enhancememt of mammogram images

1. 1
MAMMOGRAM ENHANCEMENT AND SEGMENTATION METHODS:
CLASSIFICATION, ANALYSIS, AND EVALUATION
MARIAM BILTAWI1
, NIJAD AL-NAJDAWI2
, SARA TEDMORI3
,
1
Department of Computer Science, The King Hussein School for Information Technology, Princess Sumaya University for
Technology, Al-Jubaiha 11941, Jordan, 00962(6)5359949 Ext 276, maryam@psut.edu.jo
2
Department of Computer Science, Prince Abdullah Ben Ghazi Faculty of Information Technology, Al-Balqa Applied
University, Al-Salt 19117, Jordan, 00962(5)3552519 Ext 3016, n.al-najdawi@bau.edu.jo
3
Department of Computer Science, The King Hussein School for Information Technology, Princess Sumaya University for
Technology, Al-Jubaiha 11941, Jordan, 00962(6)5359949 Ext 238, s.tedmori@psut.edu.jo
Abstract
Breast cancer is the leading cause of deaths among female cancer patients. Mammography is the most effective technique for
breast cancer screening and detection of abnormalities. However, early detection of breast cancer is dependent on both the
radiologist’s ability to read mammograms and the quality of mammogram images. The aim of this paper is to conduct a
comprehensive survey of existing mammogram enhancement and segmentation techniques. Each method is classified, analyzed,
and compared against other approaches. To examine the accuracy of the mammogram enhancement and segmentation
techniques, the sensitivity and specificity of the approaches is presented and compared where applicable. Finally, this research
provides taxonomy for the available approaches and highlights the best available enhancement and segmentation methods.
Keywords: Mammogram enhancement, Mammogram segmentation, Breast mass detection, Image Calcification, detection
1. INTRODUCTION
Globally, breast cancer is ranked first among the
leading causes of cancer affecting females. Statistics have
shown that 1 out of 10 women are affected by breast cancer
in their lifetime. There are several ways in which breast
cancer can be diagnosed, including breast self examination
(BSE), clinical breast exam (CBE), imaging or
mammography, and surgery. A mammogram is the most
effective technique for breast cancer screening and early
detection of masses or abnormalities; it can detect 85 to 90
percent of all breast cancers. The most common said
abnormalities that indicate breast cancer are masses and
calcifications.
Depending on its shape, a mass screened on a
mammogram can be either benign or malignant. Usually
benign tumors have round or oval shapes, while malignant
tumors have a partially rounded shape with a spiked or
irregular outline. Noncancerous or benign tumors include
cysts, fibro adenomas, and breast hematomas. A cancerous
or malignant tumor in the breast is a mass of breast tissue
that grows in an abnormal and uncontrolled way [8]. The
malignant mass will appear whiter than any tissue
surrounding it. Calcifications (both macro-calcification and
micro-calcification), the second abnormality that can be
seen on mammogram images, are most of the time not
malignant and not a sign of cancer. Successful diagnosis in
mammography is dependent on detecting cancer in its
earliest and most treatable stage. The challenge is to employ
computer aided detection (CAD) techniques for the purpose
of assisting radiologists in the early detection of cancer, by
processing and analyzing mammogram images. The
majority of the proposed cancer detection techniques in
literature share the common steps of image enhancement,
segmentation, quantification, registration, visualization [1],
[44]. These techniques can be differentiated by the varying
algorithms employed at each step. One of the challenges
faced by the current mammogram image detection
techniques lies in the difficulty of analyzing dense tissues.
This difficulty can be attributed to the breast region which
appears white in the mammogram images making masses
and specifically micro-calcification highly invisible
intermixed with the background tissue. Mammogram image
enhancement is the process of manipulating mammogram
images to increase their contrast and decrease the noise
present in order to aid radiologists in the detection of
abnormalities. Mammogram image segmentation is the
process of partitioning mutually homogeneous regions into
meaningful regions of interest.
A few papers have been written surveying the
techniques used in cancer detection of mammogram images.
Cheng et al., [14], summarized the enhancement,
segmentation, and detection techniques of mammogram
images. The authors categorized the micro-calcification
enhancement techniques into three categories: conventional
enhancement techniques, region-based enhancement, and
feature based enhancement. The conventional enhancement
techniques are in turn divided into contrast stretching,
histogram equalization, convolution mask enhancement, and
fixed and adaptive neighboring enhancement. Oliver et al.,
[39] reviewed the existing techniques for cancerous mass
detection and segmentation. In their paper, the authors
categorized mammogram segmentation techniques into mass
detection using a single view and mass detection using
multiple views. The mass detection using single view
segmentation in turn is divided into four categories: (1)
model-based methods, (2) region-based methods, (3)
contour-based methods, and (4) clustering methods. Only
477
The 13th International Arab Conference on Information Technology ACIT'2012 Dec.10-13
ISSN : 1812-0857

2. 2
the model based methods are considered supervised
segmentation methods, while the remaining three methods
are considered unsupervised segmentation methods. The
image segmentation techniques using multiple views are
divided into three categories: left and right mammograms,
two mammographic views (CC and MLO) of the same
breast, and same view mammograms taken at different
times. Bandyopadhyay [6] surveyed some commonly used
mass segmentation methods. Raba et al., [42] reviewed
breast region segmentation methods developed and
proposed in the 80’s, 90’s and 2000. The techniques that
were surveyed included: histogram based techniques,
gradient based techniques, polynomial modeling based
techniques, active contour based techniques, and classifiers
based techniques. In this paper, the authors’ review the
algorithms that have been proposed in the literature to
enhance and segment mammogram images that contain both
masses and micro-calcifications.
The goal of this paper is to provide a
comprehensive review, comparison, and analysis of the
available mammogram enhancement and segmentation
algorithms. The rest of this paper is organized as follows:
section 2 presents details of mammogram image
enhancement techniques, section 3 presents the
mammogram image segmentation techniques, section 4
evaluates the proposed approaches. Finally, section 5
presents the conclusion and the future work.
