12-12-2012, 05:11 PM
A Survey of Current Methods in Medical Image Segmentation
A Survey of Current.pdf (Size: 414.39 KB / Downloads: 103)
ABSTRACT:
Image segmentation plays a crucial role in many medical imaging applications by automating
or facilitating the delineation of anatomical structures and other regions of interest. We present herein
a critical appraisal of the current status of semi-automated and automated methods for the segmentation of
anatomical medical images. Current segmentation approaches are reviewed with an emphasis placed on
revealing the advantages and disadvantages of these methods for medical imaging applications. The use
of image segmentation in different imaging modalities is also described along with the difficulties encountered
in each modality. We conclude with a discussion on the future of image segmentation methods in
biomedical research.
Introduction
Diagnostic imaging is an invaluable tool in medicine today. Magnetic resonance imaging
(MRI), computed tomography (CT), digital mammography, and other imaging modalities
provide an effective means for noninvasively mapping the anatomy of a subject.
These technologies have greatly increased knowledge of normal and diseased anatomy
for medical research and are a critical component in diagnosis and treatment planning.
With the increasing size and number of medical images, the use of computers in facilitating
their processing and analysis has become necessary. In particular, computer
algorithms for the delineation of anatomical structures and other regions of interest are
a key component in assisting and automating specific radiological tasks. These algorithms,
called image segmentation algorithms, play a vital role in numerous biomedical
imaging applications such as the quantification of tissue volumes [98], diagnosis [176],
localization of pathology [208], study of anatomical structure [198], treatment planning
[90], partial volume correction of functional imaging data [128], and computerintegrated
surgery [6, 61].
Background
In this section we define terminology that will be used throughout and describe important
issues in the segmentation of medical images.
Definitions
An image is a collection of measurements in two-dimensional (2-D) or three-dimensional
(3-D) space. In medical images, these measurements or image intensities can be radiation
absorption in X-ray imaging, acoustic pressure in ultrasound, or RF signal amplitude
in MRI. If a single measurement is made at each location in the image, then the image
is called a scalar image. If more than one measurement is made (eg. dual-echo MRI),
the image is called a vector or multi-channel image. Images may be acquired in the
continuous domain such as on X-ray film, or in discrete space as in MRI. In 2-D discrete
images, the location of each measurement is called a pixel and in 3-D images, it is called
a voxel. For simplicity, we will often use the term “pixel” to refer to both the 2-D and
3-D cases.
Image Segmentation 3
When the constraint that regions be connected is removed, then determining the sets
Sk is called pixel classification and the sets themselves are called classes. Pixel classification
rather than classical segmentation is often a desirable goal in medical images,
particularly when disconnected regions belonging to the same tissue class need to be
identified. Determination of the total number of classes K in pixel classification can be
a difficult problem [97]. Often, the value of K is assumed to be known based on prior
knowledge of the anatomy being considered.
Labeling is the process of assigning a meaningful designation to each region or class
and can be performed separately from segmentation. It maps the numerical index k of
set Sk, to an anatomical designation. In medical imaging, the labels are often visually
obvious and can be determined upon inspection by a physician or technician. Computer
automated labeling is desirable when labels are not obvious and in automated processing
systems. A typical situation involving labeling occurs in digital mammography where
the image is segmented into distinct regions and the regions are subsequently labeled as
being healthy tissue or tumorous.
Dimensionality
Dimensionality refers to whether a segmentation method operates in a 2-D image domain
or a 3-D image domain. Methods that rely solely on image intensities are independent
of the image domain. However, certain methods such as deformable models, Markov
random fields, and region growing (described in Section 3), incorporate spatial information
and may therefore operate differently depending on the dimensionality of the image.
Generally, 2-D methods are applied to 2-D images and 3-D methods are applied to 3-D
images. In some cases, however, 2-D methods are applied sequentially to the slices of a
3-D image [7, 52, 103, 141]. This may arise because of practical reasons such as ease of
implementation, lower computational complexity, and reduced memory requirements.
In addition, certain structures are more easily defined along 2-D slices.
A unique situation that occurs in medical image segmentation is the delineation of
regions on a non-Euclidean domain, such as in brain cortex parcellation [148, 156]. This
is essentially segmentation on a surface of measurements. Because a surface is a 2-D
object folded in 3-D space, segmentation on a surface can not be treated as a standard
2-D or 3-D problem. The modeling of spatial characteristics along a surface is much
more difficult than in a standard imaging plane because of the irregular sampling used
by mesh representations and because of the need to compute geodesics [89]. This is an
emerging area of research and preliminary results have shown great promise.
Soft segmentation and partial volume effects
Segmentations that allow regions or classes to overlap are called soft segmentations.
Soft segmentations are important in medical imaging because of partial volume effects,
where multiple tissues contribute to a single pixel or voxel resulting in a blurring of
intensity across boundaries. Figure 1 illustrates how the sampling process can result in
partial volume effects, leading to ambiguities in structural definitions. In Figure 1b, it is
difficult to precisely determine the boundaries of the two objects. A hard segmentation
forces a decision of whether a pixel is inside or outside the object. Soft segmentations
on the other hand, retain more information from the original image by allowing for
uncertainty in the location of object boundaries. Note that the point spread function of
an imaging device can be larger than the spatial extent of a single pixel or voxel. Thus,
partial volume effects can cause boundaries to be blurred across significant portions of
an image.
Continuous or discrete segmentations
Nearly all medical images used for image segmentation are represented as discrete samples
on a uniform grid. Segmentation methods typically operate on the same discrete grid
as the image. However, certain methods such as deformable models (see Section 3.7) are
capable of operating in the continuous spatial domain, thereby providing the potential
for subpixel accuracy in delineating structures. Subpixel accuracy is desirable particularly
when the resolution of the image is on the same order of magnitude as the structure
of interest.
Segmentation on the continuous domain is not equivalent to partial volume estimation
or other soft segmentation methods. Partial volume estimation methods merely provide
the fraction of a structure which is present in a voxel. This may be sufficient for quantification
purposes but not in situations where precise localization is required, such as for
tumors in surgical or radiotherapy planning. Continuous segmentation methods actually
reconstruct how a structure passes through a voxel. Although continuous segmentation
methods have subpixel or subvoxel resolution, their precision and accuracy are still dependent
on the resolution of the original data. Furthermore, this level of precision can
be difficult to validate on real data.
Validation
In order to quantify the performance of a segmentation method, validation experiments
are necessary. Validation is typically performed using one of two different types of truth
models. The most straightforward approach to validation is by comparing the automated
segmentations with manually obtained segmentations (cf. [199, 186]). This approach,
besides suffering from the drawbacks outlined in the previous section, does not guarantee
a perfect truth model since an operator’s performance can also be flawed. The other
common approach to validating segmentation methods is through the use of physical
phantoms [102] or computational phantoms [36]. Physical phantoms provide an accurate
depiction of the image acquisition process but typically do not present a realistic
representation of anatomy.