International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5 - Case study of CAD Based Mammographic Lesions using Wavelet Decomposition
This document summarizes a research article that studied the use of wavelet decomposition to analyze mammographic lesions. The study aimed to characterize true masses versus falsely detected masses. Breast cancer rates have increased each year since 1980, though death rates have decreased due to mammography. Current computer-aided detection systems aim to assist rather than replace radiologists. The study used wavelet decomposition transforms to analyze characteristics of true versus false masses detected on mammograms. This technique could help improve computer-aided diagnosis systems by better distinguishing between malignant and benign lesions.
Succession “Losers”: What Happens to Executives Passed Over for the CEO Job?
Similar a International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5 - Case study of CAD Based Mammographic Lesions using Wavelet Decomposition
CLASSIFICATION OF LUNGS IMAGES FOR DETECTING NODULES USING MACHINE LEARNING sipij
Similar a International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5 - Case study of CAD Based Mammographic Lesions using Wavelet Decomposition (20)
College Call Girls Vyasarpadi Whatsapp 7001305949 Independent Escort Service
International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5 - Case study of CAD Based Mammographic Lesions using Wavelet Decomposition
1. International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5
Case study of CAD Based Mammographic Lesions using Wavelet
Decomposition
Elayabharathi.T1, Dr.Nagappan.A2
1
is a research Scholar in the Department of computer science &engineering at the Vinayaka mission research foundation
deemed university, Salem, Tamilnadu, India.
Abstract – This paper describes the efforts by study of the characteristics of true masses compared to the falsely detected
masses is carried out using wavelet decomposition transform. According to the cancer statistics, the breast cancer incidence rate
is increased almost every year since 1980, the death rate is shown a substantial decrease. Both trends may be attributed, in large
part to mammography which is widely recognized as the most sensitive technique for breast cancer detection
Index terms –Lesions, Wavelet, contourlet, Mammogram, CAD
1. Introduction difficulty in maximizing both sensitivity to tumoral
Conventional mammography is a film based x-ray growths and specificity in identifying their nature.
Technique referred as a screen-film mammography. Full X-ray mammography is the best current method for early
field digital mammography is a new technology in which a detection of breast cancer, with an accuracy of between
solid state detector is used instead of film for the generation 85% and 95%[3]. Identifying abnormalities such as
of the breast image. Modern applications, including calcifications and masses often requires the eye of a
computer aided detection and computer aided diagnosis, trained radiologist. As a result, some anomalies may be
computer display and interpretation, digital image and missed due to human error as a result of fatigue, etc.
transmission and storage, require a digital format of the The development of CAD systems that assist the
mammogram radiologist has thus become of prime interest, the aim
1.1 Objective being not to replace the radiologist but to offer a second
The purpose of our study was to retrospectively opinion. Eventually, the state-of-the-art could advance to
evaluate the impact on recall rates and cancer detection the point where such systems effectively substitute for
when converting from film-screen to digital mammography trained radiologists, an eventuality that is desirable for
in a small community-based radiology practice. small outfits that cannot afford to have an expert
Digital mammography offers considerable advantages over radiologist at their continuous disposal. For example, a
film-screen mammography [1-3]. Despite advantages, it has CAD system could scan a mammogram and draw red
been slow to be adopted. This reluctance is due to many circles around suspicious areas. Later, a radiologist can
factors, including the high initial capital expenditure and the examine these areas and determine whether they are
question of whether the added expense results in a better true lesions or whether they are artifacts of the scanning
mammography “product” [4-7]. The reluctance to upgrade process, such as shadows.
to digital mammography is especially true of small
community-based imaging centers, where capital is less To our knowledge, no prior study has compared cancer
prevalent and patient volumes are lower than in larger detection and recall rates at a single center before and after
metropolitan locations. the installation of a digital mammography system, keeping
1.2 statistics and discussions the interpreting radiologists constant. Such a study would
Breast cancer ranks first in the causes of cancer limit the number of uncontrolled variables, allowing
deaths among women and is second only to cervical potential outcomes to be mediated only by the introduction
cancer in developing countries[8]. The best way to of the technology and the variability in the women
reduce death rates due to this disease is to treat it at an undergoing screening.
