Hi,

I'm having problems applying the transform matrix from dof2mat to traced neurons.
A little background for my problem:  we're trying to trace neurons from many different insect brains in different alignments.  Since brain registration usually involves a decrease in image resolution, we'd like to trace the neurons in the original images, perform the registration to a template brain, and then apply the transformation matrix from said registration to the traced neuron.

The dof2mat seems to be the correct tool, however when I apply the matrix to my swc file (in Matlab), the traced neuron is significantly off from transformed image.  The misalignment is beyond just a simple translation, and appears to consist of both rotation and shear errors.

Are there any known issues that would cause the error with this workflow?  Should I instead be trying to sequentially apply the translation/rotation/scale/shear matrices from the xform file?  If so, what are the units on these sub matrices (radians on the rotation matrix, microns on the translation, etc?)

Thanks!

Dear Zane,
Originally posted by Zane Aldworth:
This is generally the best thing to do, but can be a little bit more fiddly.
I take it you are only trying to apply affine transformations?
It is most likely that you have not interpreted the matrix correctly (or at least in the same way as CMTK!). The two most common problems are:

1. a difference in the convention for row-major and column-major representation of the matrix. Essentially you may need to transpose.
2. a difference in your expectations about the transformation direction. You may actually want the inverse of what you have tried to apply.

I have regularly encountered these issues when working with different transform tools / representations. I actually use an R-based toolkit (https://jefferis.github.io/nat/) to look after the details. The minimal approach would look like this:


If you need to apply different transforms for each neuron then a little more effort is required to match neurons to transform. If you are interested then we would be happy to support your use of this toolkit over at https://groups.google.com/forum/#!forum/nat-user. 


No.
No. Best to use a single matrix multiplication of the homogenous affine transformation matrix.

Best wishes,

Greg Jefferis.

Thanks Greg - I think you pretty much covered the possible sources of errors :)

Just to add a bit of theoretical background why often one would want the inverse transformation:

Say you align image A to image B via transform X such that you can reslice B to match A.

Then if your neuron model is in the coordinates of image B, in order to get it transformed into the space of A, you need to apply the inverse of X.

That is because for reslicing, one takes the coordinate of each pixel in A, maps it to B via X, then pulls back the interpolated image intensity from B. So intensities move the opposite direction between coordinate systems as do geometries (e.g., pixel or model coordinates).

Best,
  Torsten

Originally posted by Greg Jefferis:

Thanks, Torsten.

Originally posted by Zane Aldworth:
Glad that was quickly sorted out. Queries welcome at https://groups.google.com/forum/#!forum/nat-user

Best, Greg.