Development of Unsupervised Image Segmentation Schemes for Brain MRI using HMRF model

By: Pradhan, SmitaContributor(s): Patra, Dipti [Supervisor] | Department of Electrical EngineeringMaterial type: TextTextLanguage: English Publisher: 2010Description: 113 pSubject(s): Engineering and Technology | Electrical Engineering | Image SegmentationOnline resources: Click here to access online Dissertation note: Thesis (M.Tech (R))- National Institute of Technology, Rourkela Summary: Image segmentation is a classical problem in computer visio n and is of paramount im- portance to medical imaging. Medical image segmentation is an essential step for most subsequent image analysis task. The segmentation of anatom ic structure in the brain plays a crucial role in neuro imaging analysis. The study of many br ain disorders involves accu- rate tissue segmentation of brain magnetic resonance (MR) i mages. Manual segmentation of the brain tissues, namely white matter (WM), gray matter ( GM) and cerebrospinal fluid (CSF) in MR images by an human expert is tedious for studies in volving larger database. In addition, the lack of clearly defined edges between adjace nt tissue classes deteriorates the significance of the analysis of the resulting segmentati on. The segmentation is further complicated by the overlap of MR intensities of different ti ssue classes and by the pres- ence of a spatially and smoothly varying intensity in-homog eneity. The prime objective of this dissertation is to develop strategies and methodolo gies for an automated brain MR image segmentation scheme. As an initial attempt in this direction, the brain MR image se gmentation problem is addressed in an unsupervised framework and is formulated as pixel labeling problem. Stochastic model based approach has been considered for the same. Hidden Markov Ran- dom Field (HMRF) models have been used to model the tissue cla sses of the observed degraded image. The a priori class labels are modeled as Markov Random Field (MRF) model. As the problem is addressed in an unsupervised framew ork, HMRF model pa- rameters are assumed to be unknown. It is assumed to have the a priori knowledge of MRF model parameters which are used to model the unknown clas s labels, but no knowl- edge of number of classes and image labels. The problem becom es an incomplete data problem. To handle this problem, Expectation-Maximizatio n algorithm is used. In or- der to incorporate a variable spatial characteristics whic h varies with internal part of the brain, the energy function associated with the a priori model is modified by an biased factor. This factor controls the effect of spatial informat ion to avoid identical spatial in- formation throughout the brain. The proposed modified model is named as Biased HMRF v (BHMRF) model. Intensity inhomogeneity or multiplicative bias field in brain MR image is also corrected in the proposed scheme. The results obtain ed by the proposed BHMRF- EM framework are compared with that of HMRF-EM scheme. The pr oposed scheme is found to be outperforming the later one and is observed to be a n efficient method for brain MR image segmentation corrupted by biasfield. In order to address the problem from practical stand point, a new notion of im- age segmentation is introduced by incorporating the fuzzy c lustering approach in HMRF framework. The proposed approach is formulated using fuzzy c-means (FCM) algorithm which is facilitated by a priori MRF distribution. In this regard, HMRF oriented mod- ification of the fuzzy objective function is incorporated. H MRF-EM scheme is found to be sensitive to the initial set of parameters. This has bee n overcome by proposing fuzzy clustering -EM (FCEM) algorithm that does not require to have a proper choice of initial parameters. In the proposed HMRF-FCEM scheme, comb ined strength of fuzzy clustering approach as well as HMRF model are incorporated. The result obtained by the proposed FCEM algorithm in HMRF-FCEM scheme are compare d with that of ex- isting schemes and the results are quite comparable to the la ter ones. The performance of proposed algorithm could be successfully tested with an a rbitrary set of initial model parameters. Both BHMRF-EM and HMRF-FCEM schemes could be validated for h ealthy as well as diseased brain MR images.
Tags from this library: No tags from this library for this title. Log in to add tags.
    Average rating: 0.0 (0 votes)
Item type Current location Collection Call number Status Date due Barcode
Thesis (Ph.D/M.Tech R) Thesis (Ph.D/M.Tech R) BP Central Library
Thesis Section
Reference Not for loan T109

