Dear SPMers,
We're looking at different non-linear registration methods and trying to optimise for speed, since it is the slowest step in our pipeline. Would anybody on the list be willing to offer recommendations on:
which algorithms are fastest, and
which tweaks might optimise speed?
In the literature, I've mostly found comparison papers looking at accuracy, not speed. There are a couple of examples of GPU acceleration but none that look like stable releases that I've found. We're looking into whether downsampling can work for our purposes (we're using CTA, which often has about 0.6 mm resolution; this is overkill for features like gyri but we might lose vessel info if we go too coarse).
Sorry for the rather vague query, but even a vague answer would reassure me I'm looking in the right direction.
Thanks!
Paul Wright
KCL
|