Overview

This component of the rumba tools address the following problem: given two different images defined on some set of voxels, how does one find an appropriate way to map the two intensities ? This is a non-trivial problem for several reasons:

• The subjects may be different. A given anatomical region in one subject will not generally be aligned with the same region in another subject, even if the image acquisition techniques are the same (or to put it more curdely, the two subjects may have differently shaped heads). Since differing subjects will not in general have heads that differ by a simple linear or affine transform, spatial warps consisting of polynomial or cosine functions are usually used for this.
• Even if the two data sets are taken from the same subject, during the same experiment, the subject's head may move during the experiment. The convention has been to use rigid body transformations to realign these images.
• It is desirable to present coordinates of significant regions in terms of a standard scale, where (for example), each given coordinate is known to correspond to a particular anatomical component of the brain (for example, a particular Brodmann area). In order to do this, one conventionally aligns to a template image. If different members of the research community align to the same template, the coordinates obtained in different studies are, in theory, comparable.

I'm currently investigating and implementing different algorithms for performing image registration.

Software Implementations

Programs

• align

  The align program performs the following algorithms:

  • rigid-body motion correction, using a least-squares model
  • warp-based inter-subject or subject-template registration, using a least squares cost function, and polynomial warp basis functions (degree 2-5)
  • affine normalisation registration using least squares.
  • affine registration for data sets of differing modality. This uses a cost function similar to that used in Roger Woods' AIR software.

• reslice

  The reslice program applies a spatial transform and resamples data accordingly. Choices of interpolation include trilinear, SINC, and various combinations of trilinear/SINC interpolation.

Status: stable, but still in need of refinement/improvement. There are unresolved corner cases that need to be addressed. For example, what should be done if one of two images is severely truncated, or if one image contains considerably more of the head than the brain ? This problem often causes image registration algorithms to fail.