2009 UNC HAMMER WML

From NAMIC Wiki
Jump to: navigation, search
Home < 2009 UNC HAMMER WML
Deformable registration using HAMMER (Heirarchical Attribute Matching Mechanism for Elastic Registration)
White Matter Lesion Segmentation, comparison between manual and automated results

Results from the new implementation are:

Components of Hammer Attribute Vectors
Edge points selected at different deformation stage
Similarity map (middle column) computed on moving image (right) for the selected point on the fixed image (left, cross). Point indicated by a cross on the middle and right column is the point on the moving image that is the most similar to the picked point on the fixed image.

Key Investigators

  • UNC: Dinggang Shen
  • GE Research: Xiaodong Tao, Jim Miller
  • UPenn: Christos Davatzikos (Consultant)


Objective

Design ITK classes for Hierarchical Attribute Matching Mechanism for Elastic Registration (HAMMER) and White Matter Lesion segmentation so that these algorithms can be easily implemented using the Insight Toolkit and integrated into Slicer. The goals of the week is to analyze the algorithms on the object level, map components to existing ITK classes, identify gaps, and start implementation. Will also identify testing dataset and generate baseline results for testing.

Approach, Plan

Software process embraced by NA-MIC community. We will be using NITRC as our CVS repository for development. We will develop the algorithms as well as the testing. The current binary distribution of the algorithms will serve as the baseline for our testing.

Progress

Both HAMMER and WML segmentation algorithms have been distributed in binary forms for a number of years and have been used by some research groups. We will start with the algorithms in their current forms to generate ground truth on publically available data sets. A NITRC project (http://www.nitrc.org/projects/hammerwml/) is setup for collaboration between UNC and GE Research.

On Hammer side:

  1. HAMMER attribute vector computation has been implemented as an ITK Filter.
  2. Feature point selection has been implemented.
  3. Feature similarity computation has been implemented.
  4. Test data will be selected from OASIS database.
  5. HAMMER Registration is been implemented using ITK Finite Difference framework--itk::HammerDeformableRegistrationFunction and itk::HammerDeformableRegistrationFilter.

One WML Segmentation side:

  1. Skull stripping is being developed. For WML segmentation, skull stripping is not THE most critical piece. The accuracy requirement is relaxed.
  2. Support vector machine classification will be using/extending SVMShapeAnalysis in Namic Sandbox developed by MIT group.
  3. Test data is yet to be determined.


References