Child pages
  • Image registration and ANTs tools

Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

MINC registration programs such as minctracc typically refer to 'source' and 'target' rather than 'fixed' and 'moving' images, e.g., minctracc ${options} -source $source -target $target -o source_to_target.xfm followed by mincresample source -like target -transform source_to_target.xfm -o source_to_target_resampled.mnc.  From this you might guess the (incorrect) correspondence source ~ moving, target ~ fixed, mincresample ~ antsApplyTransform (error).  However, mincresample and antsApplyTransforms give different results, which is resolved by inspecting mincresample source code to reveal that without the -invert flag the libminc function invert_general_transform is used, while mincresample -invert calls general_transform.  Thus, applying a transform such as done by antsApplyTransforms is actually achieved by mincresample -invert, and for this to make sense we must have source ~ fixed, target ~ moving.  (Also, we should prefer to use -invert when possible to avoid numerically inverting a transform!)

The resulting oddity to keep in mind is that in the MINC world, individual subjects are typically 'source' and templates 'target', while in the ANTs world, subjects are typically 'moving' and templates 'fixed', so the 'direction' in which a registration algorithm has been applied will likely be different (assuming one is interested in warping a subject to a template in both cases) between the two worlds, resulting in different transformations (since most registration algorithms are not symmetric or invertible and even ANTs's SyN transformations are only so up to various numerical approximations.).  It's the extra inversion hidden in mincresample which reconciles everything, of course.

Another issue to keep in mind is that often one wants to initialize the transformation of the moving image to some with some existing warping, say from a previous stage of registration.  (This is not equivalent to resampling the moving image first since the any cost function applied to the transform won't consider this pre-applied part, and also introduces some extra numerical error.)  ANTs tools allow --initial-moving-transform as well as --initial-fixed-transform, but `minctracc` allows only `-transform` which is applied to the source, i.e. fixed, image.  One way to get around this might be to use minctracc's -invert option to recover a transform going in the opposite direction, effectively identifying source with moving and target with fixed, but it should be checked whether this actually inverts the transform or just swaps the images in the registration process.

Jacobian determinants

Recall that given a differentiable function f : D → Rn defined on some n-dimensional region D, we can compute the Jacobian determinant function detJ(f) : D → R.  (Mathematically, this is the composition of the derivative operator to produce a linear approximation of the function - which can be represented as a matrix of partial derivatives - at each point and the usual determinant, which gives the ratio by which a linear operator changes volumes.)  The Jacobian determinant is therefore itself an image with domain D.  The Jacobian determinant of f at a point p is interpreted as a "local volume change" at p: intuitively, it's the ratio by which f changes the volume of an infinitesimal cube near p, where a negative value means the orientation (aka handedness or chirality) of the image has been reversed (a 'left-right flip').  In practice, this is computed numerically by finite difference methods.

...