[HCP-Users] inhomogeneous signal intensity
Hi HCP users, When we looked at the resting state fMRI data (minimally preprocessed), we found that there are pretty obvious signal intensity difference between brain areas. For example, the frontal areas have significantly higher signal intensity than other brain areas (see the attached figure). We are wondering whether this is considered as artifacts that need special treatment? Thanks, Tianwen ___ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users
Re: [HCP-Users] rsfc preprocessing
Thanks Matt! This is really helpful. We weren't necessarily going to use a parcellated analysis, since we have the disk space for voxel by voxel -- although it seems like that's what many of the sub graph detection results are for. However, perhaps we should do that if we are going to skip filtering. Is it a *bad* idea to use just fsl's bandpass filter instead and use voxel by voxel instead of using a regression paradigm? The reason I ask is that we are looking to replicate voxel by voxel community detection results with a different type of factor analysis. I will check out fslnets, that seems really perfect. Thanks again! Kim On Sunday, April 12, 2015, Glasser, Matthew glass...@wusm.wustl.edu wrote: Unfortunately the really good way to deal with these issues for dense data has not yet been released (basically improvements upon the MIGP algorithm). I’m not sure what the release plan is on that, so perhaps someone else can comment on that. Generally though, if you can set up your problem as a regression problem rather than a correlation problem (and your regression design matrix is from ROI(s) or a weighted component (s)), you don’t need to worry about spatial or temporal filtering (assuming you’ve properly cleaned your data using movement parameter regression and ICA+FIX—something the HCP does for you in the FIX cleaned packages) to get high quality spatial maps. Additionally, if you parcellate your data, you don’t need to do spatial or temporal filtering even when doing correlation, as parcellation is a great way to reduce unstructured noise in a neurobiologically valid way. It sounds like you might already be doing parceallated analyses. As for combining across subjects, generally one simply averages across runs after converting to Z scores and does a t-test across subjects. I’d have a look at FSLNets, as it has a good implementation of these things. Peace, Matt. From: Kimberly Stachenfeld k...@princeton.edu javascript:_e(%7B%7D,'cvml','k...@princeton.edu'); Date: Sunday, April 12, 2015 at 4:22 PM To: hcp-users@humanconnectome.org javascript:_e(%7B%7D,'cvml','hcp-users@humanconnectome.org'); hcp-users@humanconnectome.org javascript:_e(%7B%7D,'cvml','hcp-users@humanconnectome.org'); Subject: [HCP-Users] rsfc preprocessing Hi hcp-users, I'm new to resting state connectivity analysis (and this list-serve), and I have a few basic questions about applying it to the HCP data. I'm using the minimally preprocessed REST1 data. 1. The low-pass filtering seems controversial, though commonly employed -- is there at this point an agreed-upon way to remove deleterious high frequency noise? In addition, I'm having difficulty with temporally filtering the data in fsl. To bandpass data from .009 - .08 Hz, I'm running: fslmaths nii_in -bptf 77.168.68 nii_out where 77.16 = sigma_hipass = 1/(2 * TR * F_hicutoff), for TR = .72 and F_cutoff = .009 and 8.68 = sigma_lopass = 1/(2 * TR * F_locutoff), for TR = .72 and F_locutoff = .08 This *seems* correct (I at least confirmed with the feat gui that the conversion from cycle time to sigma is 1/(2*TR)). However, when I look at the data in the frequency domain, it looks like there is significant response left for frequencies below .009 Hz (picture attached) and very little between .01-.08. Does anyone know if I'm doing something incorrectly, or if the frequency cutoff for a Gaussian filter is just very gradual? 2. Any additional preprocessing is recommended, besides temporal filter and what the minimal pre-processing has already enacted? 