Hello,

What would be the prospects of stitching adjacent, partially overlapping (say 10-20%) 3d image stacks with CMTK affine registration.

This situation often occurs with tiled confocal images. Most solutions that i am aware of only consider translations (3dof), but sometimes it is also necessary to consider small rotations (normally only around 1 axis) or even small scale differences.

What sort of initialisation would be required. How would data in the floating image that did not overlap the template be managed? And could data in the region of overlap be merged/averaged in any semi-intelligent way?

Many thanks,

Greg.

Hi Greg -

So this took some thinking obviously. Basically, what I am going to tell you is based on my observation of Stephan Preibisch's work a few years back on the SPIM mosaicing.

Regarding alignment - the intensity-driven approaches are really breaking down when overlap is low. The exact definition of "low" being elusive, I would still be willing to guarantee that 10-20% isn't going to be enough. Certainly not with any of the algorithms in CMTK. My recommendation would be to use feature descriptors, like Stephan did, for example SIFT in 2D and whatever generalization they have in 3D.

Regarding the image fusion post alignment - CMTK doesn't really have anything better than averaging. I was trying to find the code for the information-weighted fusion that Stephan and I worked on way back when, but I couldn't find any. I have a vague recollection we actually did that in ITK, not CMTK, to take advantage of the faster Gaussian convolution implementation.

Sorry if I am not very helpful here, but I am afraid this is the best I can offer.

TR