User:Ramirez: Difference between revisions

From KitwarePublic
Jump to navigationJump to search
No edit summary
No edit summary
Line 19: Line 19:


== Tuning Parameters ==
== Tuning Parameters ==
=== Optimizers ===




Line 110: Line 113:


http://public.kitware.com/pipermail/insight-users/2002-October/001400.html
http://public.kitware.com/pipermail/insight-users/2002-October/001400.html
<b><u>itk::RegularStepGradientDescentOptimizer->SetMaximumStepLength()</u></b>
<b><u>itk::RegularStepGradientDescentOptimizer->SetMinimumStepLength()</u></b>
"There is no magic recipe for selecting one. You probably
want to start experimenting with a small value (e.g. 0.01)
and plot the metric evaluations during the registration
process.  If you observe that the metric values are fairly
monotonic, that means that you can safely increment the
step length. Such an increment has the advantage of reducing
the time required to reach an extrema of the cost function
(the image metric in this case).  You could restart the
registration with larger values of the step length, as long
as you don't observe a noisy and/or erratic behavior on the
Metric values.
Step length issues are discussed in the course material
from the "Image Registration Techniques" course at RPI.
  http://www.cs.rpi.edu/courses/spring04/imagereg/
for example in lecture 9:
http://www.cs.rpi.edu/courses/spring04/imagereg/lecture09.ppt"
http://public.kitware.com/pipermail/insight-users/2004-July/009558.html





Revision as of 14:34, 31 December 2005

Image Registration Components

Image Similarity Metrics


Transforms

Optimizers

Interpolators

Tuning Parameters

Optimizers

Optimizer->SetScales()


The rule of thumb is to figure out how much each one of those parameters will change for your registration, and then rescale that range to [-1:1].


In the case that you know the anticipated range of translations and rotations,


"if you are doing 2D rigid you will have a 2D transform with three parameters:

     Tx  translation in millimeters along X
     Ty  translation in millimeters along Y
     R   rotation in radians

and you anticipate that your images need a correction of the order of 10 to 50 millimeters in translation and 0.01 to 0.1 radians in rotation, then you should put scales:

     scale[0] = 1/50;     scale for Tx
     scale[1] = 1/50;     scale for Ty
     scale[2] = 1/0.1;    scale for Rotation

Of course, those will be just "good values to start with". You will still need to refine them according to the behavior of the optimizer."

http://public.kitware.com/pipermail/insight-users/2005-April/012896.html


In the case that you do not know the anticipated range of translations and rotations,


"the recommendation for the scaling of translation parameters versus rotation parameter is to use a factor proportional to the diagonal length of the image.

For your case the, you have 100 pixels with 1 mm / pixel, therefore the physical extent of your image is

       100mm  X  100mm  X 100mm

The diagonal the image bounding box is

         sqrt(3) * 100 mm

which is about

             173.2

and extra factor of 10X is usually useful, so you should probably try a factor of

    1.0 / ( 10 x 173.2 )  =  1.0 / 1732.0

You could use this same factor for the three components of the translation or you could estimate independent factor for each component in the way it is done in the VolView plugin.

Note that this factors are not expected to be computed precisely. Their purpose is simply to bring the rotational and translational parameters to a similar numerical scale.

By default, they are quite disproportionate since rotation are in radians, therefore in a range about -1:1, while translations are in millimeters, and for an image of 100mm you probably can expect translations as large as 50mm."

http://public.kitware.com/pipermail/insight-users/2004-July/009558.html


In short,


"for an 3D AffineTransform, you get 12 parameters: the first 9 are the coefficients of the matrix (representing rotation, scale and shearing) the last 3 are the components of a translation vector. You want then to provide an array of 12 values with the first 9 being =1.0 and the last three being on the range of 1.0 / the image size (in millimeters)."

http://public.kitware.com/pipermail/insight-users/2002-October/001400.html


itk::RegularStepGradientDescentOptimizer->SetMaximumStepLength()

itk::RegularStepGradientDescentOptimizer->SetMinimumStepLength()

"There is no magic recipe for selecting one. You probably want to start experimenting with a small value (e.g. 0.01) and plot the metric evaluations during the registration process. If you observe that the metric values are fairly monotonic, that means that you can safely increment the step length. Such an increment has the advantage of reducing the time required to reach an extrema of the cost function (the image metric in this case). You could restart the registration with larger values of the step length, as long as you don't observe a noisy and/or erratic behavior on the Metric values.

Step length issues are discussed in the course material from the "Image Registration Techniques" course at RPI.

  http://www.cs.rpi.edu/courses/spring04/imagereg/

for example in lecture 9:

http://www.cs.rpi.edu/courses/spring04/imagereg/lecture09.ppt"

http://public.kitware.com/pipermail/insight-users/2004-July/009558.html



Use Cases

CT-MRI Brain Registration

PET-CT Registration



ITK: [Welcome | Site Map]