2. MAMMOGRAM IMAGE
ENHANCEMENT TECHNIQUES
Mammogram image enhancement is the process of
manipulating mammogram images to increase their contrast
and decrease the noise present in order to aid radiologists in
the detection of abnormalities. The methods used to
manipulate mammogram images can be categorized into
four main categories; the conventional enhancement
techniques, the region-based enhancement techniques, the
feature-based enhancement techniques, and the fuzzy
enhancement techniques as shown in Figure 1. Conventional
enhancing techniques are fixed neighborhood techniques
and they are used to modify images based on global
properties. Although region-based methods are used in
segmentation, they are also used for enhancing the contrast
of mammogram features according to the surroundings. On
the other hand, feature based enhancement methods are
those methods that are based on wavelet domain
enhancement. And the fuzzy enhancement techniques are
methods that apply fuzzy operators and properties to
enhance mammogram features. Table 1 shows what each of
the four mammogram image enhancement categories is
primarily used for.
Enhancement
Category
Used for the
enhancement of
masses
Used for the
enhancement of
calcifications
Conventional
Enhancement
√ x
Region-based
Enhancement
√ x
Feature-based
Enhancement
√ √
Fuzzy Enhancement √ √
Table 1: Main usage of mammogram enhancement categories
2.1 CONVENTIONAL ENHANCEMENT
TECHNIQUES
The conventional enhancement techniques are
mostly used to enhance masses in mammogram images; as
an example, Bovis and Singh [9] and Antonie et al., [3] used
histogram equalization to enhance the mammogram images
before segmentation and mass detection. However, Schiabel
et al., [47] used the histogram equalization technique
accompanied with other techniques and as a part of a pre-
processing step for mammogram enhancement. Whereas,
Pisano et al., [40] used the contrast limited adaptive
histogram equalization (CLAHE) in order to determine
whether such a method can improve the detection of
stimulated speculations in dense mammograms.
Hemminger et al., [26] compared between contrast-
limited adaptive histogram equalization (CLAHE) and
histogram-based intensity windowing (HIW) in order to
determine which of them outperforms the other in the
detection of simulated masses in dense mammograms. Yu
and Bajaj [64] described a fast approach based on localized
contrast manipulation to enhance image contrast. This
method is based on fast computation of local minimum,
maximum, and average maps using a propagation scheme.
Kom et al., [29] presented a linear transformation filter for
mammogram enhancement. Their algorithm modifies the
local contrast of each pixel according to two linear functions
and two constant values.
2.2 REGION-BASED ENHANCEMENT
Similar to the conventional enhancement methods,
the region-based enhancement methods are mostly used for
the enhancement of masses. An example of this is the work
conducted by Dominguez and Nandi [23] who presented an
enhancement algorithm as a part of an automatic detection
of mammogram masses method. Sampat and Bovik [45]
propose a filtering algorithm that enhances speculations (i.e.
linear features of masses) in mammograms as a part of a
speculated mass detection technique. of the image is
computed to obtain the enhanced image.
2.3 FEATURE BASED ENHANCEMENT:
Figure 1: Categorizations of mammogram enhancement techniques
478

3. 3
Feature based enhancement methods can be used to
enhance both masses and micro-calcifications. Gagnon et
al., [24] proposed a simple multi-scale sharpening
enhancement algorithm based on the hidden zero-crossing
property of the complex symmetric Daubechies wavelets.
The algorithm was tested using low contrast digitized
mammograms. Chang and Laine [12] presented an
enhancement algorithm based on over-complete multi-scale
wavelet analysis. Dabour [19] introduced an algorithm
based on wavelet analysis and mathematical morphology for
digital mammograms enhancement. The authors tested this
algorithm on several mammograms from the MIAS
database. The algorithm was also compared with various
algorithms and the experimental results and showed a better
contrast improvement index. Rodz et al., [4] used the
wavelet-based sharpening algorithm to enhance the contrast
of mammogram images. Laine et al., [30] introduced a
method for mammographic feature analysis by multi-
resolution representations of the dyadic wavelet transform.
Scharcanski and Jung [46] described an approach for noise
suppression and enhancement of mammogram images and
that can be effective in screening dense regions of the
mammograms. Stefanou et al., [50] compared between the
adaptive enhancement algorithm and the typical method of
enhancement.
2.4 FUZZY ENHANCEMENT METHODS
Fuzzy enhancement methods can be used to
enhance masses and micro-calcifications, and as an example
of enhancing mammograms containing masses, Singh and
Al-Mansoori [49] compared between fuzzy enhancement
techniques and histogram equalization. Mohanalin et al.,
[35] presented fuzzy algorithm based on Normalized Tsallis
entropy to enhance the contrast of micro-classifications in
mammograms. Jiang et al., [28] described a combined
approach of fuzzy logic and structure tensor, to enhance
micro-calcifications in digital mammograms. Cheng and Xu
[16] proposed an adaptive fuzzy logic contrast enhancement
method to enhance mammogram features. Table 2
summarizes the mammogram enhancement methods
previously reviewed, specifying the year of the research,
enhancement category, database used, and the number of
mammograms used for testing.
