13 CHAPTER 2 A REVIEW ON BREAST ABNORMALITY SEGMENTATION AND CLASSIFICATION TECHNIQUES Segmentation or abnormality detection is the initial step in mammographic Computer Aided Detection (CAD) system. The review on different approaches to the segmentation and classification of mammographic masses and microcalcifications are described in this chapter. It also describes main features and differences of these approaches. The key objective is to point out the advantages and disadvantages of these approaches. 2.1 INTRODUCTION A segmentation algorithm is used to detect region of interest, usually the part of breast or a specific kind of abnormalities like microcalcifications or masses. It is generally known that the detection of masses is technically difficult, because masses can be simulated or obscured by normal breast parenchyma. Moreover, there is an outsized variability in these lesions, which is reflected in the morphology features (shape and size of the lesions). Variations exist in large number of features that have been used to detect and classify them. Microcalcifications are considered to be important indicators of breast cancer. However, its interpretation is very difficult and 10% - 30% of breast microcalcifications are missed during routine screening (Bird et al 1992, Burhenne et al 1994). It appears as tiny objects which can be described
33
Embed
CHAPTER 2 A REVIEW ON BREAST ABNORMALITY …shodhganga.inflibnet.ac.in/bitstream/10603/26433/7/07_chapter 2.pdf · A REVIEW ON BREAST ABNORMALITY SEGMENTATION AND CLASSIFICATION TECHNIQUES
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
13
CHAPTER 2
A REVIEW ON BREAST ABNORMALITY
SEGMENTATION AND CLASSIFICATION TECHNIQUES
Segmentation or abnormality detection is the initial step in
mammographic Computer Aided Detection (CAD) system. The review on
different approaches to the segmentation and classification of mammographic
masses and microcalcifications are described in this chapter. It also describes
main features and differences of these approaches. The key objective is to
point out the advantages and disadvantages of these approaches.
2.1 INTRODUCTION
A segmentation algorithm is used to detect region of interest,
usually the part of breast or a specific kind of abnormalities like
microcalcifications or masses. It is generally known that the detection of
masses is technically difficult, because masses can be simulated or obscured
by normal breast parenchyma. Moreover, there is an outsized variability in
these lesions, which is reflected in the morphology features (shape and size of
the lesions). Variations exist in large number of features that have been used
to detect and classify them.
Microcalcifications are considered to be important indicators of
breast cancer. However, its interpretation is very difficult and 10% - 30% of
breast microcalcifications are missed during routine screening (Bird et al
1992, Burhenne et al 1994). It appears as tiny objects which can be described
14
as granular, linear, or irregular on mammograms. Although they have higher
inherent attenuation properties, they cannot be distinguished from the high
frequency noise because of their smaller size. The microcalcifications
typically range in size from 0.1 to 1.0 mm (Thor Ole Gulsrud and John Hakon
Husoy, 2001). Microcalcifications often appear in a heterogeneous
background describing the structure of the breast tissue. Some elements of the
background, like dense tissue, could also be brighter than the
microcalcifications in the fatty part of the breast. The Regions of Interest
(ROI’s) may be of low contrast. The intensity difference between suspicious
areas and their surrounding tissues can be quite slim. Dense tissues may be
easily misinterpreted as calcifications yielding a high False Positive (FP) rate,
which is a major problem with most of the algorithms.
In this chapter, lesion segmentation and classification algorithms
are reviewed in detail. In section 2.2, the algorithms which look for
segmentations of masses and microcalcifications using mammographic
images are described. Section 2.3 describes the literature study on
mammographic classification. The evaluation methodology is given in section
2.4. The summary of reviews is given in section 2.5.
2.2 BREAST ABNORMALITY SEGMENTATION
Segmentation of breast abnormality relies on the fact that pixels
inside a mass or microcalcifications have different characteristics from the
other pixels within the breast area. The characteristics used can be gray level
values, texture features, shape features or morphological features of the
lesions. The outcome of segmentation of image is a set of segments that
collectively cover the entire image, or a set of edges extracted from the image.
