HI all,
We had submitted a paper based on cortical thickness analysis in Qdec. The reviewer's had couple concerns regarding the QDec analysis. We would appreciate your help to address/resolve those issues. Here are the concerns:
(1) What was the rationale of using 10mm smoothing? The typical fMRI values for volumteric work is 2-3 x the original voxel size. What's the reason for this value?
Our response so far: A smoothing kernel of 10 mm was chosen because this setting had been published recently in articles that performed similar analyses. 10 mm therefore was a typical parameter for this type of analysis. This setting was also the default in the QDEC GUI, also suggesting that it was a typical setting. We only performed analyses of the data for this project using this setting and did not iteratively test multiple kernel widths. Also, the FWHM for cortical thickness analyses is somewhat larger than that used in fMRI analyses because it is applied to surface, as opposed to volume-based data. Surface based data are not prone to the same artifacts of including white matter, csf or brain tissue on the other side of a sulcus as volume-based data are.
(2) What's the p-value to get in for a cluster-wise analysis and what size is a significant cluster? It's important to know these non cluster-wise values to get a sense of the parameters that determine significance.
Our response so far: QDEC does not have a parameter that can be adjusted for cluster size or extent as other whole brain analyses such as Statistical Parametric Mapping do. The researcher can only indicate a statistical cutoff. We used a p value cutoff of 0.05 after a correction for multiple comparisons that was based on the Monte Carlo permutation cluster analysis with 10,000 iterations. Additionally, our findings comprise of pretty large clusters. This obviates the reason for cluster-thresholding based on size.
Thanks in advance for the response. Binod
On 08/16/2013 05:02 PM, Binod Thapa-Chhetry wrote:
HI all,
We had submitted a paper based on cortical thickness analysis in Qdec. The reviewer's had couple concerns regarding the QDec analysis. We would appreciate your help to address/resolve those issues. Here are the concerns:
(1) What was the rationale of using 10mm smoothing? The typical fMRI values for volumteric work is 2-3 x the original voxel size. What's the reason for this value?
Our response so far: A smoothing kernel of 10 mm was chosen because this setting had been published recently in articles that performed similar analyses. 10 mm therefore was a typical parameter for this type of analysis. This setting was also the default in the QDEC GUI, also suggesting that it was a typical setting. We only performed analyses of the data for this project using this setting and did not iteratively test multiple kernel widths. Also, the FWHM for cortical thickness analyses is somewhat larger than that used in fMRI analyses because it is applied to surface, as opposed to volume-based data. Surface based data are not prone to the same artifacts of including white matter, csf or brain tissue on the other side of a sulcus as volume-based data are.
I think this is a good answer for a question that is not entirely fair. Aside from a few arguments about making GRF work, no one has ever had much of a justification for the particular choice of FWHM. It is just traditional at this point. You might point out that more smoothing is generally considered worse and that 10mm on the surface is much less than 10mm in the volume and people use 10mm in the volume is not unusual.
(2) What's the p-value to get in for a cluster-wise analysis and what size is a significant cluster? It's important to know these non cluster-wise values to get a sense of the parameters that determine significance.
Our response so far: QDEC does not have a parameter that can be adjusted for cluster size or extent as other whole brain analyses such as Statistical Parametric Mapping do. The researcher can only indicate a statistical cutoff. We used a p value cutoff of 0.05 after a correction for multiple comparisons that was based on the Monte Carlo permutation cluster analysis with 10,000 iterations. Additionally, our findings comprise of pretty large clusters. This obviates the reason for cluster-thresholding based on size.
I'm not sure what the reviewer is asking. Is he/she asking for the cluster-forming, ie, voxel-wise, threshold? Or the p-value threshold for the cluster that you used to report clusters? You should report both. Both can be set in the QDEC GUI. I think the defaults are .01 and .05 respectively. Given the clusterwise p-value, there is a corresponding critical cluster size. You can get this from a file in the QDEC output folder (in the contrast subfolder). The file will be called something like "cache.th20.abs.pdf.dat". The 2nd column gives you the cluster size (in mm2) and the 4th column gives you the clusterwise p-value. Find the size of the cluster that corresponds to .05. In one data set I looked at, the size was about 328mm^2
doug
Thanks in advance for the response. Binod
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Hi
Du A.T. Brain 2007 ”Different regional patterns of cortical thinning in AD" - used the smoothing level of 10-mm. This article can be added as reference.
Additionally "higher smoothing levels, for ex. at 20 mm, would bring partially inflated patterns due to excessive smoothing and due to failure of inference methods to control the proportion of false discoveries" - Bernal-Rusiel, Neuroimage 2010 (www.ncibi.nlm.nih.gov/pubmed/20362677) while lower smoothing levels have an increased potential of preserving clusters with low significance.
alex.
Le 19/8 12:58, Douglas N Greve a écrit :
On 08/16/2013 05:02 PM, Binod Thapa-Chhetry wrote:
HI all,
We had submitted a paper based on cortical thickness analysis in Qdec. The reviewer's had couple concerns regarding the QDec analysis. We would appreciate your help to address/resolve those issues. Here are the concerns:
(1) What was the rationale of using 10mm smoothing? The typical fMRI values for volumteric work is 2-3 x the original voxel size. What's the reason for this value?
Our response so far: A smoothing kernel of 10 mm was chosen because this setting had been published recently in articles that performed similar analyses. 10 mm therefore was a typical parameter for this type of analysis. This setting was also the default in the QDEC GUI, also suggesting that it was a typical setting. We only performed analyses of the data for this project using this setting and did not iteratively test multiple kernel widths. Also, the FWHM for cortical thickness analyses is somewhat larger than that used in fMRI analyses because it is applied to surface, as opposed to volume-based data. Surface based data are not prone to the same artifacts of including white matter, csf or brain tissue on the other side of a sulcus as volume-based data are.
I think this is a good answer for a question that is not entirely fair. Aside from a few arguments about making GRF work, no one has ever had much of a justification for the particular choice of FWHM. It is just traditional at this point. You might point out that more smoothing is generally considered worse and that 10mm on the surface is much less than 10mm in the volume and people use 10mm in the volume is not unusual.
(2) What's the p-value to get in for a cluster-wise analysis and what size is a significant cluster? It's important to know these non cluster-wise values to get a sense of the parameters that determine significance.
Our response so far: QDEC does not have a parameter that can be adjusted for cluster size or extent as other whole brain analyses such as Statistical Parametric Mapping do. The researcher can only indicate a statistical cutoff. We used a p value cutoff of 0.05 after a correction for multiple comparisons that was based on the Monte Carlo permutation cluster analysis with 10,000 iterations. Additionally, our findings comprise of pretty large clusters. This obviates the reason for cluster-thresholding based on size.
I'm not sure what the reviewer is asking. Is he/she asking for the cluster-forming, ie, voxel-wise, threshold? Or the p-value threshold for the cluster that you used to report clusters? You should report both. Both can be set in the QDEC GUI. I think the defaults are .01 and .05 respectively. Given the clusterwise p-value, there is a corresponding critical cluster size. You can get this from a file in the QDEC output folder (in the contrast subfolder). The file will be called something like "cache.th20.abs.pdf.dat". The 2nd column gives you the cluster size (in mm2) and the 4th column gives you the clusterwise p-value. Find the size of the cluster that corresponds to .05. In one data set I looked at, the size was about 328mm^2
doug
Thanks in advance for the response. Binod
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu