[Insight-users] [Rigid 3D Registration] : Tuning the optimizer and Transform choice ?

[Vincent Daanen]

dear itk-users,

First of all, let's explain the data : MRI volume, 256x256x108 voxels. 
The moving volume is the fixed volume transformed by an arbitrary transform.

When this arbitrary transform is a pure translation (-15mm to +15mm in 
each direction), registration (multiresolution) works quite well (good 
result and about 1,5 mintes) ;
but when it is complex (i.e. rotation (more or less Z dir, -5 to +5 
degrees converted in radian + translation), I ran into troubles !

First, the registration process is really long (multiresolution with 2 
level) : more than 1 hour.
So I want to tune the optimizer (I'm using VersorRigid3DTransform and 
its associated optimizer itk::VersorRigid3DTransformOptimizer).
Yesterday evening, I read the Itk software guide to find explanation 
about the optimizer but I did not find what I was searching for :
what are the use and how to set the MaximumStepLength & 
MinimumStepLength parameters ?
how to choose the values to have a quick and rather good transform a 
coarse level ?

Secondly, which is the best 3D Rigid transform of itk to use ? 
Rigid3DTransform, Euler3DTransform, QuaternionRigidTransform, 
VersorRigid3DTransform ? and why ?
oh ! btw, what is the difference between QuaternionRigidTransform and 
VersorRigid3DTransform ?

 Vince

-- 
Vincent Daanen, PhD
Research Engineer, Post-Doctoral Position
Laboratoire TIMC/IMAG (Univ. Joseph Fourier - CNRS UMR 5525)
Equipe GMCAO
Institut d'Ingénierie de l'Information de Santé (IN3S)
Faculté de Médecine - 38706 La Tronche cedex - France
Tel: +33 (0)4 56 52 00 54 - Fax: +33 (0)4 56 52 00 55
Vincent.Daanen at imag.fr

"Les problèmes ne peuvent être résolus par
ceux dont l'horizon se limite aux réalités 
quotidiennes, mais par ceux qui rêvent de 
choses qui n'ont jamais existées et qui se 
disent : Pourquoi Pas ?"

J-F Kennedy, 1963.