2011 Summer Project Week Breakout Session EMRegistration

From NAMIC Wiki
Jump to: navigation, search
Home < 2011 Summer Project Week Breakout Session EMRegistration

Summer Project Week Breakout Session: Registration for EM Segmenter

Wednesday, June 22, 2011, 3-4pm, Star room.

Synopsis:

The EM segmenter includes as a vital step the co-registration of a multiple subject atlas to obtain spatial priors for small region detection. The success of these priors depends heavily on the quality of the registration obtained. Because the registration includes a large-DOF nonrigid inter-subject registration, benchmarks for a good solution are largely absent. We seek an implementation of a registration using Slicer registration tools that will provide best possible robustness and if possible an estimate of reliability. Because this registration will be integrated into a slicer module, tools with a BSD license are strongly preferred.


Attendees (tentative):

  • Kilian Pohl
  • Dominik Meier
  • Dominique Belhachemi
  • Greg Sharp
  • Danielle Pace
  • Hans Johnson
  • Jim Miller
  • Daniel Haehn

Reference Material:

Preliminary Agenda:

  • Description of the registration task for the EM segmenter
  • Review of example case
  • Comparison of BRAINSfit and CMTK strategies
  • Discussion of design/implementation options/strategies


Meeting Notes:

  • Hans pointed out that an upcoming fix, soon to be in ITK4, will incorporate an improved placement of control points, which reduces the need for masking
  • the multiresolution approach and fine & adaptive grid are likely the reason for the good performance of CMTK on inter-subject registration
  • CMTK takes ~1hr and does not have a BSD license, bootstrapping of transforms for saving time require transform concatenation
  • skull stripped masks are not necessarily avail. for all use cases and should not be expected as avail. input
  • emulation of CMTK behavior would require concatenation of multiple nonrigid transforms, which is also forthcoming in ITK4
  • Alternatives:
    • Plastimatch was not designed for brain MRI, but it does have multiresolution
    • 2 track approach, using multiple versions that use different registration methods
    • NA-MIC sandbox options: Andryi
    • Nonrigid registration within the ANTS package