We quite often use the surface based registration option in Amira to generate an initial alignment for cases where CMTK's affine registration fails.

  - Take your starter image and smooth/downsample to e.g. 3 um voxels. 
  - Then generate a surface – you can use the Isosurface viewer module
  - We then normally simplify to ~ 18k vertices. Now you are ready for the surface aligment
  - Choose Align surfaces
    - set the **template** brain as the surface to be transformed
    - set the **sample/floating** brain as the reference surface
    - (yes this sounds backwards, but that's what CMTK wants)
  - Now align the centres
  - rigidly align the brains until you see no further improvement
  - then continue with affine registration until ditto
  - In the console, get the reference brain transformation e.g. ''JFRC2.surf getTransform''
  - These 16 numbers define a homogeneous affine transformation, which can now be used to make a CMTK registration.

We carry out the last step in R:
<code>
library(nat)
# paste in the 16 numbers
m=matrix(scan(), ncol=4)
# check that you have a 4 x 4 matrix with last row 0 0 0 1
write.cmtkreg(cmtk.mat2dof(m), "/path/to/my/reg/folder/Registration/affine/JFRC2_FLPO_20121214175357476_2097_01_ght.list")
</code>

You will need to have R and the nat package installed. See [[http://jefferis.github.io/nat]] for details.