Systems, Man & Cybernetics - April 2017 - 19

such as tumor segmentation, while the latter
creates vivid two-dimensional representations of three-dimensional volumetric data. Our research projects have
focused on the sparse and compressible
properties of signals in some representation systems with the ge neral objective of fully utilizing these properties
for the development of effective and efficient algorithms for medical image analysis and volume visualization. In this
article, we introduce one of our research
projects, which aims to address the signiﬁcant but challenging task of multilabel
brain tumor segmentation. In this work, we
proposed a superpixel-based framework for
this speciﬁc task using structured kernel
sparse representation.

considered a challenging topic, especially for the multilabel case.
Researchers have put much effort into the study of brain
tumor segmentation for decades. Published approaches
can be roughly categorized into either generative methods
or discriminative methods [3]. Generative methods explicitly model the anatomy and statistics of brain tissues and
usually use these models in the expectation-maximization
algorithm [4], [5]. Generative methods generalize well on
unseen images, but they suffer from difficulties in modeling
the prior knowledge of brain tissues and elaborate nonrigid
registration. Discriminative methods directly learn the features of task-relevant brain tissues from given training sets
for the segmentation task [6], [7]. Although discriminative
methods avoid the difficulties in modeling and registration,
they are sensitive to the quantity and quality of the available training samples. The success of sparse coding and
dictionary learning has been demonstrated in various
vision problems, including image segmentation. In [8], we
proposed a superpixel-based framework aimed at multilabel brain tumor segmentation using structured kernel
sparse representation.

Background
Uncontrollable cell proliferation in the brain is referred
to as a brain tumor.
tumor Even though a brain tumor is not a common case, with a prevalence of less than 0.1% in Western
The Proposed Framework
countries, the high mortality it causes is eye-catching [1].
An overview of the proposed framework is shown in Figure 1.
The goal of brain tumor segmentation is to separate tumor
There are three main components: multichannel superpixeltissues from nontumor regions and partition them into
level feature extraction and fusion, kernel sparse representaspatially contiguous regions according to predeﬁned critetion, and segmentation. In our approach, the superpixel,
ria [2]. Magnetic resonance (MR) imaging is the most comwhich refers to a group of pixels with similar perceptual
mon technology in brain imaging due to its advantages in
meaning, is used as the basic processing unit for efficiency
terms of safety, tissue contrast, and fewer artifacts comand compact representation. Sparse coding and dictionary
pared to other modalities, such as computed tomography.
learning, both part of kernel sparse representation, are impleThis emphasizes the signiﬁcance of developing effective
mented in a high-dimensional feature space F with the
and reproducible methods for brain tumor segmentation
help of the kernel trick, which simpliﬁes the complicated, or
based on MR images. When dealing with large medical
even intractable, inner product calimage data sets, semi-automated
culation in F to the calculation of
and automated approaches, which
known kernel functions.
require very little human intervenToday, the use of
MR i mag i ng i mproves ou r
tion, are preferable to manual segknowledge
of brain tumors by promentation, which is tedious and
medical images
viding multichannel information
time-consuming.
is often critical
[3], such as T1-weighted (T1), T2Nevertheless, brain tumor segweighted (T2), contrast-enhanced
mentation is never an easy task.
for diagnosis and
T1-weighted (T1c), and fluid-attenBrain tumors vary significantly in
treatment planning.
uated inversion recovery (FLAIR).
size, shape, and location from case
Due to the fact that T1c images
to case. In addition, a common situusually have higher spatial resoluation in MR images is that tumor
tion and present a clearer display of brain tumor strucand nontumor tissues overlap in their gray-level intensities.
ture compared to other channels, they are used as the
These present a huge challenge in semi-automated and
reference image to generate superpixels. These superpixfully automated brain tumor segmentation because the use
el regions are then applied to other channels for the
of strong priors is impossible. Brain tumor segmentation is
Ap ri l 2017

IEEE SyStEmS, man, & CybErnEtICS magazInE

Table of Contents for the Digital Edition of Systems, Man & Cybernetics - April 2017