Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Look up keyword
Like this
5Activity
0 of .
Results for:
No results containing your search query
P. 1
Multi-Label Image Segmentation for Medical Applications Based on Graph- Cuts

Multi-Label Image Segmentation for Medical Applications Based on Graph- Cuts

Ratings: (0)|Views: 358|Likes:
Published by ijcsis
Mathematical morphology is very attractive for automatic image segmentation because it efficiently deals with geometrical descriptions such as size, area, shape, or connectivity that can be considered as segmentation-oriented features. This paper presents a new approach for some theoretical and practical improvements of image segmentation based on graph cuts. The method is based on the use of the region adjacency graph produced by the watershed transform from mathematical morphology. Marker extraction identifies the presence of homogeneous regions. The combination of morphological and graph cuts segmentation permits us to speed up and define new classes of energy functions that can be minimized using graph cuts. The use of region graphs gives promising results and can potentially become a leading method for interactive medical image segmentation.
Mathematical morphology is very attractive for automatic image segmentation because it efficiently deals with geometrical descriptions such as size, area, shape, or connectivity that can be considered as segmentation-oriented features. This paper presents a new approach for some theoretical and practical improvements of image segmentation based on graph cuts. The method is based on the use of the region adjacency graph produced by the watershed transform from mathematical morphology. Marker extraction identifies the presence of homogeneous regions. The combination of morphological and graph cuts segmentation permits us to speed up and define new classes of energy functions that can be minimized using graph cuts. The use of region graphs gives promising results and can potentially become a leading method for interactive medical image segmentation.

More info:

Published by: ijcsis on Sep 05, 2010
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

12/02/2010

pdf

text

original

 
 
Multi-Label Image Segmentation for MedicalA
PPLICATIONS
B
ASED ON
G
RAPH
-
 
C
UTS
 
R.Rajeswari, P. Anandhakumar,
Research Scholar, Department of Information Technology Asst. Professor, Department of Information TechnologyMIT campus, Anna University MIT Campus, Anna UniversityChennai-44,India Chennai-44, Indiarajimaniphd@gmail.com anandh@annuniv.edu
 Abstract
: Mathematical morphology is very attractive forautomatic image segmentation because it efficiently deals withgeometrical descriptions such as size, area, shape, or connectivitythat can be considered as segmentation-oriented features. Thispaper presents a new approach for some theoretical and practicalimprovements of image segmentation based on graph cuts. Themethod is based on the use of the region adjacency graphproduced by the watershed transform from mathematicalmorphology. Marker extraction identifies the presence of homogeneous regions. The combination of morphological andgraph cuts segmentation permits us to speed up and define newclasses of energy functions that can be minimized using graphcuts. The use of region graphs gives promising results and canpotentially become a leading method for interactive medicalimage segmentation.
 
 Keywords-
Image segmentation, marker extraction, morphology
 ,
Graphs, Watershed Transform
 .
I.
 
I
NTRODUCTION
 
M
AGNETIC resonance imaging (MRI) provides detailedimages of living tissues, and is used for both brain and bodyhuman studies. Data obtained from MR images is used fordetecting tissue deformities such as cancers and injuries. Thekey to any automatic method is that it must be robust, so that itproduces reliable results on every image acquired from anyMR scanner using different relaxation times, slice thicknessesand fields of view. More recently, computer-assisted methodshave been used for specific tasks such as extraction of MSlesions from MRI brain scans [1], [2].A key problem in medical imaging is automaticallysegmenting an image into its constituent heterogeneousprocesses. Automatic segmentation has the potential topositively impact clinical medicine by freeing physicians fromthe burden of manual labeling and by providing robust,quantitative measurements to aid in diagnosis and diseasemodeling. One such problem in clinical medicine is theautomatic segmentation and quantification of brain tumors.The model-aware affinities integrated into the multilevelsegmentation
 
by weighted aggregation
 
algorithm, and applythe technique to the task of detecting and segmenting braintumor and edema in multichannel magnetic resonance (MR)volumes.[3].The flowchart of proposed system shown in Fig.1.The tumorssegmentation is obtained through the computation of aminimal graph cut of the region adjacency graph obtained onthe watershed segmentation of the image. The user has tointeractively specify the location of the tumors as well as amarker to specify the background. A post-processing step isalso proposed to smooth the segmentation result. This post-processing step consists in a morphological opening of thesegmentation.II.
 