Author
Enhancement
category
Used
database
used
mammogram
s
Yu and Bajaj [64] Conventional N/A N/A
Kom et al [29] Conventional YGOPH 61
Dominguez and
Nandi [23]
Region-based
Mini-
MIAS
N/A
Sampat and Bovik
[45]
Region-based DDSM N/A
Gagnon et al [24] Feature-based N/A N/A
Dabour [19] Feature-based MIAS N/A
Scharacanski and
Jung [46]
Feature-based MIAS N/A
Mohanalin et al
[35]
Fuzzy
MIAS
UCSF
50
197
Jiang et al [28] Fuzzy DDSM 2000
Cheng and
Huijuan Xu [16]
Fuzzy N/A N/A
Table 2: Results of mammogram enhancement techniques
3. MAMMOGRAM IMAGE
SEGMENTATION TECHNIQUES
Mammogram image segmentation is the process of
partitioning mutually homogeneous regions of a
mammogram image into meaningful regions of interest. The
algorithms used for segmentation can be categorized into
two distinct categories according to the regions to be
segmented; breast region segmentation and region of interest
(ROI) segmentation. Breast region segmentation is the
process of splitting the mammogram image into a breast
region and a background in order to focus and limit the
search for abnormalities on the breast region without the
effect of the background on the results resulting in better
detection. On the other hand, region of interest segmentation
is the process of segmenting the suspicious regions to be
analyzed for abnormalities.
Figure 2: Categorizations of mammogram image segmentation techniques
479

4. 4
While segmentation using multiple views can be
categorized into left and right mammograms, two
mammographic views (cranio-caudal (CC) and medio lateral
oblique (MLO)) of the same breast, and same view
mammograms taken at different times. Unsupervised
segmentation using a single view can in turn be categorized
into six classes, region-based segmentation, contour-based
segmentation, clustering segmentation, pseudo-color
segmentation, graph segmentation, and variant-feature
transformation. Figure2 illustrates this categorization.
3.1 BREAST REGION SEGMENTATION
Breast segmentation techniques set the focus of the
search for abnormalities on the region of the breast
excluding its background. The techniques used for
segmenting are similar to those used in the regions of
interest segmentation and can also be categorized with the
same perspectives though it’s not the interest of this paper.
As an example of approaches that can be listed under the
clustering segmentation method is the novel approach
proposed by Shahedi et al., [48] for breast region
segmentation based on local threshold. The Ojala et al., [38]
which is based on histogram thresholding, morphological
filtering, and contour modeling. This approach is applied by
Wang et al., [59] as the first step in their automatic
framework to identify differences between corresponding
mammographic images. Raba et al., [42] proposed an
automatic technique for segmenting a digital mammogram
into a breast region and a background based on a "two-
phase" approach. Bovis and Singh [9]applied the technique
proposed by Chandrasekhar and Yttikiouzel [12] adding to it
an additional step, in order to segment the breast region
from its background. Wirth and Stapinski [61] described an
approach that automatically segments the breast region and
extracts the breast contour in mammogram images using
snakes or active contours. Wei et al., [60] presented a novel
approach that extracts the contour of a region of interest in
mammogram images. Chen and Zwiggelaar [13], which is a
histogram thresholding, edge detection in scale space,
contour growing and polynomial fitting base technique for
segmenting the breast region. Yapa and Harada [63]
presented a breast skin-line estimation and breast
segmentation algorithm using fast marching approach. Table
3 summarizes the breast region segmentation methods
described above, providing the year of the research, the
database used, the number of mammograms used, the
accuracy of segmenting the breast boundary, the pectoral
muscle extraction accuracy, and the ability of the method to
extract the nipple.
3.2 REGIONS OF INTEREST
SEGMENTATION
Regions of interest segmentation is divided into
two main categories, as mentioned previously in this paper,
segmentation using a single view and segmentation using
multiple views. This section lists some algorithms under
each category.
3.2.1 Segmentation using single view
Regions of interest segmentation using a single
view are divided into supervised and unsupervised
segmentation. This section will mainly list the unsupervised
segmentation algorithms under the six categories described
previously. Each category of unsupervised segmentation is
Author
Segmentation
category
Used database
used
mammogr
am
Masses Calcifications Sensitivity Specificity
Schiable et al [47] Region-based UNESP 130 √ X 90% 92%
Cascio et al [11] Contour-based MAGIC-5 N/A √ X 82% N/A
Kom et al., [29] Clustering YGOPH 61 √ X 95.91% N/A
Zheng et al [66] Variant feature N/A 510 √ √ N/A N/A
Mencattini et al [34] Region-based DDSM 200 √ X N/A N/A
Zhang et al [65] Contour-based DDSM N/A √ X N/A N/A
Muralidhar et al [37] Contour-based DDSM N/A √ X N/A N/A
Cao et al [10] Clustering MIAS N/A √ X N/A N/A
Comer et al [18] Clustering U.South Florida N/A √ √ N/A N/A
Anguh and Silva [2] Pseudo-color U. of South Florida 50 √ √ N/A N/A
Bajger et al [5] Graph Mini-MIAS 55 √ X N/A N/A
Guan et al [25] Variant feature MIAS N/A X √ N/A N/A
Stojic et al [51] Variant feature Mini-MIAS N/A X √ N/A N/A
Bhattacharya and Das [7] Variant feature MIAS 65 X √ N/A N/A
Table-4: comprehensive summary of the previously mentioned unsupervised ROI segmentation methods using a single view
Author Year Used database
Number of used
mammograms
Breast
boundary
accuracy
Pectoral muscle
extraction accuracy
Ability of
nipple
extraction
Shahedi et al [48] 2007 Mini-MIAS 66 86% 94% √
Raba et al [42] 2005 Mini-MIAS 320 98% 86% N/A
Wirth and Stapinski [61] 2004 MIAS 32 N/A N/A N/A
Wei et al., [60] 2008 MIAS 322 N/A N/A √
Chen and Zwiggelaar [13] 2010 EPIC 240 98.4% 93.5% N/A
Yapa and Harada [63] 2008 Mini-MIAS 100 99.1% N/A √
Table 3: Results of breast region segmentation techniques
480

5. 5
specialized in segmenting abnormalities such as masses and
calcification.
3.2.1.1 Region-based methods:
Region growing segmentation techniques are used
to segment both masses and calcifications. As an example of
mass segmentation methods, Mencattini et al., [34] proposed
a mass segmentation module for a CAD system
implemented using a new region growing algorithm.
