Re: [Freesurfer] freesufer v6 beta - T2 pial refinement performance and nsigma_below setting
Dear experts,
given sub-optimal results of my surfaces I referred to in my previous post, I have tested several different settings and input data preparation to obtain optimal pial surfaces in my modified HCP pipeline. I tested all variants in 5 representative subjects.
My modified HCP pipeline uses FreeSurfer V6beta and full hires recon (-cm option) with 0.7mm3 data (including FreeSurferHiResWhite.sh and FreeSurferHiResPial.sh where effectively no up and downsampling is done since all steps are done in original 0.7 mm3 resolution).
I have played with input data and parameters of 2nd pass mris_make_surfaces -T2pial in FreeSurferHiResPial.sh Since FreeSurferHiResPial.sh does not use aseg-informed and white surface-informed normalization and brain masking of T2 (which is default in recon-all), I also tested performance of this option (using commands borrowed from recon-all -T2pial). Since -nsigma_below 3 (default in FreeSurferHiResPial.sh) with combination of freeSurfer V6beta version of mris_make_surfaces cuts out some portions of gray matter in my data (in contrast to FreeSurfer 5.3 version where it seems that nsigma_below 3 suffices!), I increased -nsigma_below to 5 (which is default value used in recon-all -T2pial).
My tested versions were following: For modified full hires HCP pipeline with FreeSurfer V6beta: 1. -nsigma_below 3 (original FreeSurferHiResPial.sh) 2. -nsigma_below 5 3. -nsigma_below 5 with additional masking of T2 by brainmask.mgz 4. -nsigma_below 5 with aseg-informed normalization of T2 (using mri_normalize) 5. -nsigma_below 5 with aseg-informed normalization of T2 (using mri_normalize) and masking of T2 by brainmask.mgz
And, as Matt suggested, I also tried: 6. standard FreeSurfer V6 beta recon-all -pial and -T2pial (apart from standard recon-all I used as ?h.white input the output of FreeSurferHiResWhite.sh , but I think it should not matter in this context).
My observations are following:
It is definitely needed to increase -nsigma_below from 3 due to significant cut out of portions of gray matter in some regions, in some subjects these regions are very large.
Increase nsigma_below to 5 prevents cutting out of gray matter, however it leads to leak of surfaces to cerebellum (as I showed in my previous post) for all my versions (to varying extent). I am not sure how much significant the extend of error is, looking subjective it seems significant, but accounting the fact that the data is of high resolution, the maximal error would be approx. 2 mm.
-nsigma_below 5 with aseg-informed normalization does not work well in some areas and results in cutting out gray matter in some regions (but not to such great extend as in nsigma_below 3)
FreeSurfer V6beta tends in some (not very large) areas to cut out gray matter. On the other hand, in some areas and in some subjects it produces largest leak of surfaces to dura and cerebellum from all my tested versions.
mri_normalize of T2 has in some regions sub-optimal performance on pial surface estimation and actually performs worse than global (no aseg-informed and no white surface-informed) normalization ( especially at temporal poles, but also in some areas at convexity).
I have indecisive results with masking. In some subjects it had almost no effect, in one subject it definitely improved the leak to dura, but in some subjects the results were quite opposite. I am not sure why the masking should matter since the mask is larger than the regions where the surfaces are extending, but I am not very familiar with internals of mris_make_surfaces (the masked T2 maybe after internal smoothing in mris_make_surfaces can propagate the voxel values even to regions relevant for surface estimation). Therefore I would tend to use the masking.
Overall best results (better than with V6beta recon-all) I obtained with -nsigma_below 5 with global (no aseg-informed and white surface-informed) normalization of T2 as it is done in FreeSurferHiResPial.sh.
My overall impresion: I was quite frustrated to observe, how much the results are dependent on particular setting of the parameters and input data. But maybe my expectations of accuracy of surfaces were unrealistically high....
Regards,
Antonin
Hi Antonin,
In our experience with FreeSurfer 5.3 we were getting very consistent results. We are going to try some of this locally and work with the FreeSurfer folks to get the T2w stream working robustly in FreeSurfer v6.0+.
Thanks for testing this,
Matt.