R
ELATED
W
ORK
 There are lot of work available related to the proposedwork, in which some of them are significant. It includes
Post processingWatershed TransformRegion Adjacency GraphTumor segmentation byComputation of Minimal Graph CutsOriginal ImageUser specifiedMarkersFig 1.Flow Chart of proposed method
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 5, August 2010142http://sites.google.com/site/ijcsis/ISSN 1947-5500
 
 
automatic and semiautomatic for segmentation of brain intodifferent tissues, including MS lesions. These approachesinclude a variety of methods such as statistical, fuzzy, neuralnetworks, and fuzzy neural networks.Kamber et al.[4]developed a brain tissue model forsegmentation of MR images of patients with MS disease. Theoffered model was three-dimensional, voxel based, whichprovided
a priori
probabilities for white matter, gray matterand CSF.The survey by Clarke
et al.
of segmentation methods for MRimages [5] describes many useful image processing techniquesand discusses the important question of validation. Thevarious image processing techniques used for segmenting thebrain can be divided into several groups: those required toperform a crude threshold-based extraction of the brain,followed by refinement of brain contours; statistical methodsfor brain segmentation, and region growing methods.
 
The work by M. Kass, A. Witkins & D. Terzopoulos [6] in1988 is the basis for the active contour model, the Snakealgorithm. The work by Mathews Jacob, Thierry Blu andMichael unser [7] presents different solutions for improvingspline based snakes, and demonstrates the minimum curvatureinterpolation property used as an argument to get rid of theexplicit smoothness constraint.The work of El Naqa et al [8] proposes a variational methodsbased on multivalued level set deformable models forsimultaneous 2D or 3D segmentation of multimodality imagesor multiple image sets from the same modality.A multidimensional segmentation and filtering methodologyfor accurate blood flow velocity field reconstruction fromphase-contrast magnetic resonance imaging (PC MRI)developed and validated byKartik S. Sundareswaran, DavidH. Frakes et.al[9].Yuri Boykov, Olga Veksler and Ramin Zabih [10] proposes anexpansion algorithm that finds a labeling within a knownfactor of the global minimum, while swap algorithm handlesmore general energy functions. Both algorithms allowimportant cases of discontinuity preserving energies for imagerestoration, stereo and motion.Saracoglu
et al.
[11] modeled the problem using a methodconsisting of three steps: image tessellation, clustering, andclassification. The image was tessellated into regions withsimilar properties using a region growing approach(tessellation step).Based on the “average” color information of the regions, clustering is performed.A novel method for segmentation and classification of M-FISH chromosome images is presented by Petros S. Karvelis,Alexandros T. Tzallas et al.[12]. The segmentation is based onthe multichannel watershed transform in order to defineregions of similar spatial and spectral characteristics. Then, aBayes classifier task-specific on region classification, isapplied.
 
The combination of the multichannel segmentationand the region-based classification is found to improve theoverall classification accuracy compared to pixel-by-pixelapproaches
.
A very powerful segmentation method that has been widelyused in image segmentation problems is the watershedtransform (WT) [13]–[19]. Since its original application ongrayscale images [13], [14], a very attractive computationalform has been derived [15] and extended to color images [16].An extensive review of the watershed algorithms can be foundin [19].The watershed transform presents some advantagesover other developed segmentation methods.1) The watershed lines form closed and connected regions,where edge based techniques usually define disconnectedboundaries that need postprocessing to produce closed regions.2) The watershed lines always correspond to obvious contoursof objects which appear in the image.The main problem of over-segmentation, can be usuallyovercome by the use of preprocessing or postprocessing,producing a segmentation that better reflects the arrangementof objects within the image.III.W
ATERSHED BASED
I
MAGE
T
RANSFORM
 The proposed system starts with the original image andwatershed transform of the image's gradient modulus is usedfor segmentation purpose.The validation is done comparing our segmentation with thephysician’s delineation. Fig 1 shows the flowchart of theproposed method.
 A.
 
