Best registration tool for 3D electron microscopy images?

We recently started 3D electron microscopy (FIB-SEM) image analysis.
We know that there are several tools in ImageJ such as elastic, bUnwarpJ, Linear Stack Alignment with SIFT and so on and also there are many other algorithm of deep learning techniques such as ssEMnet.
However, we could not find the good comparison of these methods in terms of usability.
And we could not find the good tips to apply these methods to our datasets.
Are there any discussion about the issue above?
We’d appreciate someone could give us an advise what we should try.
Thank you!

For example, one of our data is like this
Resolution x-y 5 nm/px, z 10 nm/px,
Dimension x-y-z 3000x1000x300 px

@hiroalchem I hope this reply finds you well, as your question was asked 3 weeks ago.

Generally, alignment accuracy is determined through a similarity metric like SSIM. These metrics provide a number for the otherwise subjective art of judging image alignment accuracy. For FIB-SEM in particular, since the slices are not uniform thickness (in reality), and there are many subtle differences in pixel size between slices due to foreshortening, defocus, etc., getting a ‘true’ alignment for FIB-SEM data is very difficult to determine empirically. Serious investigations comparing techniques will work with simulated data where the true alignment is known, but this may bear little relevance to your application, especially if it is a biological specimen.

All of that out of the way, the general approach is to run a rigid alignment followed by non-rigid (elastic) alignment. The landmarks used for these tools mentioned are the primary difference between them. Some use the raw grayscale while others use feature classifiers (e.g., SIFT) and are thus less prone to weird drifting as you move through a stack of images. The SIFT patent ran out recently, so it is probably going to be seeing more use moving forward in image processing software. Occasionally, large jumps can occur during automated alignment and this will require manual correction. In FIB-SEM, this can be due to severe curtaining, defocus, or bad image alignment due to sample drift during acquisition. In each case, you generally will want to restrict the rotation, as FIB-SEM data are not expected to rotate through acquisition, unlike ssEM.

Others would have to comment on the how-to of these techniques, but I hope that provides a general overview and a way to get started. I support Avizo software, so if you are interested in trying the tools available in Avizo, you could reply to me here or request a trial and mention this forum post: We are set to release some better 3DEM alignment methods by the end of the year for non-rigid alignment. It works very well for ssEM/array tomography, but also works nicely for FIB-SEM.