3. What is an intelligent way to combine correlation matrices? Averaging (Power et al, 2011)? Binarizing the correlation matrix by setting the top 10% of voxels to 1 and the rest to 0, and averaging the binarized matrices (Yeo et al, 2011)? Either? Something fancier? Any advice or additional resources would be enormously appreciated -- thanks very much!! Kim -- Kimberly Stachenfeld, BS Graduate Student 236A Princeton Neuroscience Institute Washington Road Princeton, NJ 08544 k...@princeton.edu javascript:_e(%7B%7D,'cvml','k...@princeton.edu'); ___ HCP-Users mailing list HCP-Users@humanconnectome.org javascript:_e(%7B%7D,'cvml','HCP-Users@humanconnectome.org'); http://lists.humanconnectome.org/mailman/listinfo/hcp-users -- The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail. -- Kimberly Stachenfeld, BS Graduate Student 236A Princeton Neuroscience Institute
Re: [HCP-Users] rsfc preprocessing
Came across this recently in TaskfMRIAnalysis/scripts/TaskfMRILevel1.sh: AdditionalSigma=`echo $AdditionalSmoothingFWHM / ( 2 * ( sqrt ( 2 * l ( 2 ) ) ) ) | bc -l` So even more precise.;-) On Apr 13, 2015, at 8:59 AM, Harms, Michael mha...@wustl.edu wrote: Hi, Just wanted to mention, for purposes of documenting in this thread, that technically the conversion from FWHM to sigma is: sigma = FWHM/2.355 http://en.wikipedia.org/wiki/Full_width_at_half_maximum I believe the FEAT just uses 2 rather than 2.355 in the denominator for the calculation of the sigma for its high pass temporal filter because the Gaussian filter is very gradual anyway. In the HCP pipelines, I believe that we use the technically correct factor of 1/2.355 for any conversion of FWHM to Sigma. cheers, -MH -- Michael Harms, Ph.D. --- Conte Center for the Neuroscience of Mental Disorders Washington University School of Medicine Department of Psychiatry, Box 8134 660 South Euclid Ave. Tel: 314-747-6173 St. Louis, MO 63110 Email: mha...@wustl.edu From: Kimberly Stachenfeld k...@princeton.edu Date: Sunday, April 12, 2015 4:22 PM To: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: [HCP-Users] rsfc preprocessing Hi hcp-users, I'm new to resting state connectivity analysis (and this list-serve), and I have a few basic questions about applying it to the HCP data. I'm using the minimally preprocessed REST1 data. 1. The low-pass filtering seems controversial, though commonly employed -- is there at this point an agreed-upon way to remove deleterious high frequency noise? In addition, I'm having difficulty with temporally filtering the data in fsl. To bandpass data from .009 - .08 Hz, I'm running: fslmaths nii_in -bptf 77.168.68 nii_out where 77.16 = sigma_hipass = 1/(2 * TR * F_hicutoff), for TR = .72 and F_cutoff = .009 and 8.68 = sigma_lopass = 1/(2 * TR * F_locutoff), for TR = .72 and F_locutoff = .08 This seems correct (I at least confirmed with the feat gui that the conversion from cycle time to sigma is 1/(2*TR)). However, when I look at the data in the frequency domain, it looks like there is significant response left for frequencies below .009 Hz (picture attached) and very little between .01-.08. Does anyone know if I'm doing something incorrectly, or if the frequency cutoff for a Gaussian filter is just very gradual? 2. Any additional preprocessing is recommended, besides temporal filter and what the minimal pre-processing has already enacted? 3. What is an intelligent way to combine correlation matrices? Averaging (Power et al, 2011)? Binarizing the correlation matrix by setting the top 10% of voxels to 1 and the rest to 0, and averaging the binarized matrices (Yeo et al, 2011)? Either? Something fancier? Any advice or additional resources would be enormously appreciated -- thanks very much!! Kim -- Kimberly Stachenfeld, BS Graduate Student 236A Princeton Neuroscience Institute Washington Road Princeton, NJ 08544 k...@princeton.edu ___ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail. ___ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users ___ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users
Re: [HCP-Users] rsfc preprocessing
