[Insight-users] DeformableRegistration7 and stopping criteria

[Kris Thielemans]

Dear all

I'm trying to run DeformableRegistration7 on some of my data. This uses the
LBFGSBOptimizer for a BSplineDeformableTransform using a
MeanSquaresImageToImageMetric.

I have a problem with the stopping criteria used in the file. I think this
is related to the usage of the MeanSquaresImageToImageMetric which seems to
have values which are very much image dependent. I think it's the
appropriate metric for my problem (registering multiple gates of the same
patient on the same modality) but am not sure how to use it. More detail
follows.


The example program sets

  optimizer->SetCostFunctionConvergenceFactor( 1e+7 ); // KT original 1e+12
which according to the LBFGSBOptimizer  doc is low accuracy
  optimizer->SetProjectedGradientTolerance( .001 );// KT original 1.0
  optimizer->SetMaximumNumberOfIterations( 10 ); // KT original 500
  optimizer->SetMaximumNumberOfEvaluations( 500 );
  optimizer->SetMaximumNumberOfCorrections( 5 );

I find that I have to tune these variables to the particular data-set I'm
using. For example, if I give as input images with 26 slices of the whole 3D
images, the above parameters 'work'. That is, the program runs for 10
iterations and outputs a very reasonable match. Now I use the whole data-set
(211 slices) and the program stops after the first iteration (with final
deformation parameters equal to the identity), presumably because one of the
above criteria has already been met. (I had the same with the 26 slice data
if I used the original settings).

I'm playing around with SetCostFunctionConvergenceFactor (setting 1e+5) and
decreasing the SetProjectedGradientTolerance to smaller numbers, but still
have the same problem.

(For your info, right now, I'm using NCAT simulated data, which consist
essentially of a bunch uniform regions. This might mean that the metric is
not so suitable after all. Later on I'll go to real data).

No doubt, there will be some values for which the program will actually do
something. Question is which ones... It seems that to make this slightly
more robust, these settings will have to be computed somehow from the
images. Or maybe I should be using some normalised metric.

Any suggestions?

Many thanks

Kris Thielemans
Hammersmith Imanet Ltd
part of GE Healthcare
London, UK