July31T-con/Synopsis

Each of the algorithms has an optimal voxel dimension and they are not the same. We must decide what format of the data will be used as the common starting point.

I have three goals here:

1. to summarize the new information gained during the past 9 days since we had our T-con
   
2. to capture the essence of the discussions that took place in pursuit of that information and
   
3. to propose a final solution regarding the format of the starting data to be processed by all the groups for the Santa Fe meeting that will be a reasonable compromise between the in principle optimal but completely impossible solution of all groups using the "same data" and the "every group for themselves" solution that will require more work in Sante Fe to sort out our results.

1. New Information

• The wiki has been edited to correct the image acquisition details. The data were collected on a GE scanner using parallel imaging. The prescribed slice resolution of x is not what is actually in the raw image file as reconstructed by GE. There is some proprietary preprocessing done as or after the data are taken out of K-space and before the non-linear combination of the parallel images. The data available on the scanner has a resolution of 0.9x0.9x1.7mm (no gap), FOV 240 x 240, 256 x 256. That data is the closest to "raw" that is available and is what is currently posted as dwi. The ROIs were generated at this resolution.
• There are three camps within our small group, those whose algorithms can work with this data "as is", those who need upsampled isotropic voxels and those who need downsampled isotropic voxels.
• The dwi-EdCor files are the "raw" data that have been corrected for eddy currents (using FSL)
• Marek provided the requested information required for clinical relevance of this project WRT the choices of tracts and potential hypotheses to be tested. I have edited them into the ROI page and our Project Page.

2. Essence of Email discussion

"On Jul 31, 2007, at 4:57 PM, Marc Niethammer wrote:

Hi Guido,

it is not completely clear which particular algorithm is implemented on the scanner for the MR signal reconstruction. There is certainly some form of scanner-internal up-sampling, however, to reverse-engineer this may not be trivial.

To make sure that the comparison between the methods is fair, I agree that it is desirable that all of them run with data that is, as far as possible, identical. This will only be possible to a certain extent, since some algorithms work on the diffusion weighted images directly, others use the tensors as their input, but most (all?) of the algorithms perform some form of algorithm-internal data interpolation anyway. (This is easiest to see for the classic streamlining algorithm that needs to compute values with subpixel resolution.)

Here is a possible suggestion of additional datasets we could provide. Let me know what you think.

(1) Downsampled datasets (original and eddy current corrected) to 1.7mm isotropic voxels. The downsampling would be performed on the diffusion weighted images using Gordon's teem library (using a Catmull-Rom interpolation kernel). This includes downsampled labelmaps using nearest neighbor interpolation.

(2) Estimated diffusion tensors (using simple linear estimation) for the datasets in their original resolution as well as for the downsampled datasets. This should be the input data for all methods relying on diffusion tensor input.

Best,

Marc"

"On Aug 2, 2007, at 1:27 PM, Marek Kubicki wrote:

Hi Guys, I looked at the action items from the t-con, and I have some thoughts/comments/questions for all of you, regarding my part: 1. I edited some information regarding the data- FOV is 240 mm square, not 140. We have eddy current corrected, not distortion corrected images. We do not have distortion correction tools. For the eddy current distortion correction, we took all the gradient images, and registered them together. I think its linear registration, nothing fancy. 2. I will try to get the information from GE, about how they up-sample the images, and if there is anything that can be done to reverse the process. 3. Regarding the clinical relevance, this is tough. We picked tracts that connect frontal and temporal lobes, and all of them have been shown, at some point, to be abnormal (FA lower) in schizophrenia. I will put a slide that illustrates their exact function, and why we think they are affected in schizophrenia, if this is what you are looking for, but there is not much that has been done in terms of specificity of the white matter abnormalities. So it might be tough to directly answer to your question- what we expect to find in schizophrenia on this dataset. We have very small datasets in this project, so I would not expect to find statistical group differences. Because of the huge symptom variability within the schizophrenia spectrum, we usually need at least 20 cases, plus the same number of controls, to show the difference. I guess we could look at the effect sizes, and see how many subjects would we want for each method and each tract to show group differences? We could also look at the measurement variability within the control group? Or how combination of measurements increases the sensitivity of the DTI measurements? Your thoughts? Marek"

From Guido on August 6: "Randy and all,

I did not update the Wiki, and this is indeed a very interesting discussion. Although C-F proposes that we should not force anyone to accept a specific data interpolation, I fear that any comparison a the forthcoming meeting becomes very difficult. I understand the concerns of C-F and Marc that cutting the high frequencies indeed seems to cut some information. Would some labs us upinterpolated, no-interpolated, down-interpolated data with different interpolation schemes, everyone would work with different input data which by nature already would be seen as systematic differences. The telephone conference clearly showed that some groups really can't use the raw data but would down-sample them to an isotropic grid. Forcing all the participating groups to use this non-isotropic data might "cut" some labs out of this comparison. Maybe we really have to make a compromise just for this comparison to provide a downinterpolated, close-to-isotropic dataset for everyone and leave the more sophisticated analysis of optimal "redoing of the GE upsampling" or the use of the raw GE data for additional research to be discussed at the meeting?

Randy and all, could we propose the following: a) we decide about using the non-EPI corrected data for the time being (given the short time left for the meeting, not because EPI correction might not be useful but because we first need to quantitatively show its advantage)

b) we provide downsampled images (sampling the 256x256 back to a 144x144 grid with the standard procedure as used by Marc and provide this close to isotropic dataset to everyone. This will ensure that every group has access to a standard, close to isotropic dataset (1,67x1.67x1.7).

c) leave it to every group to use the original matrix, more sophisticated down- or upsampling, and to compare - and discuss implications/differences as part of the report.

Please understand that I don't want to stop the current, exciting discussion about the optimal reformatting of the data or redoing the GE upsampling, but I would like to see this as a component of the workshop itself. I think that all these questions is a very important part of the NAMIC DTI toolkit and recommendations for users, and of course providing the tools to perform a correct/optimal data preprocessing."

Later that day Sylvain expressed the following concerns: 1) All ROIs were drawn on the original data and they will have to be downsampled too. 2) "WRT to Guido's point > c) leave it to every group to use the original matrix, more > sophisticated down- or upsampling, and to compare - and discuss > implications/differences as part of the report. I am somewhat less convinced about this. I say if we want to be consistent, then only one single dataset should be provided so that all methods start with the same data.

Unless there is *strong* opposition against this, the PNL will prepare for each case: 1- A downsampled 1.67x1.67x1.7mm data set of DWIs in NRRD format. 2- Its corresponding plain vanilla linear least square tensor fit in NRRD format. 3- The associated ROIs downsampled to same resolution in NRRD format. All other data sets should be removed from this project (original and epi corrected).

For consistency, if the candidate tracking methods uses DTI as input then the tensor data *must* be used. Differences in tensor estimation techniques should not justify poor tracking, or at least should be observable in all the tracking techniques. Similarly, If the candidate tracking methods works directly on the DWIs then the DWI data *must* be used.

There is no perfect solution to finding the ideal data set. This is the lowest common denominator to all techniques within the NAMIC community and beyond. Given the large variability in processing techniques the least we should do is use the exact same input.

Best, -Sylvain

3. Proposed Solution

Return to July31T-con Page