Ok, correction on myself. That factor of 1/2.355 (obtained using the formula Donna provided) is what we use for conversion of FWHM to Sigma for *spatial* filtering. However, it appears that for the *temporal* filtering that we still use the approximation of 1/2 = 0.5 in the relevant HCPpipelinecode for setting up the sigma for the high pass temporal filter. e.g., in inTaskfMRIAnalysis/scripts/TaskfMRILevel1.sh: fslmaths ${fake_nifti_file} -bptf `echo 0.5 * $TemporalFilter / $TR_vol | bc -l` 0 ${fake_nifti_file} and in hcp_fix hptr=`echo 10 k $hp 2 / $tr / p | dc -` ${FSLDIR}/bin/fslmaths $fmri -bptf $hptr -1 ${fmri}_hp$hp Sorry about that. cheers, -MH -- Michael Harms, Ph.D. --- Conte Center for the Neuroscience of Mental Disorders Washington University School of Medicine Department of Psychiatry, Box 8134 660 South Euclid Ave. Tel: 314-747-6173 St. Louis, MO 63110 Email: mha...@wustl.edu On 4/13/15 9:13 AM, Donna Dierker do...@brainvis.wustl.edu wrote: Came across this recently in TaskfMRIAnalysis/scripts/TaskfMRILevel1.sh: AdditionalSigma=`echo $AdditionalSmoothingFWHM / ( 2 * ( sqrt ( 2 * l ( 2 ) ) ) ) | bc -l` So even more precise.;-) On Apr 13, 2015, at 8:59 AM, Harms, Michael mha...@wustl.edu wrote: Hi, Just wanted to mention, for purposes of documenting in this thread, that technically the conversion from FWHM to sigma is: sigma = FWHM/2.355 http://en.wikipedia.org/wiki/Full_width_at_half_maximum I believe the FEAT just uses 2 rather than 2.355 in the denominator for the calculation of the sigma for its high pass temporal filter because the Gaussian filter is very gradual anyway. In the HCP pipelines, I believe that we use the technically correct factor of 1/2.355 for any conversion of FWHM to Sigma. cheers, -MH -- Michael Harms, Ph.D. --- Conte Center for the Neuroscience of Mental Disorders Washington University School of Medicine Department of Psychiatry, Box 8134 660 South Euclid Ave.Tel: 314-747-6173 St. Louis, MO63110Email: mha...@wustl.edu From: Kimberly Stachenfeld k...@princeton.edu Date: Sunday, April 12, 2015 4:22 PM To: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: [HCP-Users] rsfc preprocessing Hi hcp-users, I'm new to resting state connectivity analysis (and this list-serve), and I have a few basic questions about applying it to the HCP data. I'm using the minimally preprocessed REST1 data. 1. The low-pass filtering seems controversial, though commonly employed -- is there at this point an agreed-upon way to remove deleterious high frequency noise? In addition, I'm having difficulty with temporally filtering the data in fsl. To bandpass data from .009 - .08 Hz, I'm running: fslmaths nii_in -bptf 77.168.68 nii_out where 77.16 = sigma_hipass = 1/(2 * TR * F_hicutoff), for TR = .72 and F_cutoff = .009 and8.68 = sigma_lopass = 1/(2 * TR * F_locutoff), for TR = .72 and F_locutoff = .08 This seems correct (I at least confirmed with the feat gui that the conversion from cycle time to sigma is 1/(2*TR)). However, when I look at the data in the frequency domain, it looks like there is significant response left for frequencies below .009 Hz (picture attached) and very little between .01-.08. Does anyone know if I'm doing something incorrectly, or if the frequency cutoff for a Gaussian filter is just very gradual? 2. Any additional preprocessing is recommended, besides temporal filter and what the minimal pre-processing has already enacted? 3. What is an intelligent way to combine correlation matrices? Averaging (Power et al, 2011)? Binarizing the correlation matrix by setting the top 10% of voxels to 1 and the rest to 0, and averaging the binarized matrices (Yeo et al, 2011)? Either? Something fancier? Any advice or additional resources would be enormously appreciated -- thanks very much!! Kim -- Kimberly Stachenfeld, BS Graduate Student 236A Princeton Neuroscience Institute Washington Road Princeton, NJ 08544 k...@princeton.edu ___ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail. ___ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users The materials in this message are private and may contain Protected Healthcare
Re: [HCP-Users] rsfc preprocessing