Schiabel et al., [47] described a methodology for
segmenting suspicious masses in dense mammograms. This
methodology is based on the Watershed transformation.
Wang and Karayiannis [58] applied the watershed algorithm
to segment micro-calcification in the segmentation phase of
the approach proposed by the authors to detect micro-
calcifications employing wavelet-based sub-band image
decomposition.
3.2.1.2 Contour-based methods:
Most of the research conducted using the contour-
based methods segmented masses rather than segmenting
calcification. An example of such a method, is the work
conducted by, Zhang et al., [65] proposed a contour-based
segmentation method for mass segmentation. The authors
tested their approach on ROI marked mammograms from
the digital database for screening mammography (DDSM).
Cascio et al., [11] proposed a contour based mass
segmentation algorithm, that was tested on the
mammograms from the Medical Applications on a Grid
Infrastructure Connection (MAGIC-5 collaboration) which
consists of 3762 mammograms. Singh and Al-Mansoori [49]
compared between region growing and gradient-based
segmentation techniques. Muralidhar et al., [37] describe a
mass classification method, based on the snakules
segmentation method which is an evidence active contour
algorithm developed by the authors. The authors tested this
method using mammograms from DDSM database and the
results of using snakules are promising.
3.2.1.3 Clustering methods:
Clustering segmentation methods can segment both
masses and calcifications. And as an example of segmenting
masses using clustering, Kom et al., [29] developed a local
adaptive thresholding technique for mass segmentation.
Velthuizen [56] developed a segmentation method based on
an initial unsupervised clustering. Dominguez and Nandi
[23] applied density slicing technique proposed by
Mudigonda et al., [36] to segment masses. This technique
was used as a part of a method the authors presented for
mass detection in mammograms. Cao et al., [10] proposed
an information based algorithm for mass segmentation on
digital mammograms. This algorithm is called cshells based
on deterministic annealing (CSDA).
Cheng et al., [15] who used the iterative threshold
selection method [43] to implement the segmentation
process as a part of a novel approach based on fuzzy logic
for micro-calcification detection. Moreover, clustering can
be used for the detection of masses and micro-calcifications
together, as shown by the statistical algorithm for
mammogram segmentation presented by Comer et al., [18].
3.2.1.4 Pseudo-color segmentation
Pseudo-color segmentation methods can be used to
detect masses and micro-calcifications together. Such an
algorithm is used by the approach presented by Anguh and
Silva [2].
3.2.1.5 Graph segmentation:
Graph segmentation methods can be used to
segment masses. Bajger et al., [5] employed a graph
segmentation method in his approach to automatically
segment mammogram masses using minimum spanning
trees (MST). The authors tested their approach using two
sets of mammograms. The first set consists of 55
mammograms from the Mini-MIAS database, and the
second set consists of 37 mammograms from a local
database. Ma et al., [31] presented a method based on the
adaptive pyramid (AP) segmentation and sublevel set
analysis of mammograms. In another work, Ma et al., [32]
compared between the performance of the minimum
spanning trees (MST) based segmentation method and that
of the adaptive pyramid (AP) based segmentation method in
terms of robustness. Graph segmentation methods can also
be used to segment micro-calcifications. D’Elia et al., [20]
and D’Elia et al., [21] used the tree-structured markov
random field (TS-MRF) based segmentation method (D’Elia
el al., [22]) to segment mammograms in order to detect
abnormalities.
Author Year Segmentation category Used database
Number of used
mammograms
Bovis and Singh [9] 2000 Left and right mammograms MIAS 144
Xu et al [62] 2010 Left and right mammograms DDSM 60
Wang et al [59] 2006 Left and right mammograms MIAS N/A
Marti et al [33] 2006 Left and right mammograms N/A 64
Velikova et al [55] 2009 Two mammographic views (CC and MLO) of the same breast N/A 1063
Pu et al [41] 2008 Two mammographic views (CC and MLO) of the same breast N/A 200
Sun et al [52] 2004 Two mammographic views (CC and MLO) of the same breast N/A 100
Altrichter et al [1] 2005 Two mammographic views (CC and MLO) of the same breast DDSM N/A
Timp et al [54] 2005 Same view of mammograms taken at different times N/A 389
Timp and Karssemeijer [53] 2006 Same view of mammograms taken at different times DBSP 2873
Wirth et al [67] 2002 Same view of mammograms taken at different times MIAS N/A
Mudigonda et al [36] 2001 Same view of mammograms taken at different times N/A N/A
Table-5: ROI segmentation techniques using multiple views
481

6. 6
3.2.1.6 Variant feature transformation
Variant feature transformation can be used in
mammogram mass segmentation. Zheng et al., [66] used a
technique of single-image segmentation with Gaussian
band-pass filtering. This technique is used as a segmentation
part of a CAD system for mammogram mass detection. The
authors tested their CAD system on 510 mammograms and
the results showed that single-image segmentation method
have a high sensitivity. However, variant feature
transformation methods are mostly effective in segmenting
micro-calcification along with mammogram masses. And as
an example of segmenting micro-calcification, Guan et al.,
[25] proposed an approach based on scale invariant feature
transform (SIFT) to segment micro-calcifications
automatically in mammograms. Stojic et al., [51] modified
the multi-fractal (MF) segmentation method to improve the
segmentation of micro-calcification. Bhattacharya and Das
[7] presented a novel approach for segmenting
mammograms in order to accurately detect micro-
calcification clusters. Table-4 provides a comprehensive
summary of the previously mentioned unsupervised ROI
segmentation methods using a single view, specifying the
year of the research, the segmentation category, the database
used, the number of mammograms used, the ability of the
method to segment masses, and the ability of the method to
segment micro-calcifications, sensitivity, and specificity.
