Hi,

sub81737 from the Munchen data set appears to be left-right flipped with respect to the anatomical.

There are other peculiarities with the Munchen data set. For instance, looking at sub96591 it's hard to see that the mprage and epi images are from the same person. The mprage is very rectangular and the epi is oval-shaped. They don't align well at all. There seem to be similar problems with some of the other subjects as well.

cheers,

Brad Buchsbaum

Hi Brad,

Thanks for your comments. We've reviewed the two dataset you are referring to.

For Sub96591, we reviewed the registered EPI images (EPI to MPRAGE using 6-DOF FLIRT registration in FSL). When registered, the fit is quite good. In the original / native space images the EPI is rotated compared to the MPRAGE (likely reflecting that the functional slices were obtained at an angle). Were you looking at the original images or registered ones?

For sub81737. Looking at the registered images (EPI to MPRAGE) we did not see clear evidence for a possible left-right flip of the EPI compared to the MPRAGE, though we have contacted the site to request they verify the orientation.

Hope this helps, and thanks again for your thoughts!
Maarten

Regarding sub96591. When I run flirt with 6 dof I get poor registration (see attached). The mprage is wider by about 5-8 mm on either side. I guess it could be severe distortion but the epi looks pretty good.


Regarding flipping. This is hard to see, but examine sub64202. There are some very deep sulci in the right inferior frontal area, which makes it easier to see the alignment.

First register the original data sets:

align_epi_anat.py -epi mean_rest.nii -anat mprage_skullstripped.nii.gz -epi_base 0 -giant_move -anat_has_skull no -cost lpc -suffix flipped

Then flip the image:

3dresample -orient LPI -prefix flipped_mean.rest.nii -inset mean_rest.nii
3drefit -orient RPI flipped_mean.rest.nii

Then rerun:

align_epi_anat.py -epi flipped_mean_rest.nii -anat mprage_skullstripped.nii.gz -epi_base 0 -giant_move -anat_has_skull no -cost lpc -suffix flipped

Then load the registered mprage in to AFNI and overlay the flipped epi. Look at slice 72-75 on the aligned mprage and look for the very deep sulcus in the right frontal cortex. You should see the contour matches on the epi overlay. It helps to rapidly toggle the visibility of the epi overlay.

Note that flirt does not do well with these images as it tends to push the epi up in the z plane 5 to 10mm. This makes it hard to see the left-right flip.


cheers,

Brad Buchsbaum

Hi Brad,

Thanks for your continued effort on this one!

Regarding sub96591. We backtraced our steps and realized we used a 7 DOF linear EPI to skull-stripped MPRAGE for the images in the paper. However, the 7 vs. 6 DOF is not the full solution. If you look in our preprocessing scripts, you can see that we get the example func image that is used for later registration after skull-stripping the EPI. If you use the skull-stripped EPI in combination with the 7 DOF FLIRT registration the alignment appears to be good. I attached an image comparing 6 DOF to 7 DOF and 7 DOF on the skull-stripped EPI for this subject. The differences in results are pretty clear.

As to why 7 DOF appears to work better, one possibility is that some degree of scaling may have occurred for the EPI on the scanner. As indicated before, we are not sure whether the EPI and MPRAGE for these subjects were collected on the same day.

Regarding the left-right issue we are continuing to look into this issue with the contributors.

Hope this helps,
Thanks!
Maarten

Hi Maarten,

That extra degree of freedom can hide a lot. Try a 7 dof registration between two different brains and you'll see decent overlap.

I'll try and convince you with one more image. I ran flirt with 7 dof on a skull-stripped epi for sub96591.

Three images are attached.

1. epi aligned to mprage (slice 146)
2. mprage (slice 146)
3. overlay with 50% transparency.

I've put the cross-hair at the edge of the posterior extent of the left lateral ventricle. The cross-hair is displayed for the same coordinate on the epi. Look at the shape of the ventricles and the position of the cross-hair. The overlay looks OK because the extra dof corrects for overall differences in brain size and its hard to see the differences in detail (e.g. shape of ventricles). My sense is that these are two different brains.

anyway, thanks for looking in to this!

cheers,

Brad Buchsbaum

Regarding the flipping of 64202 and 81737, I think Brad may be right. I am using basically the same methodology as him, so that might be expected. For visualization, it is helpful to look at edges within the images. Attached is an image of 64202 as received and left-right flipped antomical dataset and both aligned to the EPI data, as Brad suggested. For AFNI users, the edges can be seen in the underlay (background anatomical) dataset by using the Display control panel, edge detect option and even better with the @AddEdge script or the -AddEdge option to align_epi_anat.py.

Sometimes the left-right flipping is more obvious, as with major asymmetric differences in ventricle sizes. Here it's much more subtle (it's really kind of amazing Brad saw this at all). By looking at some of the more superior slices, it becomes more obvious the sulcal edges do not align well while the flipped volume does align well at those edges.

The real engine of the alignment, 3dAllineate, can show the values of the costs for all the cost functionals available. Here the values for the unflipped and flipped data are shown side by side for each metric:

non-flipped flipped
ls = 0.782876 0.805848
sp = 0.694957 0.71488
mi = -0.179521 -0.205214
crM = 0.389297 0.371718
nmi = 0.974591 0.970862
je = 6.8855 6.83786
hel = -0.0326018 -0.0367718
crA = 0.635821 0.623007
crU = 0.513464 0.494868
lss = 0.217124 0.194152
lpc = 0.290404 0.16722
lpa = 0.709596 0.83278
lpc+ = 0.805524 0.673313
ncd = 1 1

Each line shows the abbreviation for the cost functionals. To align the data, we used the lpc, the local Pearson correlation. To compute an alignment the 12-parameter affine transformation is optimized to minimize the cost. The lpc cost for the flipped data shows a lower cost, but so did many of the other well known metrics such as the normalized mutual information or the correlation ratio cost functionals. Note the costs are not comparable across cost functionals, so one couldn't compare the lpc with nmi directly. Similar results were seen for both the datasets Brad mentioned.

Hi Brad,

Yep, we see your point - that is not a great alignment. We are still waiting to hear back from the contributors to get a final answer about the correctness of that subject's data, as well as the possible left-right flip.

Thanks,
Maarten

In response to the feedback provided by 1000 Functional Connectomes Project dataset users (see open discussion forums for details), the data for the following sites are currently under review to ensure proper right/left orientation: Dallas, Leipzig, Munchen. We are working with each site to trace the data path from magnet acquisition to data transfer to the organization and uploading of datasets to the NITRC website, to verify right/left orientation. We should have updates on these sites over the course of the next 7-10 days.

We are very grateful to our users for their various inputs on the datasets. A key goal of the project is the development and promotion of quality assurance procedures capable of detecting errors that can arise in data handling and transfer processes. We strongly encourage continued input from users to ensure the maximal accuracy of the 1000 Functional Connectomes Project database. Further, we encourage users to continue to provide input in the forums regarding recommended procedures for accomplishing this.

Hi Brad,

we managed to trace the L/R MPRAGE - EPI discrepancy of the Munchen dataset to a L/R flip of the EPI somewhere in the process of converting from SPM analyze to nifti. The flip was obscured by the fact that SPM seems to correct for the flip when displaying the EPI.

We uploaded the corrected dataset.

Thanks again for bringing this to our attention!
Best,
Maarten