[Insight-users] ITKBsplineSyN learning rate

[Luis Ibanez]

Hi Ema


On Mon, Jan 20, 2014 at 11:17 AM, Emma Saunders
<emmasaunders123 at gmail.com>wrote:

>
> Can anyone clarify the method to set the learning rate for ITKBsplineSYN
> registration
>


Looking at lines 417-424 of itkSyNImageRegistrationMethod.hxx
might be helpful.

https://github.com/InsightSoftwareConsortium/ITK/blob/master/Modules/Registration/RegistrationMethodsv4/include/itkSyNImageRegistrationMethod.hxx#L417


They are:

417   RealType scale = this->m_LearningRate / maxNorm;
418
419   typedef Image<RealType, ImageDimension> RealImageType;
420
421   typedef MultiplyImageFilter<DisplacementFieldType, RealImageType,
DisplacementFieldType> MultiplierType;
422   typename MultiplierType::Pointer multiplier = MultiplierType::New();
423   multiplier->SetInput( updateField );
424   multiplier->SetConstant( scale );

That is, in here the displacement field is normalized, to 1.0 vector length,
and scaled to the value of the m_LearningRate.

So, the LearningRate value is indeed the maximum "step" length, in physical
units
(e.g. millimeters), that the deformation field will have.

One way to approach the calibration is to set the Learning Rate to half the
dimension
of a pixel. You can do this by using your knowledge of the pixel spacing in
the image.

In this way, you would set up the registration method to move less than one
pixel
at every iteration. This is a smooth way to move, but it may also result in
a larger
number of iterations. That's a trade-off that you may have to evaluate in
the context
of your application.

This criteria of "moving less than one pixel per iteration" is similarly
applied in the
front propagation of Level Sets.


So, if your image has pixel spacing of 3mm,

then a good learning rate to start experimenting with,
could be

                                       3mm / 2.0  =  1.5


    HTH


        Luis
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-users/attachments/20140123/51c05905/attachment.html>