From: freesurfer-bounces@nmr.mgh.harvard.edu on behalf of Antonin Skoch ansk@ikem.cz Reply-To: Freesurfer support list freesurfer@nmr.mgh.harvard.edu Date: Wednesday, November 30, 2016 at 3:31 PM To: freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] freesufer v6 beta - T2 pial refinement performance and nsigma_below setting
Dear experts,
given sub-optimal results of my surfaces I referred to in my previous post, I have tested several different settings and input data preparation to obtain optimal pial surfaces in my modified HCP pipeline. I tested all variants in 5 representative subjects.
My modified HCP pipeline uses FreeSurfer V6beta and full hires recon (-cm option) with 0.7mm3 data (including FreeSurferHiResWhite.sh and FreeSurferHiResPial.sh where effectively no up and downsampling is done since all steps are done in original 0.7 mm3 resolution).
I have played with input data and parameters of 2nd pass mris_make_surfaces -T2pial in FreeSurferHiResPial.sh Since FreeSurferHiResPial.sh does not use aseg-informed and white surface-informed normalization and brain masking of T2 (which is default in recon-all), I also tested performance of this option (using commands borrowed from recon-all -T2pial). Since -nsigma_below 3 (default in FreeSurferHiResPial.sh) with combination of freeSurfer V6beta version of mris_make_surfaces cuts out some portions of gray matter in my data (in contrast to FreeSurfer 5.3 version where it seems that nsigma_below 3 suffices!), I increased -nsigma_below to 5 (which is default value used in recon-all -T2pial).
My tested versions were following: For modified full hires HCP pipeline with FreeSurfer V6beta: 1. -nsigma_below 3 (original FreeSurferHiResPial.sh) 2. -nsigma_below 5 3. -nsigma_below 5 with additional masking of T2 by brainmask.mgz 4. -nsigma_below 5 with aseg-informed normalization of T2 (using mri_normalize) 5. -nsigma_below 5 with aseg-informed normalization of T2 (using mri_normalize) and masking of T2 by brainmask.mgz
And, as Matt suggested, I also tried: 6. standard FreeSurfer V6 beta recon-all -pial and -T2pial (apart from standard recon-all I used as ?h.white input the output of FreeSurferHiResWhite.sh , but I think it should not matter in this context).
My observations are following:
It is definitely needed to increase -nsigma_below from 3 due to significant cut out of portions of gray matter in some regions, in some subjects these regions are very large.
Increase nsigma_below to 5 prevents cutting out of gray matter, however it leads to leak of surfaces to cerebellum (as I showed in my previous post) for all my versions (to varying extent). I am not sure how much significant the extend of error is, looking subjective it seems significant, but accounting the fact that the data is of high resolution, the maximal error would be approx. 2 mm.
-nsigma_below 5 with aseg-informed normalization does not work well in some areas and results in cutting out gray matter in some regions (but not to such great extend as in nsigma_below 3)
FreeSurfer V6beta tends in some (not very large) areas to cut out gray matter. On the other hand, in some areas and in some subjects it produces largest leak of surfaces to dura and cerebellum from all my tested versions.
mri_normalize of T2 has in some regions sub-optimal performance on pial surface estimation and actually performs worse than global (no aseg-informed and no white surface-informed) normalization ( especially at temporal poles, but also in some areas at convexity).
I have indecisive results with masking. In some subjects it had almost no effect, in one subject it definitely improved the leak to dura, but in some subjects the results were quite opposite. I am not sure why the masking should matter since the mask is larger than the regions where the surfaces are extending, but I am not very familiar with internals of mris_make_surfaces (the masked T2 maybe after internal smoothing in mris_make_surfaces can propagate the voxel values even to regions relevant for surface estimation). Therefore I would tend to use the masking.
Overall best results (better than with V6beta recon-all) I obtained with -nsigma_below 5 with global (no aseg-informed and white surface-informed) normalization of T2 as it is done in FreeSurferHiResPial.sh.
My overall impresion: I was quite frustrated to observe, how much the results are dependent on particular setting of the parameters and input data. But maybe my expectations of accuracy of surfaces were unrealistically high....
Regards,
Antonin _______________________________________________ Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer 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.
freesurfer@nmr.mgh.harvard.edu
-
Antonin Skoch -
Matt Glasser