Email exchange between Arno Klein and Andrew Worth of Neuromorphometrics (7/8-10/2013):
Arno:
the subcortex-labeled oasis-trt-20 brains have much bigger dimensions than the original oasis brains (e.g., 256x256x287 vs. 256x256x160) and are oriented differently. can you recommend a way to put the labels in the original space of the oasis scans?
Andy:
We don't specifically pad. We do it the same way as in Cardviews, but I had to make the code more robust to handle additional cases where the original scan was acquired differently than what the CMA was used to. Here's what the code says we do (it's been a long while since I looked at this):
The following is based on three 3D landmarks chosen by the operator; Anterior Commisure (AC), Posterior Commisure (PC) and "Mid-Sagittal" Point (MS).
First, find oms, the closest point on the ac-pc line to the ms point.
Then, nvec is calculated as the cross product between vec (ms-oms) and pvec (ac-oms), i.e. "up" crossed with "front".
To re-orient the scan, we first rotate nvec about the Z axis to make it point towards the +X direction (the subject's left). This is the Z rotation. Then we rotate NVec about Y so it points exactly in the +X direction to get the Y rotation. Then we Rotate vec towards into the -Z direction, to get the X rotation.
Arno:
oh, no. they've been rotated as well as translated, so what i've been trying won't work. i guess then i will have to resort to registration.
Andy:
Yes, they are rotated, translated, AND interpolated at a different resolution. I don't like the way this messes with the original data, but it was an attempt to match what was already present from the CMA when I originally wrote NVM.
I think I figured out where the padding comes from. The old slice resolution, which is normally less than the in-plane resolution, is used to calculate the number of slices for the new scan. For the OASIS scans, columns become slices and this changes the resolution when it is re-sliced after the rotation. However, the OASIS scans are 1x1x1mm, so this makes it less of a problem. However however, in NVM, we always blow up the scan to be 512x512 in-plane (the better to see the anatomy, the CMA always said). So this does have the effect of changing the resolutions and adding the padding.
For the usual CMA scans, this was not really a problem because the scans were acquired coronally so the rotations and translations were small.
Email exchange between Arno Klein and Andrew Worth of Neuromorphometrics (7/8-10/2013):
Arno: the subcortex-labeled oasis-trt-20 brains have much bigger dimensions than the original oasis brains (e.g., 256x256x287 vs. 256x256x160) and are oriented differently. can you recommend a way to put the labels in the original space of the oasis scans?
Andy: We don't specifically pad. We do it the same way as in Cardviews, but I had to make the code more robust to handle additional cases where the original scan was acquired differently than what the CMA was used to. Here's what the code says we do (it's been a long while since I looked at this):
The following is based on three 3D landmarks chosen by the operator; Anterior Commisure (AC), Posterior Commisure (PC) and "Mid-Sagittal" Point (MS).
First, find oms, the closest point on the ac-pc line to the ms point. Then, nvec is calculated as the cross product between vec (ms-oms) and pvec (ac-oms), i.e. "up" crossed with "front".
To re-orient the scan, we first rotate nvec about the Z axis to make it point towards the +X direction (the subject's left). This is the Z rotation. Then we rotate NVec about Y so it points exactly in the +X direction to get the Y rotation. Then we Rotate vec towards into the -Z direction, to get the X rotation.
Arno: oh, no. they've been rotated as well as translated, so what i've been trying won't work. i guess then i will have to resort to registration.
Andy: Yes, they are rotated, translated, AND interpolated at a different resolution. I don't like the way this messes with the original data, but it was an attempt to match what was already present from the CMA when I originally wrote NVM.
I think I figured out where the padding comes from. The old slice resolution, which is normally less than the in-plane resolution, is used to calculate the number of slices for the new scan. For the OASIS scans, columns become slices and this changes the resolution when it is re-sliced after the rotation. However, the OASIS scans are 1x1x1mm, so this makes it less of a problem. However however, in NVM, we always blow up the scan to be 512x512 in-plane (the better to see the anatomy, the CMA always said). So this does have the effect of changing the resolutions and adding the padding.
For the usual CMA scans, this was not really a problem because the scans were acquired coronally so the rotations and translations were small.