Hi, Just wanted to mention, for purposes of documenting in this thread, that technically the conversion from FWHM to sigma is: sigma = FWHM/2.355 http://en.wikipedia.org/wiki/Full_width_at_half_maximum I believe the FEAT just uses 2 rather than 2.355 in the denominator for the calculation of the sigma for its high pass temporal filter because the Gaussian filter is very gradual anyway. In the HCP pipelines, I believe that we use the technically correct factor of 1/2.355 for any conversion of FWHM to Sigma. cheers, -MH -- Michael Harms, Ph.D. --- Conte Center for the Neuroscience of Mental Disorders Washington University School of Medicine Department of Psychiatry, Box 8134 660 South Euclid Ave. Tel: 314-747-6173 St. Louis, MO 63110 Email: mha...@wustl.edu From: Kimberly Stachenfeld k...@princeton.edu Date: Sunday, April 12, 2015 4:22 PM To: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: [HCP-Users] rsfc preprocessing Hi hcp-users, I'm new to resting state connectivity analysis (and this list-serve), and I have a few basic questions about applying it to the HCP data. I'm using the minimally preprocessed REST1 data. 1. The low-pass filtering seems controversial, though commonly employed -- is there at this point an agreed-upon way to remove deleterious high frequency noise? In addition, I'm having difficulty with temporally filtering the data in fsl. To bandpass data from .009 - .08 Hz, I'm running: fslmaths nii_in -bptf77.168.68 nii_out where 77.16 = sigma_hipass = 1/(2 * TR * F_hicutoff), for TR = .72 and F_cutoff = .009 and 8.68 = sigma_lopass = 1/(2 * TR * F_locutoff), for TR = .72 and F_locutoff = .08 This seems correct (I at least confirmed with the feat gui that the conversion from cycle time to sigma is 1/(2*TR)). However, when I look at the data in the frequency domain, it looks like there is significant response left for frequencies below .009 Hz (picture attached) and very little between .01-.08. Does anyone know if I'm doing something incorrectly, or if the frequency cutoff for a Gaussian filter is just very gradual? 2. Any additional preprocessing is recommended, besides temporal filter and what the minimal pre-processing has already enacted? 3. What is an intelligent way to combine correlation matrices? Averaging (Power et al, 2011)? Binarizing the correlation matrix by setting the top 10% of voxels to 1 and the rest to 0, and averaging the binarized matrices (Yeo et al, 2011)? Either? Something fancier? Any advice or additional resources would be enormously appreciated -- thanks very much!! Kim -- Kimberly Stachenfeld, BS Graduate Student 236A Princeton Neuroscience Institute Washington Road Princeton, NJ 08544 k...@princeton.edu ___ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail. ___HCP-Users mailing listHCP-Users@humanconnectome.orghttp://lists.humanconnectome.org/mailman/listinfo/hcp-users
Re: [HCP-Users] rsfc preprocessing
That seems like a good reason -- can you refer me somewhere for the better alternatives? We're on a bit of a timeline so I am hesitant to wait for the HCP version. Thanks again for all the help! On Mon, Apr 13, 2015 at 12:34 PM, Glasser, Matthew glass...@wusm.wustl.edu wrote: Well, I wouldn’t do it that way since there are a variety of better alternatives that deal with the issue in a more targeted way. In any case, we should have a better solution for dense analyses in the future. Peace, Matt. From: Kimberly Stachenfeld k...@princeton.edu Date: Monday, April 13, 2015 at 6:48 AM To: Matt Glasser glass...@wusm.wustl.edu Cc: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: Re: rsfc preprocessing Thanks Matt! This is really helpful. We weren't necessarily going to use a parcellated analysis, since we have the disk space for voxel by voxel -- although it seems like that's what many of the sub graph detection results are for. However, perhaps we should do that if we are going to skip filtering. Is it a *bad* idea to use just fsl's bandpass filter instead and use voxel by voxel instead of using a regression paradigm? The reason I ask is that we are looking to replicate voxel by voxel community detection results with a different type of factor analysis. I will check out fslnets, that seems really perfect. Thanks again! Kim On Sunday, April 12, 2015, Glasser, Matthew glass...