2010 Winter Project Week ProstateSeg

• 

  Projects List

• 

  Prostate segmentation result with the old (RandomWalk) algorithm - already integrated into Slicer

• 

  Prostate segmentation result using the new shape-based algorithm, using a command-line module (ProstateSegSB)

Key Investigators

• Andras Lasso, Gabor Fichtinger (Queen's University)
• Yi Gao, Allen Tannenbaum (Georgia Tech)
• Andriy Fedorov (BWH)

Objective

Implement a Slicer module from the shape-based prostate segmentation algorithm developed by Yi Gao et al.

Approach, Plan

Implement as a command-line module that can be downloaded and installed as a Slicer extension. Add automatic testing.

Progress

• The segmentation Slicer module has been created (available at http://svn.na-mic.org/NAMICSandBox/trunk/Queens/ProstateSeg/ShapeBased)
• Tested segmentation on 3 images, execution time is 10-20 minutes
• Shape learning algorithm (for segmentation of any other shape than prostate) is added to the repository, and tested on Linux and Windows
• To do: add as an extension module (partially done), add test, tutorial

References

• Yi Gao, Romeil Sandhu, Gabor Fichtinger, Allen Tannenbaum, A Coupled Global Registration and Segmentation Framework with Application to the Magnetic Resonance Prostate Imagery, IEEE Trans Med Imaging (in review)

Notes

Training

• each directory contains an executable together with a bash script
• Registration
  • alignTrainingShapes.bash (execution time is about 30min) (AF: with the default value of 50000 points, execution time is few (>3?) hours)
    • make isotropic (z direction)
    • register
      • ShapeBased\_reg\src\imageRegByPointSet\c\affine\ => 1 executable
        • pairwise image registration (there are many supporting files in ShapeBased\_reg\src; the result is one executable)
        • input: two images from ShapeBased\_reg\trainingShapes
        • output: transformed moving image in uchar nrrd image format
    • make anisotropic (maybe this step could be skipped, to have an atlas with isotropic images)
    • results are copied to ShapeBased\_reg\alignTrainingShapes
  • alignTrainingShapesNonIso.bash: faster but not that accurate
• Convert from binary to level set
  • ShapeBased\_reg\alignTrainingShapes\toSFLS\ => 1 executable + 1 script (toSFLS.bash)
    • Input: nrrd binary image
    • Output: level set description
  • For each binary image a level set image is generated and saved to ShapeBased\_reg\alignTrainingShapes\toSFLS
• Learning using PCA
  • ProstateSeg\ShapeBased\_reg\alignTrainingShapes\toSFLS\learn => 1 executable
    • Input: shapeList.txt list of all level set files
    • Output: mean shape and i-th eigen shape (multiplied by the eigen value),
  • Execution time is about 1 minute, repeated for each eigen shape
• Images are flipped, but the images to be segmented (or the training shapes) could be flipped instead.

Segmentation

• ProstateSeg\ShapeBased\bin\wholeSeg => 1 executable (wholeseg)
  • Input: image to be segmented, and two points (at the left and right side of the prostate, on a center axial slice in IJK space)
  • Output: ?

To do

• Send intermediate training data (to test the training steps) => Yi
• Slicer module for segmentation:
  • Add/use the boost library in an extension module
  • Filter for .sm file extension
  • Add model files to the package (or upload the model to the project wiki)
• Add a quick test