External Email - Use Caution
Martin, thank you for these comments. These are extremely valuable!
Although my patients had tumors, these are cured (this is why I have so few of them…) and have no residual mass effect/deformity beyond biopsy tracks / small resection cavities. Thus far, I was able to get pretty accurate looking segmentations of deep nuclei on aseg. Segmentation issues were associated with the surface structures (e.g. dura/falx being misclassified as gray or brain tissue adjacent to resection cavities are correctly classified as brain but GM/WM often misclassified).
I’m OK restricting the analysis to ventricular and deep GM volume, since what I’m trying to measure is marked atrophy of these structure after chemotherapy which I guess is in the range of 1-5%/year. The changes are mostly marked enough to enable a visual qualitative analysis, however, quantitative data would be clinically more helpful to clarify when, and how much change to expect after treatment.
Do you think increasing the number of time points analyzed (I have 4-6 scans per year for each patient) could eliminate the noise from head positioning/slicing as long as the volumetric trends remain consistent? Again, the data would be normalized per individual and I would not use absolute volume for any comparative analysis.
Thank you again. This message board is of incredible value!
Best regards, David
----------------------------------------------------------------------
Message: 1 Date: Wed, 09 Oct 2019 10:48:48 +0200 From: Martin Reuter <mreuter@nmr.mgh.harvard.edumailto:mreuter@nmr.mgh.harvard.edu> Subject: Re: [Freesurfer] Questions re slice thickness, aseg and longitudinal analysis To: Freesurfer support list <freesurfer@nmr.mgh.harvard.edumailto:freesurfer@nmr.mgh.harvard.edu> Message-ID: <7a4d059cb5cb42f390e03f523677a3362d17cb81.camel@nmr.mgh.harvard.edumailto:7a4d059cb5cb42f390e03f523677a3362d17cb81.camel@nmr.mgh.harvard.edu> Content-Type: text/plain; charset="UTF-8"
Hi David,
I am not very optimistic:
5mm is too thick for FreeSurfer (recommendation is 1 up to 1.5). You will certainly get something, but it can be very unreliable and completely wrong. Especially longitudinally these thick slices will induce large variance due to different head positioning (and different slice angulations) in the scanner.
Furthermore, FreeSurfer does not take Gad-Enhanced images. Also it will not work if tumor lesions are present.
About your questions:
1. Surfaces update the aseg, but if you are only interested in the volumes, you can skip this expensive step (potentially at the cost of slightly higher noise levels in your measurements).
2. I think not (see above). 5mm is too low.
3. Theoretically yes, but I have never tested if the scripts will do it. You could run up to the aseg in the cross, then create base (up to aseg) and then run the longs up to aseg. Not even sure you really need the base aseg. You might be able to just run the initial base registration step, obtain the transformations and median norm.mgz image, could be sufficient for the long runs.
4. No. Gad images won't work.
Best, Martin
External Email - Use Caution Freesurfers,
First of all, I'd like to express my gratitude to the community for the support that keeps researchers like myself afloat!
I have a unique set of oncology patients that I want to evaluate for brain atrophy in a retrospective longitudinal analysis. I was thinking about using Aseg.auto results to assess longitudinal volume changes, but before I invest all the time I wanted to check with the community whether this makes any sense at all:
The dataset that looks like this: - 22 patients (no control dataset [yet]) - 10-25 MRIs per patient acquired over 2-8 years in relatively uniform intervals - Patients had most of their scans on the same scanner, but scanners differed widely between patients - All patients have axial T1 post gadolinium scans of 1x1x5mm resolution (3D acquisition available in <10%) - About 80% of scans have an axial pre-contrast T1 sequence - All scans are skullstripped (third party algorithm)
I'm looking for crude changes, no subtleties; volumes of interest are: - Whole brain volume - White matter volume - Ventricular volume (mainly lateral ventricle) - Subcortical gray matter volume (whole thalamus most importantly)
I ran a few test analyses and to my surprise I was able to generate pretty acceptable surfaces, however, topology fixing took about 24H per scan, and I feel aseg.auto contained all the volumetric data I was really interested in.
My concrete questions are: 1) Does the full autorecon pipeline affect Aseg.auto? If there is no benefit, I could reduce the per scan analysis time from 28 hours to 1-2 h. 2) Would this low-resolution dataset be accepted by reviewers if used for Aseg? Should I do any quantitative validation beyond a visual quality analysis of Aseg? 3) Can I perform a longitudinal analysis only for the Aseg results? 4) Is it OK to use T1-gad images for the analysis?
I'd appreciate any input!
Best regards, David O. Kamson, MD PhD Neuro-oncology fellow Johns Hopkins Hospital & National Institutes of Health
_______________________________________________ Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi David,
nobody can really tell you what will happen if you use the software outside its tested use cases. 5mm thick slices, with tumor resection is not what it has been designed to work with. And as I said, gad enhanced also not. It might give you meaningful results or not. The only way to find out is to try it and make sure to carefully inspect the segmentation visually (in all three views). Don’t be fooled by the nice looking within-plane images. They are smooth averages across 5mm. Missing resolution in the 3rd direction will allow you do only detect large volume changes due to loss of sensitivity. How much you loose is unclear.
