[ITK-users] What is an accepted duration of a 3D registration? Should I expect the MSE to "converge" to 0?

[Siavash Khallaghi]

Hello Marina,

I have some registration experience and have been using ITK since 2008,
maybe I can help you:


Marina Bendersky wrote
> Should I expect the MSE metric to converge to 0, or something closer
> to 0 (instead of the current values, in the ninety thousands...)?

Maybe and maybe not, it depends on the fixed and moving images and the
nature of the spatial transformation between the two images. The MSE at the
true solution might or might not be zero depending on:

1. If the intensities differ at the true solution. For example, if there is
a shift in intensities between fixed and moving images, this would result in
a non-zero MSE at the true solution.

2. If local deformations exist between fixed and moving images, but a
rigid/affine transformation is used in the registration pipeline.

With this said, you can check your data and perform some form of
pre-processing to correct them.


Marina Bendersky wrote
> This 3D registration runs on a MacBookPro laptop for 5.5 hours, so I
> was wondering, is this "normal"? How long should a 3D registration
> take?

There are a number of reasons why a registration would take a long time. One
prominent reason is poor initialization. If the optimizer starts fara way
from the solution, then it would take a lot of iterations (if it manages to
avoid local minima) to converage. A "useful" registration for
prototyping/research is one that converges between 1-10 minutes so that you
can tune its parameters while maintaining your sanity.

HTH,

Siavash



--
View this message in context: http://itk-insight-users.2283740.n2.nabble.com/ITK-users-What-is-an-accepted-duration-of-a-3D-registration-Should-I-expect-the-MSE-to-converge-to-0-tp7587806p7587807.html
Sent from the ITK Insight Users mailing list archive at Nabble.com.