Tissue Image Segmentation

1
Tissue Image Segmentation

• - Presenter Lin Yang
• - Advisor Dr. David J. Foran
• - A General Framework for Segmenting Imaged
  Pathology Specimens Using Level-set and
  Gaussian Hidden Markov Random Fields

2
Problem Statement

• Image Segmentation
• Region based method
• Segmentation by clustering mean shift
• Segmentation by graph theory
• Segmentation by MRFs, Gaussian Mixture Models and
  EM algorithm
• Contour based method
• Active contour models
• Traditional KWT snake
• GVF snake
• Geodesic snake
• Level set based snake
• Active contour without edge

3
The Choice of Filter Bank(1)

• The Gabor filter bank
• The Leung Malik (LM) filter bank

4
The Choice of Filter Bank(2)

• The Schmid filter bank
• The Maximum Response (MR) filter bank

5
MRF Segmentation Model

• Assume a set of observed (y) and hidden (x)
  random variables
• f?y represents the low-level features
• ??x represents the labels of each pixel
• Now the segmentation problem can be modeled as a
  MAP(maximum a posterior) estimation

6
Gibbs prior

• Gibbs prior
• Intuitive Understanding
• Hammersley-Clifford theorem

7
Gaussian Mixture Model

• Given feature f, the Gaussian Mixture Model is
  defined as follows

8
Initialization and EM

• Applying EM algorithm to get the MLE estimation
  of the parameters set W

9
Complete Cost Function

• The complete cost function combining the Gaussian
  mixture models and the Gibbs priors will have the
  following forms
• Notice that the parameters are the results of EM
  algorithm

10
Optimization Algorithm (1)

• Stochastic optimization
• Simulate Annealing
• Gibbs Sampling
• Global Minimum
• Algorithm
• Code from Matlab

11
Optimization Algorithm (2)
12
Experimental Results(1)

• Synthetic Image

13
Experimental Results(2)

• Standard Texture Image

14
Level Set Based Active Contour

• Traditional Snake
• Topological change
• Difficulty with initialization problem GVF
  snake partially solve this problem
• Level Set or Geodesic Snake
• Topology changes can be easily handled and
  initial positions are not sensitive
• Computation is complex, speed is slow and the
  implementation is relatively difficult
• Multiphase level-set framework very fast
• Snake with MRF
• Apply snake on the likelihood map of MRF can mix
  the advantages of MRF and snake

15
Experimental Results(3)
16
Experimental Results(4)
17
Performance Evaluation

• Features are more important than classification
  algorithm
• Deformable Model
• None of the gradient based or even region based
  deformable model alone works well in our real
  case
• Gaussian Mixture Model
• The result is not very good because it will
  over-segment the image
• MRF based GMM will improve the result because the
  introduction of Gibbs prior
• Clustering Based Segmentation
• Actually provide satisfactory results for texture
  only segmentation
• Has some problem with homogenous segmentation
  when combined with intensity information
• Total unsupervised approach is very hard for our
  application

18
Pros and Cons

• Advantages
• Actually perform very well for our application.
• Can be combined with many different segmentation
  models
• Still active field and even show up in CVPR 2005.
• Disadvantages
• Speed, speed and speed
• Hundreds of, if not thousands of, literatures are
  proposed for increasing the speed.
• Matlab implementation and C/C implementation,
  big difference, the C implementation takes only
  no more than 1 minute for one image with 600600
  pixels
• Gaussian Models are not always, if not never,
  hold for many medical image processing
  applications

19
Reference

• Chad Carson, Serge Belongie, Hayit Greenspan and
  Jitendra Malik, Blobworld Image Segmentation
  Using Expectation-Maximization and Its
  Application to Image Querying, IEEE Tran. on
  Pattern Anal. and Mach. Intell., vol 24, no. 8,
  pp1027-1037
• C. Bouman and B. Liu, Multiple Resolution
  Segmentation of Textured Images,'' IEEE Trans. on
  Pattern Anal. and Mach. Intell., vol. 13, no. 2,
  pp. 99-113, Feb. 1991.
• C. A. Bouman and M. Shapiro, A Multiscale Random
  Field Model for Bayesian Image Segmentation,''
  IEEE Trans. on Image Processing, vol. 3, no. 2,
  pp. 162-177, March 1994
• R. O. Duda, P. E. Hart, and D. G. Stork, Patten
  Classification, 2nd Edition, Wiley, 2000.
• David A. Forsyth and Jean Ponce, Computer Vision
  A Modern Approach, 1st Edition, Prentice Hall,
  2003.
• Mario A. T. Figueiredo, Bayesian Image
  Segmentation Using Wavelet-Based Priors, CVPR,
  vol. 1 pp 437-443, 2005.
• R. Malladi, J. A. Sethian, B. C. Vemuri, "Shape
  Modeling with Front Propagation A Level Set
  Approach," IEEE Trans. on Pattern Anal. and Mach.
  Intell., vol. 17 No. 2 158-175, Feburary 1995.
• T. F. Chan, L. A. Vese, "A Level Set Algorithm
  for Minimizing the Mumford-Shah Functional in
  Image Processing," Proceedings of the IEEE
  Workshop on Variational and Level Set Methods,
  pp. 161-171, 2001.
• Y. Zhang, M. Brady, S. Smith, Segmentation of
  brain MR images through a hidden Markov random
  field model and the expectation-maximization
  algorithm, IEEE Transactions on Medical Imaging,
  Vol. 20, no 1, pp. 45 57, Jan 2001
• T. Leung and J. Malik, Representing and
  recognizing the visual appearance of materials
  using three-dimensional textons, International
  Journal of Computer Vision, 43(1)29-44, June
  2001

20
Thank You