Segmentation

1
Segmentation

• Core 1-3 Meeting, May. 22-23, 2008 - SLC, UT

2
Georgia Tech/JHU
Prostate Segmentation Registration Framework
Yi Gao (Georgia Tech), Allen Tannenbaum (Georgia
Tech), Gabor Fichtinger (JHU)?
3
Background

• Under the roadmap project Brachytherapy Needle
  Positioning Robot Integration.
• Auto/Semiauto segmentation.
• Registration
• Between modalities US/MRI
• Before/during therapy

4
Segmentation

• Two approaches
• Random Walks(RW)?
• RW post process
• Toward automatic segmentation
• Spherical wavelet shape based method

5
Random walks

• Random Walks(RW)?
• Result ?
• Less interaction
• C code

6
Shape based method

• Spherical wavelet shape based method
• Shape learning
• ITK Spherical wavelet transformation
• Shape based segmentation

7
Shape based method

• Spherical wavelet shape based method
• Shape learning
• ITK Spherical wavelet transformation
• Shape based segmentation

8
Shape learning

• Align segmented shapes.
• Registration under Similarity transform.
• Learn aligned shapes.
• Statistical learning PCA, KPCA, GPCA

9
Shape learning

• Align segmented shapes.
• Registration under Similarity transform.
• Learn aligned shapes.
• Statistical learning PCA, KPCA, GPCA

10
Registration

• Common region extraction
• Used as landmark
• Rigid/Deformable registration
• Particle filter/Kalman filter
• Optimal mass transportation

11
Landmark extraction

• Chan-Vese on manifold
• Extract featured region on surface.
• Feature defined by a function.

Color depicts a scalar function defined on a
surface.
12
Landmark extraction, cont.
13
Landmark based Registration

• Concave belly of prostate
• Common among all prostate
• Used as soft-landmark in registration.

14
UNC/MIND
Lesion Segmentation
Marcel Prastawa (Utah), Guido Gerig (Utah),
Jeremy Bockholt (MIND)?
15
(No Transcript)
16
MIND Lupus Lesion
17
Iowa/MIND
Bayesian Classification of Lupus Lesions
Vincent A. Magnotta (Iowa), Jeremy Bockholt
(MIND), Peter Pellegrino (Iowa)?
18
Algorithm Overview

• Tissue classification algorithm coupled with
  lesion identification
• Required Inputs
• T1, T2, and FLAIR images that have been spatially
  normalized and bias field corrected
• Definition of the brain
• Currently uses BRAINS Autoworkup pipeline to
  fulfill these requirements

19
Algorithm

• Uses K-means classification
• Initial estimate of GM, WM, and CSF based on
  minimum, mean, and standard deviation from T1
  weighted image
• Kmeans segmentation into GM, WM, and CSF from T1
  weighted image
• Lesion from FLAIR Images
• Threshold FLAIR image based on mean and standard
  deviation within the brain
• Eliminate lesion voxels adjacent to CSF
• Remaining lesion voxels from the Kmeans
  classification are used to relabel the Kmeans
  labelmap with a Lesion value

20
Bayesian Classification

• Define exemplars for classes
• Randomly sample 1000 points from GM, WM, CSF, and
  Lesion labels
• Used to define the means and variance for the
  classes
• Define class priors
• Extract each class from labelmap generated in
  previous step and filter with a 2mm gaussian
  filter
• Run multi-modal Bayesian classifier
• T1, T2, and FLAIR images input

21
Results
22
MIT/Harvard
Tissue Classification
Kilian Pohl (MIT/BWH), Brad Davis (Kitware),
Sylvain Bouix (Harvard), Marek Kubicki (Harvard),
Martha Shenton (Harvard), Sandy Wells (BWH),
Polina Golland (MIT)?
23
Slicer 3 Module
24
EM-Segmenter

• Intensity normalization
• Structure hierarchy
• Registration
• Atlas-to-subject
• Multimodal
• Applications
• Tissue classification
• Structure parcelation
• MS lesions segmentation

25
Georgia Tech/Harvard
Label Space Segmentation
Jimi Malcolm (Georgia Tech), Allen Tannenbaum
(Georgia Tech), Yogesh Rathi (Harvard)?
26
Label Space

• Problem Constructing an anatomical model for
  multiple, covarying regions
• - Slice from labeled brain
• State of the art
• - Signed distance maps develop artifacts
  along interface between regions, small variations
  on interface cause large perturbations far away
• - Binary vectors background bias during
  registration
• - LogOdds natural probabilistic
  interpretation, but uses the above intermediate
  representations thus incurring similar problems

27
Label Space

• Label Space
• - regular simplex
• - natural algebraic manipulation
• - direct probabilistic interpretation
• - unbiased toward any label

28
Label Space

• Experiments
• - smoothing, interpolation
• - registration
• - probabilistic atlases

29
Questions?