Liang 1993: Tissue classification and segmentation of MR images

[ CogSci Summaries home | UP | email ]
http://www.jimdavies.org/summaries/

Liang, Z. (1993). Tissue classification and segmentation of MR images. IEEE Engineering in Medicine and Biology. March, pp 81--85.

@Article{Liang1993,
  author = 	 {Liang, Zhengrong},
  title = 	 {Tissue classification and segmentation of MR images},
  journal = 	 {IEEE Engineering in Medicine and Biology},
  year = 	 {1993},
  OPTkey = 	 {},
  OPTvolume = 	 {},
  OPTnumber = 	 {},
  OPTpages = 	 {81--85},
  OPTmonth = 	 {March},
  OPTnote = 	 {},
  OPTannote = 	 {}
}

Author of the summary: Jim Davies, 2006, jim@jimdavies.org

Cite this paper for:

image segmentation

image classification

edge detection

validation

region identification

thresholding

Tissue classification of MR images group image elements and refer to the groups as classes.

Segmentation: Assignment of labels to each voxel (3D pixel) in an image.

Doing this automatically often is done through comparison to a model.

Validation or classification is the determination of the appropriate number of classes.

The two main kinds of segmentation:

edge detection: detects changes in gradients and calls them edges. Inaccurate sometimes because it's sensitive to noise, particularly between regions with small contrast differences.
region identification: Thresholding is classifying each voel with similar intensities. The region growing method starts with one voxel and grows until the surrounding voxels are a different intensity. Another way to do it is to start with arbitrary regions, and merge and split them depending on their inter-region similarity and region uniformity. The merging and splitting technique breaks an image up like a quad-tree, putting together similar regions and decomposing others.

Textures are problematic for these methods. Edge detectors will create edges in a texture, and region ID methods will make little regions for parts of textures. A way to avoid this is to characterize each voxel by the texture properties of its neighbor.

This paper then describes in detail an unsupervised method that assumes the values of voxels belonging to an image class the follows a statistical distribution and that all voxels fit a finite mixture. It fits voxel values to the mixture using the ML (maximum likelihood) principle, and the number of classes using information criteria.

conclusions:

the key feature of statistical approaches is determinign the number of image classes and the model parameters of these classes from the image data.
free parameters from the user should be avoided.

Summary author's notes:

none

Back to the Cognitive Science Summaries homepage
Cognitive Science Summaries Webmaster:

JimDavies (jim@jimdavies.org)