early stage. Early diagnosis of breast cancer requires an 2. Background
effective procedure to allow physicians to differentiate Considerable effort has been expended to develop
between benign tumors from malignant ones. CAD systems to aid the trained radiologist identify
Developing computer-aided diagnosis (CAD) systems to areas with possible pathology on an image. Most of
help with this task is a non-trivial problem, and current these efforts have concentrated on X-ray mammography
methods employed in pursuit of this goal illustrate the and chest radiography. A number of CAD schemes
Issn 2250-3005(online) September| 2012 Pag1318
2. International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5
have been investigated in literature. These include: 3. Methods
By careful consideration of the design of various
• Subtraction techniques that identify anomalies by CAD schemes, it is possible to categorize the techniques
comparison with normal tissue employed under three broad headings:
• Topographic techniques that perform feature
extraction and analysis to identify anomalies • Data reduction - the image is examined in order
• Filtering techniques that use digital signal to identify the ROIs.
processing filters, often developed especially to • Image enhancement - the ROIs are subjected to
augment anomalies for easy detection processes that enhance or augment the visibility of
• staged expert systems that perform rule-based pathological anomalies, such as microcalcifications
analysis of image data in an attempt to provide a and lesions.
correct diagnosis
• Diagnosis - the ROIs are subjected to one or more
The majority of CAD systems attempt to of the broad categories of procedures mentioned in
identify anomalies by either looking for image Section 2 in order to arrive at a diagnosis, most
differences based on comparison with known normal commonly in the form of “benign” or “malignant”
tissue (subtraction techniques)[4] or by image feature
identification and extraction of features that correlate These categories are extremely broad, and there
with pathological anomalies, such as in texture analysis may exist CAD systems that subject images to
(topographic techniques)[11, 4, 6, 36]. Most systems techniques that do not fall under one of them. However,
proceed in stages, first examining the image data and most of the CAD systems employ methods that can be
extracting pre-determined features, then localizing classified under one or more of them.
regions of interest or ROIs which can be examined
further for potential anomalies. High degrees of 3.1 Data Reduction
sensitivity have been achieved using several of these Data reduction is the process by which an image
techniques, but many have been hampered by high is decomposed into a collection of regions that appear to
false-positive rates and hence low specificity. The contain anomalies that differ from the surrounding
problem of false positives is compounded further by tissue. These regions are usually a strict subset of the
the fact that false positive rates are reported per original image and are subregions of the original image
image, not per case. Since many radiological that may contain ROIs. By doing this, the CAD system
examinations i n c l u d e more than one image, the actual need only process those subregions identified by the data
number of false positives may be a multiple of those reduction step, rather than the entire input image. Data
reported. reduction accomplishes two objectives
A number of different approaches have been simultaneously[34]:
employed in an effort to reduce false positive rates, many
of them focusing on the use of artificial neural • An increase in throughput v i a a reduction in input
networks (ANNs). A common metric used for evaluating data
the performance of CAD systems, the receiver operating • A reduction in false positives by limiting the scope
curve or ROC (see Appendix A), is commonly used to of the detection algorithms in the rest of the CAD
evaluate a CAD scheme’s degree of tradeoff between system to the ROIs only. With less of the original
sensitivity and specificity. The area under this curve, image to worry about, the CAD system gains
Az , is a measure of overall performance, with a value of specificity since less image means less false-
Az closer to 1 indicating better performance. Since positives in general, assuming that the detection
algorithms work as intended.
sensitivity in most techniques is quite high, specificity
It is clear that the most obvious way to perform
often becomes the limiting factor, with techniques data reduction is to have a trained radiologist identify
displaying higher specificity performing at higher Az the ROIs for the CAD system. This can be
values. accomplished through a graphical interface to the CAD
This study decomposes several techniques and system that allows the radiologist to specify suspicious
identifies their salient features and characteristics w i t h regions. It should be noted that some CAD systems do
respect to performance. The extent of the array o f not require this step at all due to the nature of their
techniques e x a m i n e d herein is by no means all- diagnostic process, such as that those that employ
inclusive; rather, a number of techniques are described subtraction techniques.
and their performance evaluated.