Each one of the pixels in a region is related with respect to some
characteristics or computed property, such as color, intensity, or texture.
15
Neighboring regions are considerably different with respect to the
characteristic(s).
In computer vision terminology, segmentation techniques can be
divided into unsupervised and supervised approaches. Supervised
segmentation, also known as model-based segmentation, relies on prior
knowledge about the object and background regions to be segmented. The
prior information is used to determine if specific regions are present within an
image or not. Alternatively, unsupervised segmentation partitions an image
into a set of regions which are distinct and uniform with respect to specific
properties, such as grey features, texture or colour features. Classical
methods used to solving unsupervised segmentation are divided in three major
groups (Fu and Mui, 1981). These are region based methods, contour based
methods and clustering methods. The rows of Table 2.1 shows the reviewed
works arranged according to their unsupervised segmentations. A detailed
description of the methods in each category is given in subsequent sections.
2.2.1 Region Based Methods
The main goal of segmentation is to partition an image into regions.
Region based segmentation relies on the principle of homogeneity, which means
there should be at least one feature which remains uniform for all pixels within a
region. The basic formulation for Region-based segmentation is:
a) 1
n
i iR R (2.1)
b) is a connected region, i=1,2,...,niR
c) iR for all i=1,2,...,njR (2.2)
d) ( ) for i=1,2,...,niP R TRUE (2.3)
e) i i jP(R ) for any adjacent region R and RjR FALSE (2.4)
16
where,
R – Region, n – number of pixels, - null set, ( )iP R - logical predicate.
More than 30 years have passed, since Zucker reviewed region
growing algorithms (Zucker 1976). Region growing is based on the
propagation of an initial seed point according to a specific homogeneity
criterion, iteratively increasing the size of the region. Since then, region
growing has seen a number of improvements, primarily due to the integration
of boundary information in the algorithm. Region growing algorithms have
been widely used in mammographic mass segmentation with the aim of
extracting the potential lesion from its background. Since early nineties,
researchers from the University of Chicago investigated the introduction of
shape information into the homogeneity criterion.
William Mark Morrow et al (1992) have developed an adaptive
method for enhancing the contrast of mammographic features of varying size
and shape. The method uses each pixel in the image as a seed to grow a
region. The size and shape of the region adapt to local image gray-level
variations, corresponding to an image feature. The contrast of every region is
calculated with respect to its individual background. Contrast is, then,
improved by applying an empirical transformation based on each region’s
seed pixel value, its dissimilarity, and its background.
With the aim to integrate the radiologist’s experiences, Huo et al
(1995) have developed a semi-automatic region growing approach. In this
approach, the growing step was automatically computed after a radiologist
had manually placed the seed point. Later, Matthew Kupinski and Maryellen
Giger (1998) have developed radial gradient index method and a probabilistic
method for segmentation of lesions. These methods are seeded segmentation
methods. In both methods, a series of image partitions is created using gray-
level information as well as prior knowledge of the shape of typical mass
lesions.
17
Naga Mudigonda et al (2001) have developed a method for the
detection of masses in mammographic images. The method employs Gaussian
smoothing and sub sampling operations as initial processing steps. The mass
portions are segmented by establishing intensity links from the central
portions of masses into the surrounding areas. Chu et al (2002) have
proposed a region growing approach that represented as a growing tree whose
root is the selected seed. Active leaves are removed in the connection area
between adjacent regions to avoid merging adjacent structures. The authors
affirmed that this graph-based segmentation more closely matches
radiologists’ outlines of masses.
Jiang et al (2007) have developed a genetic algorithm to
automatically classify and detect microcalcification clusters. The genetic
algorithm technique is characterized by transforming input images into a
feature domain. Here, every pixel is represented by its mean and standard
deviation inside a surrounding window of size 9×9 pixel. In the feature
domain, chromosomes are constructed to populate the initial generation. The
features are extracted to enable the genetic algorithm to search for optimized
classification and detection of microcalcification.