Watershed Based on Dissimilarity Measures
The watershed transform definitions can be slightly modifiedin such a way that it produces a segmentation according todifferent path based criteria. For image segmentation purposesthe watershed transform is computed on the image's gradientmodulus, usually the morphological gradient. The estimationof the gradient has thus also an importance for the quality of the segmentation. A first extension of the classical watershedtransform of the gradient image is based on the localdissimilarity between neighbor pixels. This approach, calledwatershed by dissimilarity was originally proposed by Pardas[20] and Lotufo et.al in [21].
 
Given a pixel adjacency graph G = (V;E;W), we considerthe following edges weights mappinge
i,j
Є
E , w
i,j
= (1/d(i,j ).|pi—pj) |+1)where i and j are two nodes of the graph, pi and pj are the greylevel values of neighbor pixels of the image and d(i, j) is thedistance between the two pixels.These edges weights represent the local estimation of theimage's gradient modulus. Instead of computing themorphological gradient of the image at each pixel, the gradientis here estimated "between" the pixels. This weight mapespecially provides a better detection of thin contrastedobjects.
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 5, August 2010143http://sites.google.com/site/ijcsis/ISSN 1947-5500
 
 
This weight map has the main advantage to achieve a betterdetection of small details compared to the classicalmorphological gradient.IV.
 
G
RAPHS IN
I
MAGE
S
EGMENTATION
 This section presents the graphs commonly encountered inimaging applications. A first level is the so-called "pixeladjacency graph". In a second level, an unsupervised low levelsegmentation of the image (i.e. a segmentation that providesmuch more regions than objects in the image) is used to builda "region adjacency graph”.
 A.
 
Pixel Adjacency Graphs
In this case, the set of nodes of the graph is the set of pixels of the image, and the edges link neighboring pixels. Foreach image and an adjacency system is needed to build thecorresponding pixel adjacency graph. Different commonadjacency systems are illustrated in figure.
Figure 1. 4 neighborhood adjacency system (V4).Figure 2. 8 neighborhood adjacency system (V8).Figure 3. 6 neighborhood adjacency system for 3D
 
 B.
 
 Region Adjacency Graph
A region adjacency graph can be obtained from low-levelsegmentation of the image by numerous methods. It can beobtained from a first unsupervised clustering of the image, forinstance one can apply the watershed transform [22, 23], flatzones labeling [24, 25], or k-means clustering [26] to obtainsuch low-level segmentations.The minimal cut problem on region adjacency graph is usedinstead of pixel adjacency graph since it is advantageouslyused to speed up and extend the presented techniques.V.
 
IMPLEMENTATION
 
AND
 
RESULTSThe region adjacency graph is obtained from the watershedtransform computed from all minima of the morphologicalgradient of the original MRI image using Meyer's algorithm.[27]. Then a region adjacency graph is extracted and used forthe next optimization steps. The user specified markers areused to compute a minimal graph cut separating the markersspecifying the myocardium and the external tissues. Themarkers that specify the myocardium denoted as the set of regions Mm. The markers specifying the tissues surroundingthe myocardium are denoted by M
ext.
 
F(r
i
, r
 j
) is defined as the set of edges of the pixel graph connectingtwo regions r
i
and r
 j
of the low-level watershed segmentation.
 
F(r
i
, r
 j
)={e
m,n
Є
E | m
 
Є
 
i ,
n
Є
 
 j
 
} (1)The strictly positive and decreasing function g used as an edgeindicator for the a minimal surface isg(||I(p)||)= (1/(1+||I(p)||)
(2)where the k parameter was set to 2.The edges weights of the region adjacency graph
G
R=
(V
R,
E
R,
W
R
) are then set such that the weight of a graph cut equalsthe energy function of a surface defined as
w
r
 i ,
, r
 j
=
Σ
(g( || I(p)||)) (3)The myocardium boundaries are finally extracted bycomputing a minimal graph cut of the region adjacency graphwith weights given by (3). The minimal cut is obtained on theregion adjacency graph with two additional nodes s and t,respectively connected to the markers of the myocardium andthe markers of the external tissues. The edges weights of thegraph are summarized in Table I.
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 5, August 2010144http://sites.google.com/site/ijcsis/ISSN 1947-5500

Activity (5)

You've already reviewed this. Edit your review.
1 thousand reads
1 hundred reads
kloud liked this
Sahar Sabbaghi liked this
shivapss liked this

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->