ITK/Release 4/Enhancing Image Registration Framework/Tcon 2010-09-07

Attendees

• Cory Quammen
• Gabe Hart
• Nick Tustison
• Andy
• Brian Avants
• Luis Ibanez

Technical Topics

• Transform hierarchy
  • How to compose multiple transforms into a single
    • ResampleImageFilter only deals with itk::Transform
    • WrapImageFilter only deals with a deformation field
    • A new filter is needed, that takes as input a collection of Transforms and deformation fields and apply them concatenated.
• Potential Names (for this new class)
  • WarpImageMultiTransformFilter
  • ConcatenatedTransformImageTransformFilter
• See the Gaussian down-sampling as another Transformation
  • Avoid storing the entire pyramid in memory (saving memory consumption).
• Generalize the representation of an image by using a Sparse representation of the image.
  • Introduce an image sampling class that generates a Sparse image from an image.
  • Then pass this Sparse Image type to the Metrics.
  • Both for the Fixed and Moving images ?
• How to consolidate a "smart" sampling to allow for
  • Dense sampling
  • Sparse sampling
  • Hide it in the iterator ?
  • Implement a Random iterator for Meshes (random point access) ?
  • Unify the representation of Meshes and Images ? (use SpatialObjects? )

• Projective transforms for CV community
  • http://ibia.umit.at/ResearchGroup/Phil/web/Simple2D3DRegistrationFramework.html

• Maximize MI( I(x) , J(T(x)) ) by gradient methods:
  • \partial Metric / \partial Image \partial Image / \partial Transform \partial Transform / \partial x

Use Cases

• Be able to transform meshes (stored in VTK files) through a combination of
  • Affine Transforms
  • Deformation Field
  • Without having to do more than one interpolation (e.g. via concatenation of Transform).
• Be able to transform Images through a combination of
  • Affine Transforms
  • Deformation Field
  • Without having to do more than one interpolation (e.g. via concatenation of Transform).
• Perform symmetric registration (affine and deformable)(un-biased)
  • Registration in which Fixed and Moving images can be exchanged and the result of the registration will be the same.
  • Implementation: Extract the interpolation from the Metric.
  • Every metric must compute the derivative of the Metric with respect to both
    • The space of the Fixed Image
    • The space of the Moving Image
  • Use an intermediate space to which both images are registered
  • Then two transforms are computed: from the central space to each one of the two images.
• Fit Intensity Models to images
  • E.g. Fitting a Gaussian (PSF) model to a microscopy image
  • Parametric image model
  • Some parameters from the Optimization space will correspond to the image parametric model.
• Geometrical-Model to Image Registration
• Better support for multiplicity (working together in a common registration problem).
  • Multiple Optimizers ?
  • Multiple Metrics ?
• Parameter Mask
  • Selecting a subset of parameters from a larger set.
    • E.g. In a 3D affine transform enable first only the translation parameters
  • Is this related to "bounding" some (or all?) elements in the parameter array ?