Projects:ProstateSegmentation - Revision history

Ygao at 01:00, 16 November 2013

2013-11-16T01:00:08Z

Ygao at 15:47, 12 January 2012

2012-01-12T15:47:38Z

Ygao: /* Publications */

2010-05-07T16:15:39Z

Publications

Kikinis: /* Publications */

2009-12-04T17:10:32Z

Publications

Melonakos at 18:27, 20 October 2009

2009-10-20T18:27:17Z

Melonakos at 18:27, 20 October 2009

2009-10-20T18:27:00Z

Melonakos at 19:12, 1 September 2009

2009-09-01T19:12:47Z

Melonakos at 17:32, 20 July 2009

2009-07-20T17:32:10Z

Melonakos at 19:12, 11 July 2009

2009-07-11T19:12:48Z

Melonakos at 19:10, 11 July 2009

2009-07-11T19:10:04Z

@@ Line 1: / Line 1: @@
-  Back to [[Algorithm:BU|Boston University Algorithms]]
+  Back to [[Algorithm:Stony Brook|Stony Brook University Algorithms]]
 __NOTOC__
 = Prostate Segmentation =

@@ Line 1: / Line 1: @@
-  Back to [[Algorithm:GATech|Georgia Tech Algorithms]]
+  Back to [[Algorithm:BU|Boston University Algorithms]]
 __NOTOC__
 = Prostate Segmentation =

@@ Line 82: / Line 82: @@
 = Publications =
-''In Press''
+In Press
-Y. Gao, A. Tannenbaum; Shape based MRI prostate image segmentation using local information driven directional distance Bayesian method. SPIE Medical Imaging 2010 (in submission)
+Y. Gao, R. Sandhu, G. Fichtinger, A. Tannenbaum; A coupled global registration and segmentation framework with application to magnetic resonance prostate imagery. IEEE TMI (in submission)
 = Recent Work =

@@ Line 86: / Line 86: @@
 Y. Gao, A. Tannenbaum; Shape based MRI prostate image segmentation using local information driven directional distance Bayesian method. SPIE Medical Imaging 2010 (in submission)
   Project Week Results: [[2009_Summer_project_week_prostate_registration|Jun 2009]]
 [[Category: Segmentation]] [[Category: Prostate]]

← Older revision		Revision as of 18:27, 20 October 2009
Line 84:		Line 84:
	''In Press''		''In Press''

−	Y. Gao, A. Tannenbaum; Shape based MRI prostate image segmentation using local information driven directional distance Bayesian method. SPIE Medical Imaging 2010 (in submission).	+	Y. Gao, A. Tannenbaum; Shape based MRI prostate image segmentation using local information driven directional distance Bayesian method. SPIE Medical Imaging 2010 (in submission)

← Older revision		Revision as of 18:27, 20 October 2009
Line 81:		Line 81:

	= Publications =		= Publications =
		+
		+	''In Press''
		+
		+	Y. Gao, A. Tannenbaum; Shape based MRI prostate image segmentation using local information driven directional distance Bayesian method. SPIE Medical Imaging 2010 (in submission).

← Older revision		Revision as of 19:12, 1 September 2009
Line 83:		Line 83:


		+	Project Week Results: [[2009_Summer_project_week_prostate_registration\|Jun 2009]]

	[[Category: Segmentation]] [[Category: Prostate]]		[[Category: Segmentation]] [[Category: Prostate]]

@@ Line 10: / Line 10: @@
 Random Walks for prostate segmentation
-'''Overview'''
+== Overview ==
 The algorithm starts with a few initial seeds marking object and background. Then to decide the category of each of the other(non-seed) pixels, a random walker is released and it walks on the lattice of the image. The difficulty of walking from on pixel position to another is inversely related with the difference between the intensities of the two pixels. Given those settings, the probability of the random walker first hitting each kind of seeds are calculated. And the category of the pixel is assigned to the one with the greatest first hit probability.
@@ Line 16: / Line 16: @@
 Ideally, through this process the category of each pixel in the image could be decided.
-''Algorithm detail''
+== Algorithm detail ==
 The philosophy of random walk segmentation being stated above, it's practically impossible to implement that way. (Say, releasing many random walkers at each pixel and get the first hit probability using Monte-Carlo method.)
@@ Line 28: / Line 28: @@
 In the above equation the <math>\triangle </math> is a Laplace-Beltrami operator where the spatial variant resistance is considered to be metric of the space.
-''Shape based segmentation''
+== Shape based segmentation ==
 The shape based method could constrain the resulting segmentation within the learned variance of the shape. The algorithm consists two stages, learning the shapes and segment new case.
-''Shape learning''
+== Shape learning ==
 The training set consists 29 data sets, each of which is manually segmented. Not only the prostate is segmented, but also the rectum. The reason for segmenting rectum will be given later. The segmented results are saved in label map where 0 indicates background, 1 means prostate while 2 is the rectum.
@@ Line 48: / Line 48: @@
 Principle component analysis was carried on to learn the above representations of the shapes. And the eigen-vectors with high eigen-values are extracted as the basis of the shape space. Next, we segment a new image within this learned shape space, denoted as S.
-''Segment new shape''
+== Segment new shape ==
 The new image to be segmented is first preprocessed to high light the region of prostate. Then, within the shape space S, we look for the one that after similarity transformation, best fits the preprocessed image. The fitness is defined by the Binomial energy proposed by Yezzi et. al. in ''A Statistical Approach to Snakes for Bimodal and Trimodal Imagery''. Next we detail the pre-processing and the segmentation.
-''Pre-processing''
+== Pre-processing ==
 In the learning process we not only learn the shape, but also we could know the mean and standard divination of the intensity of prostate. For the given new image, we compute the likelihood image using the mean and std. To get the posterior image, uniform prior is used widely in literature. However that will include other regions with similar intensities. With another pieces of information, one point in the prostate, we could get better posterior image. That will further aid the segmentation. Next we detail the construction of prior.
@@ Line 66: / Line 66: @@
 From which we can see that the distance modulated posterior eliminates the far field interference while maintain the shape(not favoring a spherical shape).
-''Segmentation of the pre-processed image''
+== Segmentation of the pre-processed image ==
 Given the posterior map, we want to find one shape in the shape space to minimize the Binomial energy. Because the shape in the learned shape space is represented by basis, instead of evolving the curve/surface, we only need to optimize over the coefficients of the basis. So the problem is turned into a finite dimensional non-linear optimization problem.