users > RE: Registering subregion of brain to whole brain
Sep 28, 2015  05:09 AM | Torsten Rohlfing
RE: Registering subregion of brain to whole brain
Hi Kate -

In general, registration of images with significant cropping at notoriously hard to register nonrigidly, in particular when there are different extents of cropping across images.

Some general tricks, not all of which may be supported by the munger script that you are using:

1) The lower your degrees of freedom, the higher your chances of success. That is, a rigid transformation will more likely be robust to cropping more than an affine, followed by relatively coarse nonrigid (low number of control points, high relative constraint weights). The lowest likelihood of reasonable registration would be for a highly nonrigid registration (large number of control points, low or no constraints). As a consequence, you should use the lowest number of degrees of freedom and highest constraint weights that you can get away with.

2) The direction of the transformation matters - in my experience, registration is somewhat more robust using a cropped image as the "reference" or "fixed" image (that is usually the first one given to CMTK's registration tools) rather than the "floating" or "moving" image. Of course for nonrigid registration, you may require a specific direction of the registration transformation based on your application.

3) For linear (rigid or affine) registration, the "registrationx" tool has an option ("--symmetric" I believe) that will simultaneously compute the registration and its inverse, and optimize a symmetric similarity measure that is invariant under fixed/moving image switch. I have found this to be more robust when images have different fields of view or other degradations. No such function exists for nonrigid registration, unfortunately, at least in CMTK.

4) CMTK's registration tools all allow for a "padding value" to be defined. This is a value that you choose to represent non-existent data in your images (could be, say, -1 for images that are otherwise strictly positive-valued), which means you can have "cropping" of pixels inside your images' field of view. Not sure this is something that's helpful for you, but thought I'd mention it.

Hope some of this is useful for you.

Best,
  Torsten

Threaded View

TitleAuthorDate
Kate Tabor Sep 26, 2015
Tim Grocott Sep 26, 2015
Torsten Rohlfing Sep 28, 2015
Kate Tabor Sep 28, 2015
RE: Registering subregion of brain to whole brain
Torsten Rohlfing Sep 28, 2015