3.2.2 Segmentation using multiple views
Breast images can be taken from different angles,
the most common views are; mediolateral oblique (MLO)
view which is the most important and the cranio-caudal
view (CC). According to the previous image views, image
segmentation techniques using multiple views can be
divided into three categories: left and right mammograms,
two mammographic views (CC and MLO) of the same
breast, and same view mammograms taken at different
times. In the left and right mammograms, the evaluation is
done by checking the symmetry of the fibroglandular tissue
in the two breasts. In the two mammographic views (CC and
MLO) of the same breast, the evaluation is done by
checking the fibroglandular tissue in CC and MLO images
of the same breast. However, in the same view
mammograms taken at different times, the evaluation is
done by checking the changes of the fibroglandular tissue of
the breast at different times.
3.2.2.1 Left and right mammograms:
Bovis and Singh [9] applied the region splitting
technique on the two images obtained from a bilateral
subtraction technique. While, Xu et al., [62] presented a
CAD algorithm for mass detection using bilateral
asymmetry. In this algorithm, the left and right
mammograms are aligned, then a bilateral subtraction
technique is applied. Wang et al., [59] proposed an
automatic registration framework, in order to identify the
differences between corresponding mammograms. The
authors used mammograms from MIAS database to test
their framework. Marti et al., [33] presented a novel method
for mammogram image registration. The authors tested this
method using 64 mammograms containing malignant
masses.
3.2.2.2 Two mammographic views (CC and MLO) of
the same breast:
Velikova et al., [55] proposed and applied the
Bayesian network multi-view system for the detection of
abnormalities. Pu et al [41] developed a computer scheme
based on an ellipse-fitting algorithm. This scheme is used to
build a Cartesian coordinate systemin order to match the
breast masses depicted in the two mammograms. Sun et al
[52] presented ipsilateral multi-view CAD scheme to detect
masses in digital mammograms. Altrichter et al., [1]
proposed a procedure for joint analysis of breast’s two
views, by combining the results of algorithms detecting
mass and micro-calcifications detection. The authors tested
their algorithm using mammograms from the DDSM
database. Results showed that this technique can reduce the
number of false negatives significantly.
3.2.2.3 Same view of mammograms taken at different
times:
Timp et al [54] developed an automatic regional
registration method for mass detection in mammograms; the
authors tested this method using 389 mammograms. Wai
and Brady [57] presented a landmark-based registration
framework of mammograms. However, Timp and
Karssemeijer [53] developed a regional registration
technique to link a suspicious location on prior and current
mammograms. Wirth et al., [67] proposed a non-rigid
mammogram registration approach to match mammograms.
The authors tested the proposed approach using
mammograms from the MIAS database. Table-5
summarizes the ROI segmentation techniques using multiple
views discussed above, specifying the year of the research,
the segmentation category, the database used, and the
number of mammograms used.
4. EVALUATION
In this research, the performance of the
Mammogram enhancement and segmentation methods is
evaluated using different databases. However, most of the
methods were evaluated using two main databases, the
MIAS and the DDSM. The MIAS database is a digital
mammography database produced by the UK research
group. The MIAS database contains left and right digital
mammograms for 161 patients with a total of 322 images of
50X50 micron resolution, and with 8-bit pixel depth. The
images included all types of breast tissues; normal, benign
and malignant. The DDSM database contains 2500 studies;
each study includes two images of each breast. An
abnormality in a mammogram is diagnosed either positive,
i.e. predict that the person has cancer, or negative, i.e.
predict that a person does not have cancer. To aid this
diagnosis in CAD systems, mammogram enhancement and
segmentation techniques should be applied. These
enhancement and segmentation algorithms can be evaluated
482

7. 7
using different techniques. The most popular technique is
dependent on the naked eye and is also known as the visual
perception and still remains a challenging task. In order to
set a standardized benchmark for subjective visual quality
assessment, the International Telecommunications Union
has proposed a set of test procedures defined in the
International Telecommunications Union Recommendation
(ITU-R [27]).This recommendation sets the guidelines for
the subjective assessment test conditions such as the
viewing distance, the test duration, and the observers’
recruitment.
Other techniques can be employed to evaluate the
enhancement and segmentation algorithms such as,
sensitivity and specificity, receiver operating characteristic
(ROC) and free-ROC (FROC). The results of diagnosing a
mammogram can be classified as follows:
 True positive (TP): sick people correctly diagnosed sick.
 False positive (FP): healthy people incorrectly diagnosed sick.
 True negative (TN):healthy people correctly diagnosed
healthy.
 False negative (FN): sick people incorrectly diagnosed
healthy.
Sensitivity and specificity are performance evaluation
statistical measures [17], where sensitivity (true positive
fraction (TPF)) is the fraction of sick people who are
correctly diagnosed as positive, and specificity (true
negative fraction (TNF)) is the fraction of healthy people
who are correctly diagnosed as negative. These measures
specify the accuracy of the system for identifying the actual
positive and actual negative patients respectively. However,
the false-positive-fraction (FPF) and the false-negative-
fraction represent the frequencies of incorrect diagnoses,
therefore: False-negative-fraction = 1 - TPF, and False-
positive-fraction = 1 - TNF. There is an interrelationship
between these measures which makes it important and
meaningful to specify a single pair, either sensitivity and
specificity, or TPF and FPF, but it is inadequate to use only
one measure alone. Equation 1 shows sensitivity measure,
while equation 2 shows specificity measure.
Sensitivity = number of TP / (number of TP + number of
FN)
Specificity = number of TN / (number of TN + number of
FP)
The disadvantage of using a pair of either
sensitivity and specificity, or TPF and FPF for a test is to
have higher sensitivity (higher TPF) but lower specificity
(higher FPF), i.e. one test is more accurate for actually
positive patients and other is more accurate for actually
negative patients, which makes it difficult to indicate which
test is better. This limitation can be resolved using ROC
curves. ROC analysis is used to provide a comprehensive
description of diagnostic accuracy by estimating and
reporting all combinations of sensitivity and specificity.
