Projects:DBP2:Harvard:Registration Documentation:Compare

From NAMIC Wiki
Jump to: navigation, search
Home < Projects:DBP2:Harvard:Registration Documentation:Compare

This page aims to create a concise comparison of registration methods to show the difference in quality of registration between FSL 4.1 and Slicer 3.3. All registrations were done with case 01031 T2w as the target and case 01053 T2w as the source image. We've used only masked images for T2w to DTI baseline because anything else is expected to fail.

Other types of registrations and a link to the files used can be found here.

Note: for FSL all files need to be in nifti format. For me, FSL was interpreting my masked baseline image strangely so I had to flip the nrrd before converting to nifti using unu flip. If you encounter problems while trying this yourself, view your nifti files in FSL before registration to see if they are oriented properly.



T2 to Baseline registration

In this registration we take a masked T2 image as our source, and register it to the corresponding masked baseline image target using FSL's FLIRT and FNIRT tools, and in Slicer 3 by the 'Registration -> Rigid Registration' and 'Registration -> Deformable BSpline Registration' modules.

Rigid

Rigid registration comparison: Very similar, only small differences viewable. Both are properly aligned.
















Rigid registration is very similar between the two tools. Differences are only visible when switching back and forth in slicer. The Slicer registration seems slightly more diffuse, but both seem to be properly aligned for further bspline registration.

BSpline

BSpline warp comparison: FSL is quite good, Slicer is OK, but much worse. Both have slight troubles in lower frontal and temporal, and some issues at back of brain.

















Bspline registration shows some differences between the two tools. FSL produces a good registration, only really having trouble in the bottom few slices of the brain when looking at the temporal lobe, and at the very front of the brain, though it does match much of the DTI distortion. Slicer does not fare as well, working only in the middle portion of the brain. It does not match the DTI distortion as well, leaving more of the frontal lobe and back of the brain mismatched. Also, FSL produces a deformation volume that can be used to transform a label map, while Slicer's output transform does not seem to accurately match the volume created when directly outputting a volume, meaning label maps cannot be accurately transformed.


Case to case registration

Rigid

Rigid registration comparison: Mostly good, slight differences in tilt, with masks giving better results than non-masks.












All a rigid registration will manage is aligning the two brains in space, it won't account for any anatomic differences. Slicer's non-masked rigid registration is close to in line, though with the front slightly rightward and the top slightly leftward of the target. FSL's non-masked rigid registration is close to in line, though with the front slightly downward and leftward of the target. Slicer's masked rigid registration is very close. There's a slight tip rightward of the top. FSL's masked rigid registration seems right on.

Affine

Rigid-Affine registration comparison: Non-masked are sized by skull and therefore a little off. Non-masked are done better, mostly well alighed with FSL performing better than slicer.












Here we've taken the Rigid images and used them as the source, so that the images you see above have been run through both rigid and affine registration. Here there should be some better sizing, as an affine transform will allow more leeway than a rigid. Don't expect exact matches at this stage. Slicer's masked affine registration is, oddly, more turned than the rigid, with the front tilting definitively left and the top slightly right. The ventricles match up reasonably well here, however. FSL's masked affine registration is for the most part well matched on length, width and orientation. Slicer's non-masked affine registration is well aligned though appears high and posterior to where I would place it by eye. This seems due to the matching of height and width by the skull instead of the brain. FSL's non-masked affine registration is similarly well aligned but seemingly misplaces, with the front appearing better matched but the posterior worse.

Bspline

Rigid-affine-Bspline registration comparison: Pretty poor for the most part. Outer boundaries are matched, but little if anything inside is in the same place. Perhaps this is expecte dfor case to case registration.












Here we've taken the results of the Rigid-Affine registration and done a bspline registration on them, meaning images have been through three steps. Here is where good registrations should be found, if ever. Slicer's masked bspline registration seems ok, though there's maybe a leftward bow to the midsagittal plane, and the back of the brain seems much larger than the target. FSL's masked bspline registration seems too large all around, though very well matched on ventricles. Slicer's non-masked bspline registration looks perfectly matched on the skull, but horrible on the brain. FSL's non-masked bspline registration looks well matched on size and skull, but very few gyri are aligned. Even the sylvian isn't really matched up.