Best, Martin
On 10. Oct 2019, at 21:11, David Kamson dkamson1@jhmi.edu wrote: Martin, thank you for these comments. These are extremely valuable!
Although my patients had tumors, these are cured (this is why I have so few of them…) and have no residual mass effect/deformity beyond biopsy tracks / small resection cavities. Thus far, I was able to get pretty accurate looking segmentations of deep nuclei on aseg. Segmentation issues were associated with the surface structures (e.g. dura/falx being misclassified as gray or brain tissue adjacent to resection cavities are correctly classified as brain but GM/WM often misclassified).
I’m OK restricting the analysis to ventricular and deep GM volume, since what I’m trying to measure is marked atrophy of these structure after chemotherapy which I guess is in the range of 1-5%/year. The changes are mostly marked enough to enable a visual qualitative analysis, however, quantitative data would be clinically more helpful to clarify when, and how much change to expect after treatment.
Do you think increasing the number of time points analyzed (I have 4-6 scans per year for each patient) could eliminate the noise from head positioning/slicing as long as the volumetric trends remain consistent? Again, the data would be normalized per individual and I would not use absolute volume for any comparative analysis.
Thank you again. This message board is of incredible value!
Best regards, David
Message: 1 Date: Wed, 09 Oct 2019 10:48:48 +0200 From: Martin Reuter <mreuter@nmr.mgh.harvard.edu mailto:mreuter@nmr.mgh.harvard.edu> Subject: Re: [Freesurfer] Questions re slice thickness, aseg and longitudinal analysis To: Freesurfer support list <freesurfer@nmr.mgh.harvard.edu mailto:freesurfer@nmr.mgh.harvard.edu> Message-ID: <7a4d059cb5cb42f390e03f523677a3362d17cb81.camel@nmr.mgh.harvard.edu mailto:7a4d059cb5cb42f390e03f523677a3362d17cb81.camel@nmr.mgh.harvard.edu> Content-Type: text/plain; charset="UTF-8"
Hi David,
I am not very optimistic:
5mm is too thick for FreeSurfer (recommendation is 1 up to 1.5). You will certainly get something, but it can be very unreliable and completely wrong. Especially longitudinally these thick slices will induce large variance due to different head positioning (and different slice angulations) in the scanner.
Furthermore, FreeSurfer does not take Gad-Enhanced images. Also it will not work if tumor lesions are present.
About your questions:
- Surfaces update the aseg, but if you are only interested in the
volumes, you can skip this expensive step (potentially at the cost of slightly higher noise levels in your measurements).
I think not (see above). 5mm is too low.
Theoretically yes, but I have never tested if the scripts will do
it. You could run up to the aseg in the cross, then create base (up to aseg) and then run the longs up to aseg. Not even sure you really need the base aseg. You might be able to just run the initial base registration step, obtain the transformations and median norm.mgz image, could be sufficient for the long runs.
- No. Gad images won't work.
Best, Martin
External Email - Use CautionFreesurfers,
First of all, I'd like to express my gratitude to the community for the support that keeps researchers like myself afloat!
I have a unique set of oncology patients that I want to evaluate for brain atrophy in a retrospective longitudinal analysis. I was thinking about using Aseg.auto results to assess longitudinal volume changes, but before I invest all the time I wanted to check with the community whether this makes any sense at all:
The dataset that looks like this:
- 22 patients (no control dataset [yet])
- 10-25 MRIs per patient acquired over 2-8 years in relatively
uniform intervals
- Patients had most of their scans on the same scanner, but
scanners differed widely between patients
- All patients have axial T1 post gadolinium scans of 1x1x5mm
resolution (3D acquisition available in <10%)
- About 80% of scans have an axial pre-contrast T1 sequence
- All scans are skullstripped (third party algorithm)
I'm looking for crude changes, no subtleties; volumes of interest are:
- Whole brain volume
- White matter volume
- Ventricular volume (mainly lateral ventricle)
- Subcortical gray matter volume (whole thalamus most importantly)
I ran a few test analyses and to my surprise I was able to generate pretty acceptable surfaces, however, topology fixing took about 24H per scan, and I feel aseg.auto contained all the volumetric data I was really interested in.
My concrete questions are:
- Does the full autorecon pipeline affect Aseg.auto? If there is no
benefit, I could reduce the per scan analysis time from 28 hours to 1-2 h. 2) Would this low-resolution dataset be accepted by reviewers if used for Aseg? Should I do any quantitative validation beyond a visual quality analysis of Aseg? 3) Can I perform a longitudinal analysis only for the Aseg results? 4) Is it OK to use T1-gad images for the analysis?
I'd appreciate any input!
Best regards, David O. Kamson, MD PhD Neuro-oncology fellow Johns Hopkins Hospital & National Institutes of Health
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu mailto:Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
David Kamson -
Martin Reuter