ROC curves are presented by plotting sensitivity as a
function of FPF (1-specificity). On the other hand, ROC
curves depend on the skill of the radiologist, and, the
imaging procedure’s technical aspects such as, spatial
resolution, noise and contrast (Swets, 1979). However,
FROC analysis accommodates multiple lesions on each
image by allowing multiple reports, and is represented by
plotting the fraction of lesions detected as the vertical axis
and the average number of false-positive detections per
image as the horizontal axis which is not normalized,
instead, it is extended to an arbitrary large number of FP
reports per image. FROC analysis provides greater statistical
power than conventional ROC analysis, and the results of its
analysis depend on the number of locations allowed by the
data analyst. Table-3 through Table-5 summarizes the
results.
5. CONCLUSION
This paper presented a comprehensive
classification and evaluation of mammogram enhancement
and segmentation algorithms. The enhancement techniques
were categorized into four distinct techniques. From the
review, it is obvious that the results produced from the fuzzy
enhancement techniques are best suited for enhancing both
masses and micro-calcifications. On the other hand, the
mammogram segmentation techniques were categorized into
two distinct categories, and the regions of interest
segmentation techniques were further categorized into sub-
categories. From the review of the literature, it can be
inferred that the variant feature transformation category
listed under the unsupervised single view segmentation
techniques are better suited for the segmentation of masses
and micro-calcifications. However, segmentation using
multiple views can also give good results.
References
[1] Altrichter, M., Ludanyi, Z., Horvath, G., “Joint analysis
of multiple mammographic views in cad systems for breast
cancer detection”. In: Proc. of Image Analysis. 14th
Scandinavian Conference, 2005.
[2] Anguh, M.M., Silva, A.C., “Multiscale Segmentation
and Enhancement in Mammograms”. In: Proc. of
SIBGRAPI'97, 1997.
[3] Antonie, M., Zaiane, O.R., Coman, A., “Application of
Data Mining Techniques for Medical Image Classification”.
In: Proc. of Second Intl. Workshop on Multimed. Data Min.
2001.
[4] Rodz, T., Kurdziel, M., Popiela, T.J., Sevre, E.O.D.,
Yuen, D.A.. “Detection of clustered microcalcifications in
small field digital mammography”. Comput. Methods
Programs Biomed. 81(1), pp 56-65, 2006.
[5] Bajger, M., Ma, F., Bottema, M.J., “Automatic tuning of
MST segmentation of mammograms for registration and
mass detection algorithms”. Digit. Image Comput. Pp 400-
407, 2009.
[6] Bandyopadhyay, S.K.. “Survey on Segmentation
Methods for Locating Masses in a Mammogram Image”.
Int. J. Comput. Appl. 9, pp 25-28, 2010.
[7] Bhattacharya, M., Das, A., “Fuzzy Logic Based
Segmentation of Microcalcification in Breast Using Digital
Mammograms Considering Multiresolution”. In: Proc. of
the Int. Mach. Vis. and Image Process. Conf (IMVIP '07), pp
98-105, 2007.
483

8. 8
[8] Blackwell, M., Nikou, C., DiGioia, A.M., Kanade, T.,
“An Image Overlay system for medical data visualization.
Med.”, Image Analysis. 4, pp 67-72, 2000.
[9] Bovis, K., Singh, S., “Detection of masses in
mammograms using texture features”, In: Proc. of the 15th
Int. Conf. on Pattern Recognit. (ICPR'00), pp 267-269,
2000.
[10] Cao, A., Song, Q., Yang, X., Wang, L., “Breast Mass
Segmentation based on Information Theory”, In: Proc. of
the 17th Int. Conf. on Pattern Recognit. (ICPR'04), 3, 2004.
[11] Cascio, D. et al, S.C., Torres, E.L., “Mammogram
Segmentation by Contour Searching and Mass Lesions
Classification With Neural Network” IEEE Transact. Nucl.
Sci. 53, pp 2827-2833, 2006.
[12] Chandrasekhar, R., Attikiouzel, Y., “A simple method
for automatically locating the nipple on mammograms”,
IEEE Transact. Med. Imaging. 16, pp 483- 494, 1997.
Chang, C.M., Laine, A., “Enhancement of Mammograms
from Oriented Information”, In: Int. Conf. on Image
Process. (ICIP'97). 3, pp 524-527, 1997.
[13] Chen, Z., Zwiggelaar, R., “Segmentation of the Breast
Region with Pectoral Muscle Removal in Mammograms”,
In: Proc. of Med. Image Underst. and Analysis
(Warwick2010), 2010.
[14] Cheng, H.D., Cai, X.P., Chen, X.W., Hu, L.M., Lou,
X., “Computer-aided detection and classification of
microcalcifications in mammograms: a survey”, Pattern
Recognit. 36, pp 2967-2991, 2003.
[15] Cheng, H.D., Lui, Y.M., Freimanis, R.I., “A novel
approach to microcalcification detection using fuzzy logic
technique”. IEEE Trans. Med. Imaging. 17, pp 442-450,
1998.
[16] Cheng, H.D., Xu, H., “A novel fuzzy logic approach to
mammogram contrast enhancement”, Inf. Sci. Appl.: An Int.
J. pp 148, 167-184, 2002.
[17] Chu, K., “An introduction to sensitivity”, specificity,
predictive values likelihood ratios. Emerg. Medicine, pp
175-181, 1999.
[18] Comer, M.L., Liu, S., Delp, E.J., “Statistical
Segmentation of Mammograms”, In: Proc. of the 3nd Int.
Workshop on Digit. Mammogr, pp 475-478, 1996.
[19] Dabour, W., “Improved Wavelet Based Thresholding
for Contrast Enhancement of Digital Mammograms”, In:
Proc. of the 2008 Int. Conf. on Comput. Sci. and Softw.
Eng., 2008.