Thesis (M.Tech (R))- National Institute of Technology, Rourkela

Image segmentation is a classical problem in computer visio
n and is of paramount im-
portance to medical imaging. Medical image segmentation is
an essential step for most
subsequent image analysis task. The segmentation of anatom
ic structure in the brain plays
a crucial role in neuro imaging analysis. The study of many br
ain disorders involves accu-
rate tissue segmentation of brain magnetic resonance (MR) i
mages. Manual segmentation
of the brain tissues, namely white matter (WM), gray matter (
GM) and cerebrospinal fluid
(CSF) in MR images by an human expert is tedious for studies in
volving larger database.
In addition, the lack of clearly defined edges between adjace
nt tissue classes deteriorates
the significance of the analysis of the resulting segmentati
on. The segmentation is further
complicated by the overlap of MR intensities of different ti
ssue classes and by the pres-
ence of a spatially and smoothly varying intensity in-homog
eneity. The prime objective
of this dissertation is to develop strategies and methodolo
gies for an automated brain MR
image segmentation scheme.
As an initial attempt in this direction, the brain MR image se
gmentation problem
is addressed in an unsupervised framework and is formulated
as pixel labeling problem.
Stochastic model based approach has been considered for the
same. Hidden Markov Ran-
dom Field (HMRF) models have been used to model the tissue cla
sses of the observed
degraded image. The
a priori
class labels are modeled as Markov Random Field (MRF)
model. As the problem is addressed in an unsupervised framew
ork, HMRF model pa-
rameters are assumed to be unknown. It is assumed to have the
a priori
knowledge of
MRF model parameters which are used to model the unknown clas
s labels, but no knowl-
edge of number of classes and image labels. The problem becom
es an incomplete data
problem. To handle this problem, Expectation-Maximizatio
n algorithm is used. In or-
der to incorporate a variable spatial characteristics whic
h varies with internal part of the
brain, the energy function associated with the
a priori
model is modified by an biased
factor. This factor controls the effect of spatial informat
ion to avoid identical spatial in-
formation throughout the brain. The proposed modified model
is named as Biased HMRF
v
(BHMRF) model. Intensity inhomogeneity or multiplicative
bias field in brain MR image
is also corrected in the proposed scheme. The results obtain
ed by the proposed BHMRF-
EM framework are compared with that of HMRF-EM scheme. The pr
oposed scheme is
found to be outperforming the later one and is observed to be a
n efficient method for brain
MR image segmentation corrupted by biasfield.
In order to address the problem from practical stand point, a
new notion of im-
age segmentation is introduced by incorporating the fuzzy c
lustering approach in HMRF
framework. The proposed approach is formulated using fuzzy
c-means (FCM) algorithm
which is facilitated by
a priori
MRF distribution. In this regard, HMRF oriented mod-
ification of the fuzzy objective function is incorporated. H
MRF-EM scheme is found
to be sensitive to the initial set of parameters. This has bee
n overcome by proposing
fuzzy clustering -EM (FCEM) algorithm that does not require
to have a proper choice of
initial parameters. In the proposed HMRF-FCEM scheme, comb
ined strength of fuzzy
clustering approach as well as HMRF model are incorporated.
The result obtained by
the proposed FCEM algorithm in HMRF-FCEM scheme are compare
d with that of ex-
isting schemes and the results are quite comparable to the la
ter ones. The performance
of proposed algorithm could be successfully tested with an a
rbitrary set of initial model
parameters.
Both BHMRF-EM and HMRF-FCEM schemes could be validated for h
ealthy as
well as diseased brain MR images.

There are no comments on this title.

to post a comment.


Implemented and Maintained by Biju Patnaik Central Library.
For any Suggestions/Query Contact to library or Email: library@nitrkl.ac.in OR bpcl-cir@nitrkl.ac.in. Ph:91+6612462103
Website/OPAC best viewed in Mozilla Browser in 1366X768 Resolution.

Powered by Koha