MRI Tissue Classification with Neighborhood Statistics: A Nonparametric, EntropyMinimizing Approach'

1
MRI Tissue Classification with Neighborhood
Statistics A Nonparametric, Entropy-Minimizing
Approach.

• Tolga Tasdizen, Suyash P. Awate, Ross T. Whitaker
  and Norman L. Foster
• University of Utah
• MICCAI 2005

2
Outline

• State-of-the-art background
• The new algorithm
• Distribution estimation
• Segmentation estimation
• Comparison with state-of-the-art
• Results

3
Segmentation of Brain MRI

• State-of-the-art
• Parametric statistical model of single pixel
  intensity (usually Gaussians) for each tissue
  class
• MRF models of spatial smoothness
• Bias field correction
• Brain atlas

4
Intensity model of tissues
5
Brain atlas

• Register with atlas
• Use as prior information

6
Bias correction

• Low frequency intensity change
• Needs to be removed for successful segmentation

7
Spatial smootness

• Include information from neighbourhood to handle
  the noise
• Markov Random Fields

8
The new algorithm

• Estimates the true intensity using non-parametric
  model
• Calculates the segmentation by minimizing entropy
• Deals with the noise by including the
  neighbourhood in the voxel intensity model
• Bias correction can be included

9
Intensity distribution estimation

• Gaussians are not ideal model of intensity
  distribution of a tissue
• Estimate the true intensity using Parzen Window
• Include the neighbourhood 7-dimensional
  distribution gt deal with the noise

10
Algorithm

• Initialise segmentation using atlas
• Estimate (7D)intensity distribution for each
  tissue class
• Calculate new segmentation to minimise entropy
• Iterate between 2 and 3 until entropy converges

11
Distribution Estimation

• pk(Zz) likelihood function probability of
  observing the neighbourhood z given tissue type k
  at center voxel t.
• Distribution estimation by Parzen window given
  the segmentation
• Ak(t) randomly chosen set of voxels labeled as
  tissue k
• G gaussian distribution

12
Segmentation estimation

• Tk set of voxels assigned to tissue k
• New segmentation For each voxel choose the class
  with highest likelihood
• Minimizes (approximation of) entropy H

13
Results

• Compare with state-of-the-art method
• The method uses gaussian tissue intensity
  distribution, maximum likelihood and EM algorithm
• Reference
• K. Van Leemput, F. Maes, D. Vandermeulen, P.
  Suetens. Automated model-based tissue
  classification of MR images of the brain. IEEE
  Trans. Med .Imag. 1999

14
Results

• Segmentation better than state-of-the-art
• Disadvantage - speed

15
Results