[20] D’Elia, C., Marrocco, C., Molinara, M., Poggi, G.,
Scarpa, G., Tortorella, F., “Detection of Microcalcifications
Clusters in Mammograms through TS-MR’F Segmentation
and SVM-based Classification”, I Proc. of the 17th Int.
Conf. on Pattern Recogn., 2004.
[21] D’Elia, C., Marrocco, C., Molinara, M., Tortorella, F.,
“Detection of Clusters of Microcalcifications in
Mammograms a Multi Classifier Approach”. In: Proc. of the
2008 21st IEEE Int. Symp. on Comput. Based Med. Syst.
(CBMS '08), 2008.
[22] D’Elia, C., Poggi, G., Scarpa, G., “A tree-Structured
Markov Random Field Model for Bayesian Image
Segmentation”. IEEE Transact. Image Process., 12, pp
1259-1273, 2003.
[23] Dominguez, A.R., Nandi, A.K., “Detection of masses
in mammograms via statistically based enhancement,
multilevel-thresholding segmentation, and region selection”,
Comput. Med. Imaging Graphics. 32, pp 304-315, 2008.
[24] Gagnon, L., Lina, J.M., Goulard, B., “Sharpening
Enhancement of Digitized Mammograms with Complex
Symmetric Daubechies Wavelets”. IEEE and EMBS Conf.,
1995.
[25] Guan, Q., Zhang, J., Chen, S., Todd-Pokropek, A.,
“Automatic Segmentation of Micro-calcification Based on
SIFT in Mammograms”. In: Proc. of the 2008 Int. Conf. on
BioMed. Eng. and Inform. (BMEI '08). 2, pp 13-17, 2008.
[26] Hemminger, B.M., Zong, S., Muller, K.E., Coffey,
C.S., DeLuca, M.C., Johnston, R.E., Pisano, E.D.,
“Improving the detection of simulated masses in
mammograms through two different image-processing
techniques:, J. Acad, pp 845-55, 2001.
[27] ITU-R, "Recommendation ITU-R BT.500-11:
“Methodology for the subjective assessment of the quality
of television pictures", Int. Telecommun. Union. Geneva,
2002.
[28] Jiang, J., Yao, B., Wason, A.M., “Integration of fuzzy
logic and structure tensor towards mammogram contrast
enhancement”, Comput. Med. Imaging Graphics. 29, pp 83-
90, 2005.
[29] Kom, G., Tiedeu, A., Kom, M., “Automated detection
of masses in mammograms by local adaptive thresholding”,
Comput. in Biol. and Medicine. 37, pp 37-48, 2007.
[30] Laine, A., Fan, J., Schuler, S., “A Framework for
Contrast Enhancement by Dyadic Wavelet Analysis, pp 91-
100, 1994.
[31] Ma, F., Bajger, M., Bottema, M.J., “Automatic mass
segmentation based on adaptive pyramid and sublevel set
analysis”. Digit. Image Comput.: Tech. and Appl., pp 236-
241, 2009.
[32] Ma, F., Bajger, M., Bottema, M.J., “Robustness of two
methods for segmenting salient features in screening
mammograms”. In Proc of the 9th Bienn. Conf. of the
Australian Pattern Recognit. Soc. on Digit. Image Comput.
Tech. and Appl., 2007.
[33] Marti, R., Raba, D., Oliver, A., Zwiggelaar, R.,
“Mammographic Registration: Proposal and Evaluation of a
New Approach”. In: 8th
Int. Workshop on Digit., pp 213-
220, 2006.
[34] Mencattini, A., Rabottino, G., Salmeri, M., Lojacono,
R., Colini, E., “Breast Mass Segmentation in
Mammographic Images by an Effective Region Growing
Algorithm.” In: Proc. of the 10th Int. Conf. on Adv.
Concepts for Intell. Vis. Syst., 2008.
[35] Mohanalin, Kalra, P.K., Kumar, N., “An automatic
method to enhance microcalcifications using Normalized
Tsallisen tropy.”, Signal Process., 90, pp 952-958, 2010.
[36] Mudigonda, N.R., Rangayyan, R.M. and Desautels,
J.E., “Detection of breast masses in mammograms by
density slicing and texture flow-field analysis”, IEEE Trans.
Med. Imaging. 20, pp 1215-1227, 2001.
[37] Muralidhar, G.S., Markey, M.K., Bovik, A.C.,
“Snakules for Automatic Classification of Candidate
Spiculated Mass Locations on Mammography”, Southwest
Symp. on Image Anal. and Interpret. 197-200, 2001.
[38] Ojala, T., Näppi, J., Nevalainen, O., “Accurate
segmentation of the breast region from digitized
mammograms”, Comput. Med. imaging graphics: the
484

9. 9
official J. of the Comput. Med. Imaging Soc. 25, pp 47-59,
2001.
[39] Oliver, A., Freixenet, J., Martí, J., Pérez, E., Pont, J.,
Denton, E.R.E., Zwiggelaar, R., “A review of automatic
mass detection and segmentation in mammographic
images”. Med. Image Anal. 14, pp 87–110, 2010.
[40] Pisano, E.D., Zong, S., Hemminger, B.M., DeLuca, M.,
Johnston, R.E., Muller, K., Braeuning, M.P., Pizer, S.M.,
“Contrast limited adaptive histogram equalization Image
Processing to Improve the Detection of Stimulated
Speculation in Dense Mamms”. J. Digit. Imaging., pp 11,
193-200, 1998.
[41] Pu, J., Zheng, B., Leader, J.K., Gur, D., “An ellipse-
fitting based method for efficient registration of breast
masses on two mammographic views”. Med. Phys., pp 35,
487-94, 2008.
[42] Raba, D., Oliver, A., Marti, J., Peracaula, M., Espunya,
J., “Breast segmentation with pectoral muscle suppression
on digital mammograms.”, In: Iber. Conf. on Pattern
Recognit. and Image Anal. , pp 471-478, 2005.