Peter Filev et al (2008) have developed a computerized regional
registration and characterization system for analysis of microcalcification
clusters. The system consists of two stages. In the initial stage, a regional
registration procedure is used to identify the local area that may contain the
cluster. A search program is used to detect cluster candidates within the local
area. In the second stage, a temporal classifier is used to classify the region.
Alfonso et al (2009) have developed two methods called dynamic-
programming-based method and a constrained region-growing method to
segment the mass contours. The simplified versions of these contours were
employed to extract a set of six features designed for characterization of mass
18
margins (contrast between foreground and background region, two measures
of the fuzziness of mass margins, coefficient of deviation of edge strength, a
measure of spiculation based on edge-signature information and a measure of
spiculation based on relative gradient orientation). Three accepted classifiers
(Fisher's linear discriminant, Bayesian classifier and a support vector
machine) were used to predict the lesions.
Claudio Marrocco et al (2010) have developed a novel system for
detection of clusters of microcalcifications. The mammogram first extracts the
elementary homogeneous regions of interest in this system. An analysis of
such regions is then performed by means of a two-stage, coarse-to-fine
categorization based on both heuristic rules and classifier combination. Giulia
Rabottino et al (2011) have developed a region growing technique to segment
the lesion region. The shape and texture features are extracted, and fuzzy
classifier is used to classify the lesion regions.
2.2.2 Contour Based Methods
Image segmentation techniques based on contour based method
have been used in the early work of Roberts (1965). Contour-based
approaches usually start with edge detection, followed by a linking process
that seeks to exploit curvilinear continuity. However, the identification of
regions based on the edge detection is far from trivial. The algorithms for
edge detection do not usually possess the ability of the human vision system
to complete interrupted edges. Therefore, sometimes edges which are not the
transition from one region to another are detected. The properly detected
edges often present gaps at places where the transitions between regions are
not abrupt enough. Hence, detected edges might not essentially form a set of
closed connected curves that surround distinct regions.
19
The Table 2.1 shows number of publications is trying to detect
masses in mammogram using contour based methods. The algorithms for
finding edges are based on filtering the image in order to enhance relevant
edges prior to the detection stage. The location of edges, in Petrick et al (1996),
is based on a Gaussian–Laplacian edge detector, after which the image is
enhanced by an adaptive density-weighted contrast enhancement filter.
Judy Kilday et al (1993) have developed an interactive
segmentation procedure to identify the tumor boundary using a thresholding
technique. The several features are extracted based on the gross and fine
shape describing properties of the tumor boundaries. Joachim Dengler et al
(1993) have used a two-stage algorithm for spot detection and shape
extraction. The topology and the number of the spots are determined by using
weighted difference of Gaussians filter and the shape by means of
morphological filters. Parr et al (1994) used Gabor filters to locate the
spicules of stellated lesions.
Kobatake and Yoshinaga (1996) have described an approach, which
starts with a sub-image containing a possible mass lesion. It looks for spicules
using gradient information in three steps: First, the morphological line-
skeletons are extracted in order to detect long and thin anatomical structures
(like spicules). Second, a modified hough transform is used to extract lines
passing near the centre of the mass, and finally the algorithm automatically
select objects based on the number of line skeletons that satisfy the second step.
Fauci et al (2005) and Cascio et al (2006) have looked for the
contours of the mass using an iterative algorithm. At each local maxima a
threshold was selected which is used to draw an intensity contour. The
threshold value is based on user interaction and histogram information.
Consequently, the area of the selected region is refined by adjusting the
threshold. Stelios Halkiotis et al (2007) have used the mathematical
20
morphology tools for the extraction of microcalcifications even if the
microcalcifications are located on a non-uniform background.