A VESSEL SEGMENTATION METHOD FOR MRA DATA
BASED ON PROBABILISTIC MIXTURE MODEL
Pei Lu*, Cheng Wang*, Shoujun Zhou
†
* Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, email: pei.lu@siat.ac.cn
† Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, email: shoujz@163.com
Keywords: Vessel segmentation, mixture model, histogram
analysis, EM algorithm, MRF.
Abstract
Accurate segmentation and visualization of cerebral vessels
have important significance to the diagnosis and treatment of
relevant brain diseases. However, some segmentation
algorithms are only competent for the standard data. In this
paper, a segmentation method based on probabilistic mixture
model is proposed to solve clinical problems. Through
histogram analysis of the magnetic resonance angiography
(MRA) data offered by Guangzhou General Hospital of the
Chinese PLA, a mixture model formed by six probabilistic
distributions (one Exponential, one Rayleigh, and four
Normal distributions) was built to fit the histogram curve.
Least squares and expectation maximization (EM) algorithms
have been used for parameters estimation. At last, the
segmentation was enhanced by maximum a posteriori
probability (MAP) and Markov random field (MRF)
algorithm. The effectiveness of the proposed method has been
validated by segmentation tests on a series of clinical MRA
data with good performance.
1 Introduction
Brain diseases that are caused mostly by cerebrovascular
accidents occupy a large part of human diseases. Imaging
techniques can give a 3D visualization of brain structures.
Nowadays, medical imaging equipment with high resolution
such as computed tomography angiography(CTA) and
magnetic resonance angiography(MRA) are widely used in
clinical practice. However, accurate extraction of cerebral
blood vessels from the imaging data is especially helpful in
computer aided diagnosis and treatment.
Many segmentation methods have been proposed for the
extraction of cerebral vessels. Among these methods,
statistical models have drawn a lot of attentions[1-6]. In 1999,
a statistically based segmentation algorithm for extracting 3D
cerebral vessels from MRA data has been established by
Wilson and Noble[1]. They divided the MRA data histogram
into three regions, modeled by two normal distributions and
one uniform distribution. Hassouna improved the mixture
model with a Rayleigh distribution and three normal
distributions which provide an accurate fitting[3]. A statistical
method based on MAP–MRF with multi-pattern
neighborhood system was proposed by Shoujun Zhou for
segmenting fine cerebral vessels with complicated context[6].
However, the above mixture model algorithms are not fit for
MRA data offered by some hospitals.
Aiming at extracting cerebral vessels from MRA data gained
from Guangzhou General Hospital of the Chinese PLA, a new
mixture model consists of six sections is proposed in this
paper. Based on histogram analysis, the MRA data was firstly
divided into two components. The first component was
modeled by an exponential distribution, while the last five
parts of the second component were modeled like described
in Hassouna’s method[3]. Least square and expectation
maximization (EM) algorithms were used for parameters
estimation of the mixture model. After that, MAP–MRF
algorithm was used to enhance the segmentation quality of
the vessels from the background noise.
This paper is organized as follows. In section 2, the mixture
model is built firstly based on the histogram analysis, and
then the parameters are estimated with least square and EM
algorithms in details. Besides, the final segmentation
confirmed with MAP-MRF method is discussed. Section 3
presents the vessel segmentation results, and the conclusions
are drawn in Section 4.
2 The proposed method
The proposed vessel segmentation method consists of three
parts. Firstly, the mixture model is built according to the
analysis of the histogram. Secondly, least squares and EM
algorithms are used to estimate the parameters of the mixture
model. At last, MAP-MRF is presented to extract the vessels
from the background voxels.
2.1 The mixture model
Fig. 1. A slice of the 3D MRA data.