[43] Ridler, T.W.,Calvard, S., “Picture thresholding using an
iterative selection method”, IEEE Trans.Syst.,Man, pp 630-
632, 1978
[44] Sajda, P., Spence, C., Parra, L., “A multi-scale
probabilistic network model for detection, synthesis
and compression in mammographic image analysis”, Med.
Image Anal. 7, pp 187-204, 2003.
[45] Sampat, M.P., Bovik, A.C., “Detection of Spiculated
Lesions in Mammograms”. In: Proc. of the 25th Annu. Int.
IEEE Conf. on Eng. in Medicine and Biol. Soc., pp 810-813.
[46] Scharcanski, J., Jung, C.R., “Denoising and enhancing
digital mammographic images for visual screening”.
Comput. Med. Imaging Graphics. 30, pp 243-254, 2006.
[47] Schiabel, H., Santos, V.T., Angelo, M.F.,
“Segmentation technique for detecting suspect masses in
dense breast digitized images as a tool for mammography
CAD schemes.” In: Proc. of the 2008 ACM Symp. on Appl.
Comput., 2008.
[48] Shahedi, M.B.K., Amirfattahi, R., Azar, F.T., Sadri, S.,
2007. “Accurate Breast Region Detection in Digital
Mammograms Using a Local Adaptive Thresholding
Method”. Workshop on Image Anal. for Multimed. Interact.
Serv. 26. 2007.
[49] Sing, S., Al-Mansoori, R., “Identification of Regions of
Interest in Digital Mamgrs.” J. Intell. Syst. 10, pp 183-217,
2000.
[50] Stefanou, H., Kakouros, S., Cavouras, D., Wallace, M.,
“ Wavelet-based Mammographic Enhancement”. In: Proc.
of the Fifth Int. Netw. Conf. (INC). Samos, Greece, 2005.
[51] Stojic, T., Reljin, I., Reljin, B., “Adaptation of
multifractal analysis to segmentation of microcalcifications
in digital mammograms”. Phys. A-Stat. Appl., 367, pp 494-
508, 2006.
Swets, A., “ ROC analysis applied to the evaluation of
medical imaging techniques”. Investig. Radiol. 14(2), pp
203-206, 1979.
[52] Sun, X., Qian, W., Song, D., “Ipsilateral-mammogram
computer-aided detection of breast cancer”. Comput. Med.
Imaging Graphics. 28, pp 151-158, 2004.
[53] Timp, S., Karssemeijer, N., “ Interval change analysis
to improve computer aided detection in mammography”.
Med. Image Anal. 10, pp 82-95, 2006
[54] Timp, S., van Engeland, S., Karssemeijer, “A regional
registration method to find corresponding mass lesions in
temporal mammogram pairs. Med. Phys. 32(8): pp 29-38,
2005.
[55] Velikova, M., Samulski, M., Lucas, P.J.F.,
Karssemeijer, N., “Improved Mammographic CAD
Performance Using Multi-View Information a Bayesian
Network Framework”. Phys. Medicine Biol., 54, pp 1131-
1147, 2009.
[56] Velthuizen, R.P., “Computer Diagnosis of
Mammographic Masses”. In: Proc. of the 29th Appl. Imag.
Pattern Recognit. Workshop (AIPR '00), 2000.
[57] Wai, L.C.C., Brady, M., “ Curvilinear Structure Based
Mammographic Registration”. In: Comput. Vis. for Biomed.
Image Appl.: Curr. Tech. and Future Trends (ICCV
Workshop), Bejing, China, pp 261-270, 2005.
[58] Wang, T.C., Karayiannis, N.B., “Detection of
microcalcifications in digital mammograms Using
Wavelets”. IEEE Trans. Med Imaging. 17, pp 498-509,
1998.
[59] Wang, K., Qin, H., Fisher, P.R., Zhao, W., “Automatic
registration of mammograms using texture-based anisotropic
features”. In: Proc. of 2006 IEEE Int. Symp. on Biomed.
Imaging: From Nano to Macro, pp 864-867, 2006.
[60] Wei, L., sheng, L.Y., Xianyi, R., Peng, D., “A New
Contour Detection in Mammogram Using Sequential Edge
Linking. In: Proc. of the 2008 Second Int. Symp. on Intell.
Inf.”, 2008.
[61] Wirth, M.A., Stapinski, A., “Segmentation of the
breast region in mammograms using snakes.” In: Proc. of
the 1st Canadian Conf. on Comput. and Robot Vis. (CRV
'04). 2004.
[62] Xu, S., Liu, H., Xu, X., Song, E., Zeng, J., “Bilateral
Asymmetry Detection in Mammograms Using Non-rigid
Registraion and Pseudo-color coding”. In: Int. Conf. on
Electr. and Control Eng. (ICECE2010), pp 544-547, 2010.
[63] Yapa, R.D., Harada, K., “ Breast Skin-Line Estimation
and Breast Segmentation in Mammograms using Fast-
Marching Method”. Int. J. Biol. Med. Sci. 3, pp 54-62, 2008.
[64] Yu, Z., Bajaj, C., “ A fast and adaptive method for
image contrast enhancement”. In: Int. Conf. on Image
Process. (ICIP), pp 1001-1004, 2004.
[65] Zhang, Y., Tomuro, N., Furst, J., Raicu, D.S., “A
Contour-based Mass Segmentation in Mammograms”. In:
Proc. of the Soc. Res. Symp, 2009.
[66] Zheng, B., Chang, Y.H., Gur, D., “Computerized
Detection of Masses in Digitized Mammograms Using
Single-Image Segmentation and a Multilayer Topographic
Feature Analysis. Acad. radiol., 2, pp 959-66, 1995.
[67] Wirth, M.A., Narhan, J., Gray, D., “Nonrigid
mammogram registration using mutual information”. In:
Proc. of SPIE Med. Imaging: Image Process. USA.4684,
pp 562-573., 2002.
485