@wusm.wustl.edu wrote: Unfortunately the really good way to deal with these issues for dense data has not yet been released (basically improvements upon the MIGP algorithm). I’m not sure what the release plan is on that, so perhaps someone else can comment on that. Generally though, if you can set up your problem as a regression problem rather than a correlation problem (and your regression design matrix is from ROI(s) or a weighted component (s)), you don’t need to worry about spatial or temporal filtering (assuming you’ve properly cleaned your data using movement parameter regression and ICA+FIX—something the HCP does for you in the FIX cleaned packages) to get high quality spatial maps. Additionally, if you parcellate your data, you don’t need to do spatial or temporal filtering even when doing correlation, as parcellation is a great way to reduce unstructured noise in a neurobiologically valid way. It sounds like you might already be doing parceallated analyses. As for combining across subjects, generally one simply averages across runs after converting to Z scores and does a t-test across subjects. I’d have a look at FSLNets, as it has a good implementation of these things. Peace, Matt. From: Kimberly Stachenfeld k...@princeton.edu Date: Sunday, April 12, 2015 at 4:22 PM To: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: [HCP-Users] rsfc preprocessing Hi hcp-users, I'm new to resting state connectivity analysis (and this list-serve), and I have a few basic questions about applying it to the HCP data. I'm using the minimally preprocessed REST1 data. 1. The low-pass filtering seems controversial, though commonly employed -- is there at this point an agreed-upon way to remove deleterious high frequency noise? In addition, I'm having difficulty with temporally filtering the data in fsl. To bandpass data from .009 - .08 Hz, I'm running: fslmaths nii_in -bptf 77.168.68 nii_out where 77.16 = sigma_hipass = 1/(2 * TR * F_hicutoff), for TR = .72 and F_cutoff = .009 and 8.68 = sigma_lopass = 1/(2 * TR * F_locutoff), for TR = .72 and F_locutoff = .08 This *seems* correct (I at least confirmed with the feat gui that the conversion from cycle time to sigma is 1/(2*TR)). However, when I look at the data in the frequency domain, it looks like there is significant response left for frequencies below .009 Hz (picture attached) and very little between .01-.08. Does anyone know if I'm doing something incorrectly, or if the frequency cutoff for a Gaussian filter is just very gradual? 2. Any additional preprocessing is recommended, besides temporal filter and what the minimal pre-processing has already enacted? 3. What is an intelligent way to combine correlation matrices? Averaging (Power et al, 2011)? Binarizing the correlation matrix by setting the top 10% of voxels to 1 and the rest to 0, and averaging the binarized matrices (Yeo et al, 2011)? Either? Something fancier? Any advice or additional resources would be enormously appreciated -- thanks very much!! Kim -- Kimberly Stachenfeld, BS Graduate Student 236A Princeton Neuroscience Institute Washington Road Princeton, NJ 08544 k...@princeton.edu ___ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users -- The materials in this message are private and may contain
Re: [HCP-Users] rsfc preprocessing
I listed several that are publicly available. Peace, Matt. From: Kimberly Stachenfeld k...@princeton.edu Date: Monday, April 13, 2015 at 11:41 AM To: Matt Glasser glass...@wusm.wustl.edu Cc: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: Re: rsfc preprocessing That seems like a good reason -- can you refer me somewhere for the better alternatives? We're on a bit of a timeline so I am hesitant to wait for the HCP version. Thanks again for all the help! On Mon, Apr 13, 2015 at 12:34 PM, Glasser, Matthew glass...@wusm.wustl.edu wrote: Well, I wouldn’t do it that way since there are a variety of better alternatives that deal with the issue in a more targeted way. In any case, we should have a better solution for dense analyses in the future. Peace, Matt. From: Kimberly Stachenfeld k...@princeton.edu Date: Monday, April 13, 2015 at 6:48 AM To: Matt Glasser glass...