Segmentation of Brain Magnetic Resonance Angiography Images Based on
MAP-MRF with Multi-pattern Neighborhood System
Shoujun ZHOU
1
, Wufan CHEN
2*
, Fucang JIA
1
, Qingmao HU
1*
, Yaoqin XIE
1
, Mingyang CHEN
1
Shang PENG
1
1. Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055.
E-mail: sj.zhou@siat.ac.cn
2. School of Biomedical Engineering, Southern Medical University, Guangzhou 510515.
E-mail: chenwf@fimmu.com
Abstract: Existing maximum a posteriori probability and Markov random field (MRF) models have limitations
associated with that the ordinary neighborhood system being unable to differentiate subtle changes due to
several-to-one correspondence within the neighborhood. Aiming at overcoming the limitations and applications to
segmentation of cerebral vessels from magnetic resonance angiography images, we proposed a multi-pattern
neighborhood system and corresponding energy equation to enable the MRF model for segmenting fine cerebral
vessels with complicated context. In the implementation, a candidate space of cerebral vessels was employed to
reduce the time-consumption, which was based on a threshold of the response to multi-scale filtering. A set of
phantoms simulating segmentation challenges of vessels have been devised to quantitatively validate the algorithm.
In addition, ten three-dimensional clinical datasets have been used to validate the algorithm qualitatively. It has
been shown that the proposed method could yield smaller error and improve the spatial resolution of MRF model.
Key Words: Cerebrovascular segmentation, Markov random field, Markov neighborhood system, magnetic
resonance angiography
1 INTRODUCTION
Every year, many people suffer a cerebrovascular
accident, usually a stroke. Because cerebral vessels are
complex three-dimensional (3D) anatomical structure,
accurate segmentation, especially for small and dim
blood vessels, is challenging [1].
A variety of methods have been proposed for
segmenting cerebral vessels from magnetic resonance
angiography (MRA) images. These methods can be
roughly classified as: multi-scale filtering [2], deformable
models [3], statistical models [4,5,6] and hybrid methods
[7,8]. Multi-scale filtering might fail when vessels are
close in space or when vessel diameters change abruptly.
Deformable model approaches might be problematic to
segment vessels from low-contrast MRA images, and
heavily depend on model parameters as well as
assumption of typical vessel shapes [3]. Statistical
approaches process MRA images with multi- modalities
in each region-of-interest (such as cerebral vessels and
brain tissues), and each modality is associated with a
particular mode of marginal probability distribution
(MPD). Wilson and Noble [4] modeled the MPD and
used expectation-maximization (EM) algorithm to
estimate parameters. For the statistical classification, the
challenges are to deal with classes having overlapped
intensities, high intra-class variances, and close class
means. Efforts have been made to combine the spatial
shape with the class probability distributions [8] to form
the so-called hybrid methods. Recently, the spatial
contextual information was incorporated through 3D
Markov random field (MRF) [5,6].
This paper considered the large amount of computation
of 3D MRF, a preprocessing step was used to acquire an
initial vessel candidate space which was a reduced
computation space with its size far smaller than that of the
original brain data space. Then, a multi-pattern
neighborhood system (MP-NBS) was defined, and the
clique energy function of MRF was proposed to avoid the
several-to-one correspondence problem. During
MRF-based segmentation, the low level process, i.e.,
maximum likelihood estimation (MLE), used finite
mixture (FM) models [4] and was applied to initialize
MRF; while the high level process, i.e., the MP-NBS and
corresponding clique energy function, was devised to
detect the vessels. Finally, the Iterated Conditional
Modes (ICM) [9] was used to iteratively solve for the
MAP-MRF model in the candidate vessel space so as to
reduce the computation time. In the experiment, the
proposed method was validated using phantom data and
ten clinical brain MRA data
2 METHOD
2.1 Preprocessing and Initialization
In this subsection, calculation of MPD parameters of
TOF-MRA dataset was explored to establish the FM model.
Hassouna et al. [5] divided the TOF-MRA data intensity
histogram into three regions. In ascending order of
intensities, tissues were (1) cerebrospinal fluid, bone, and
air; (2) brain tissues, including both the grey and white
matter, and parts of the eyes; and (3) subcutaneous fat and
arteries. Similar to their work, we used one Rayleigh (with
parameter R ) and two Gaussian distributions (with
____________________________________
978-1-4673-5887-3/13/$31.00 ©2013 IEEE