2008 Winter Project Week Image Registration Update
Back to AHM_2008
Breakout session on "Registration methods in the NAMIC Toolkit"
Attendees
Please add your name below, if your plan on attending.
- Guido Gerig
- Jim Miller
- Stephen Aylward
- Ross Whitaker
- Luis Ibanez
- Alex hanfei Gouaillard
- Casey Goodlett
- Bill Lorensen
- Luca Antiga
- Clement Vachet
- Ron Kikinis
- Serdar Balci
- Daniel Blezek
- Li Shen
- Kilian Pohl
- Xiaodong Tao
- Marcel Prastawa
- Francois Budin
- Padma Sundaram
- Jeremy Bockholt
- Mark Scully
- David Gobbi
Topics
Please submit your ideas for topics to be discussed.
- Demonstration of a registration module in Slicer (Stephen)
- Needs, Categorization (Guido)
- What is available (Guido)
- Use in a pipeline (Guido)
- Choice of appropriate/optimal algorithm, Validation (Guido)
- Testing (Luis)
- Heuristics for selecting components(Luis)
Presentations
If you would like to give a short (5-10 minute) presentation, please describe it below.
- Opening remarks (Guido)
- Demonstration of a registration module in Slicer (Stephen, Matt Turek, and Luis Ibanez)
- Discussion of directions for future work (assign priorities and responsibilities)
- Passing and tracking transforms in Slicer (Jim)
- Regularization of BSplines (Stephen)
- Adding "don't process" zones in BSplines (Stephen)
- Objects as masks: mrml -=> SpatialObject conversion (Stephen)
- Reducing BSpline memory requirements (Brad)
- Specializing particular metric/transform combinations (Luis)
- Batch processing registration
- Atlas formation
- Atlas application
- Validation
- STAPLE
Related Links
Add papers, presentations, wiki pages, etc.
Notes and Summary
Introduction by Guido Gerig.
Detailed presentation by Stephen Aylward w.r.t. the development of a registration toolkit to Slicer-3 (see status slides above), developed by the team Aylward/Ibanez/Turek.
A new 3D registration kit with rigid, affine and b-spline transformations is added to Slicer-3. Image match metrics include MattesMI (mutual information), normalized correlation and mean squared error, and three interpolators are available (linear, bspline, windowed sinc). A user has the choice of 3 optimization methods (gradient, multi-resolution with 3 levels, and 1+1 evolutionary followed by gradient).
The algorithm performs a series of steps from initial alignment over rigid, affine to BSpline. The image is finally transformed by cascading the set of transformations into one transformation step using the interpolation mode of choice.
Stephen Aylward showed a demonstration of a nonlinear registration of a subject-to-atlas matching with a case that showed difficulties using alternative registration techniques.
The BSpline mode uses a user-defined sampling of space. The current implementation does not yet allow for very dense sampling, e.g. every 2nd mm, as ultimately required for EMS to segment brain substructures at the finest resolution level. Further optimization of memory management and reducing #operations were discussed (Polina Golland's group and engineering core) and these Core-1 and Core-2 developers will get together to discuss options.