@wusm.wustl.edu Cc: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: Re: rsfc preprocessing Thanks Matt! This is really helpful. We weren't necessarily going to use a parcellated analysis, since we have the disk space for voxel by voxel -- although it seems like that's what many of the sub graph detection results are for.However, perhaps we should do that if we are going to skip filtering. Is it a *bad* idea to use just fsl's bandpass filter instead and usevoxel by voxel instead of using a regression paradigm? The reason I ask is that we are looking to replicate voxel by voxelcommunity detection results with a different type of factoranalysis. I will check out fslnets, that seems really perfect. Thanks again! Kim On Sunday, April 12, 2015, Glasser, Matthew glass...@wusm.wustl.edu wrote: Unfortunately the really good way to deal with these issues for dense data has not yet been released (basically improvements upon the MIGP algorithm). I’m not sure what the release plan is on that, so perhaps someone else can comment on that. Generally though, if you can set up your problem as a regression problem rather than a correlation problem (and your regression design matrix is from ROI(s) or a weighted component (s)), you don’t need to worry about spatial or temporal filtering (assuming you’ve properly cleaned your data using movement parameter regression and ICAFIX—something the HCP does for you in the FIX cleaned packages) to get high quality spatial maps. Additionally, if you parcellate your data, you don’t need to do spatial or temporal filtering even when doing correlation, as parcellation is a great way to reduce unstructured noise in a neurobiologically valid way. It sounds like you might already be doing parceallated analyses. As for combining across subjects, generally one simply averages across runs after converting to Z scores and does a t-test across subjects. I’d have a look at FSLNets, as it has a good implementation of these things. Peace, Matt. From: Kimberly Stachenfeld k...@princeton.edu Date: Sunday, April 12, 2015 at 4:22 PM To: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: [HCP-Users] rsfc preprocessing Hi hcp-users, I'm new to resting state connectivity analysis (and this list-serve), and I have a few basic questions about applying it to the HCP data. I'm using the minimally preprocessed REST1 data. 1. The low-pass filtering seems controversial, though commonly employed -- is there at this point an agreed-upon way to remove deleterious high frequency noise? In addition, I'm having difficulty with temporally filtering the data in fsl. To bandpass data from .009 - .08 Hz, I'm running: fslmaths nii_in -bptf77.168.68 nii_out where 77.16 = sigma_hipass = 1/(2 * TR * F_hicutoff), for TR = .72 and F_cutoff = .009 and 8.68 = sigma_lopass = 1/(2 * TR * F_locutoff), for TR = .72 and F_locutoff = .08 This seems correct (I at least confirmed with the feat gui that the conversion from cycle time to sigma is 1/(2*TR)). However, when I look at the data in the frequency domain, it looks like there is significant response left for frequencies below .009 Hz (picture attached) and very little between .01-.08. Does anyone know if I'm doing something incorrectly, or if the frequency cutoff for a Gaussian filter is just very gradual? 2. Any additional preprocessing is recommended, besides temporal filter and what the minimal pre-processing has already enacted? 3. What is an intelligent way to combine correlation matrices? Averaging (Power et al, 2011)? Binarizing the correlation matrix by setting the top 10% of voxels to 1 and the rest to 0, and averaging the binarized matrices (Yeo et al, 2011)? Either? Something fancier? Any advice or additional resources would be enormously appreciated -- thanks very much!! Kim -- Kimberly Stachenfeld, BS Graduate Student 236A Princeton Neuroscience Institute Washington Road Princeton, NJ 08544 k...@princeton.edu
Re: [HCP-Users] rsfc preprocessing
