
Hi Alex,
unfortunately the current answer depends on what part of the brain you are interested in. For cortex, it appears that the raw CNR of mp-rage is significantly better than the multi-channel (T1/PD/T2*) CNR. On the other hand, subcortically for things like pallidum/putament, the multi-channel segmentation has much higher CNR.
We do have multi-channel aseg segmentation in place, but haven't used it a ton (since we have so much mp-rage data around).
cheers, Bruce
On Mon, 24 Oct 2005, Fornito, Alexander wrote:
Hi, We're about to start a longitudinal study and are keen to get accurate measures of structural change. Currently we are planning to acquire a T1 and T2, and have some spare time for another acuqisition. I've noticed in a recent paper (Fischl, et al (2004) Neuroimage,23, S69-S84) you discuss the benefits of combining T1 and PD images to aid segmentation. Just wandering if there are any plans to incorporate multichannel segmentation into freesurfer, and if so, would you recommed acquiring two T1s and averaging, or a single T1 and PD? Thanks, Alex
Alex Fornito M.Psych/PhD (clin. neuro.) candidate Melbourne Neuropsychiatry Centre and Department of Psychology The University of Melbourne alexander.fornito@wh.org.au
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
a.m.fjell＠psykologi.uio.no -
Antao Du -
Bruce Fischl -
Doug Greve -
Fornito, Alexander -
Greg Harris