Well, I wouldn’t do it that way since there are a variety of better alternatives that deal with the issue in a more targeted way. In any case, we should have a better solution for dense analyses in the future. Peace, Matt. From: Kimberly Stachenfeld k...@princeton.edu Date: Monday, April 13, 2015 at 6:48 AM To: Matt Glasser glass...@wusm.wustl.edu Cc: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: Re: rsfc preprocessing Thanks Matt! This is really helpful. We weren't necessarily going to use a parcellated analysis, since we have the disk space for voxel by voxel -- although it seems like that's what many of the sub graph detection results are for.However, perhaps we should do that if we are going to skip filtering. Is it a *bad* idea to use just fsl's bandpass filter instead and usevoxel by voxel instead of using a regression paradigm? The reason I ask is that we are looking to replicate voxel by voxelcommunity detection results with a different type of factoranalysis. I will check out fslnets, that seems really perfect. Thanks again! Kim On Sunday, April 12, 2015, Glasser, Matthew glass...@wusm.wustl.edu wrote: Unfortunately the really good way to deal with these issues for dense data has not yet been released (basically improvements upon the MIGP algorithm). I’m not sure what the release plan is on that, so perhaps someone else can comment on that. Generally though, if you can set up your problem as a regression problem rather than a correlation problem (and your regression design matrix is from ROI(s) or a weighted component (s)), you don’t need to worry about spatial or temporal filtering (assuming you’ve properly cleaned your data using movement parameter regression and ICAFIX—something the HCP does for you in the FIX cleaned packages) to get high quality spatial maps. Additionally, if you parcellate your data, you don’t need to do spatial or temporal filtering even when doing correlation, as parcellation is a great way to reduce unstructured noise in a neurobiologically valid way. It sounds like you might already be doing parceallated analyses. As for combining across subjects, generally one simply averages across runs after converting to Z scores and does a t-test across subjects. I’d have a look at FSLNets, as it has a good implementation of these things. Peace, Matt. From: Kimberly Stachenfeld k...@princeton.edu Date: Sunday, April 12, 2015 at 4:22 PM To: hcp-users@humanconnectome.org hcp-users@humanconnectome.org Subject: [HCP-Users] rsfc preprocessing Hi hcp-users, I'm new to resting state connectivity analysis (and this list-serve), and I have a few basic questions about applying it to the HCP data. I'm using the minimally preprocessed REST1 data. 1. The low-pass filtering seems controversial, though commonly employed -- is there at this point an agreed-upon way to remove deleterious high frequency noise? In addition, I'm having difficulty with temporally filtering the data in fsl. To bandpass data from .009 - .08 Hz, I'm running: fslmaths nii_in -bptf77.168.68 nii_out where 77.16 = sigma_hipass = 1/(2 * TR * F_hicutoff), for TR = .72 and F_cutoff = .009 and 8.68 = sigma_lopass = 1/(2 * TR * F_locutoff), for TR = .72 and F_locutoff = .08 This seems correct (I at least confirmed with the feat gui that the conversion from cycle time to sigma is 1/(2*TR)). However, when I look at the data in the frequency domain, it looks like there is significant response left for frequencies below .009 Hz (picture attached) and very little between .01-.08. Does anyone know if I'm doing something incorrectly, or if the frequency cutoff for a Gaussian filter is just very gradual? 2. Any additional preprocessing is recommended, besides temporal filter and what the minimal pre-processing has already enacted? 3. What is an intelligent way to combine correlation matrices? Averaging (Power et al, 2011)? Binarizing the correlation matrix by setting the top 10% of voxels to 1 and the rest to 0, and averaging the binarized matrices (Yeo et al, 2011)? Either? Something fancier? Any advice or additional resources would be enormously appreciated -- thanks very much!! Kim -- Kimberly Stachenfeld, BS Graduate Student 236A Princeton Neuroscience Institute Washington Road Princeton, NJ 08544 k...@princeton.edu ___ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail. --