Issn 2250-3005(online) September| 2012 Pag1319
3. International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5
can then be used by a radiologist; however, the CAD
scheme proposed in
[2] uses pseudo-color mapping[10] to convert the
grayscale to a color image. This is done since human
vision can only discern a limited number of grayscale
levels. The end results is a pseudo-color breast map in
which the lesions have been highlighted in different
colors and confirmed by visual inspection by a trained
radiologist. Anguh et al[2] claim that this multiscale
segmentation and enhancement method detects
virtually all lesions identified by an expert radiologist in
the process of visual inspection in initial tests on 25
mammograms.
4. Wavelet-Based Enhancement
Koren et al[18] developed a contrast
Figure 1: overview of front-end data reduction module with enhancement method based on the adaptation of
images specific enhancement schemes for distinct
mammographic features, which were then used to
3.2 Image Enhancement combine the set of processed images into an enhanced
Mammographic image enhancement methods are image. In their scheme, the mammo- graphic image is
typically aimed at either improvement of the overall first processed for enhancement of microcalcifications,
visibility of features or enhancement of a specific sign masses and stellate lesions. From the resulting
of malignancy. Various schemes for doing this exist, enhanced image, the final enhanced image is
with most of them based in signal processing synthesized by means of image fusion[20]. Specifically,
techniques used either in their original form (such as their algorithm consisted of two major steps:
simple histogram equalization) or adapted for specific
use in mammography. 1. the image is first subjected to a redundant B-spline
A number of generic image enhancement methods wavelet transform decomposition[18]
exist. Histogram equalization and fuzzy image from which a set of wavelet coefficients is obtained
enhancement[35] are just two examples. Though a 2. the wavelet coefficients are modified distinctly for
whole slew of image enhancement techniques exist in the each type of malignancy (microcal- cifications,
general domain, very few are specifically targeted at the stellate lesions or circumscribed masses).
enhancement of mammographic images. Section 4 3. the multiple sets of coefficients thus obtained are
describes one of them. fused into a single set from which the reconstruction
3.3 Statistical Techniques is computed The algorithm is illustrated in Figure
2, as applied to a digitized mammogram that they
Anguh et al[2] propose a multiscale method for obtained from the University of Florida database.
segmenting and enhancing lesions of various sizes in The theoretical treatment for the mathematics
mammograms. The first stage applies a multiscale involved in this scheme is beyond the scope of this
automatic threshold esti- mator based on histogram study. However, it is interesting to note that the
moments to segment the mammogram at multilevels. enhance image produced by this scheme is “more
The second stage then converts the segmented image easily interpreted by a radi- ologist compared to
using pseudo-color mapping to produce a color images produced via global enhancement
image[2]. The final result is analogous to a breast map techniques”[18]. It is yet to be seen what
which provides an adequate basis for radiological breast improvement this enhancement scheme can
tissue differentiation and analysis in digital contribute t o existing CAD schemes.
mammography. Their paper provides a treatment on the
mathematical theory of moments before present an
algorithm for the multiscale thresholding of the
mammogram. The result of this thresholding technique
is a mammographic map or breast map based on
various thresholds with varying object sizes. This map
Issn 2250-3005(online) September| 2012 Pag1320
4. International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5
where 0 ≤ u ≤ 1.
Figure 2. Overview of the image fusion algorithm based
on B-Spline Wavelet Transform
5. Classification
Apply B´zier splines[17] to both lesion
e
detection and characterization, where lesion detection
is achieved by segmentation using a threshold
computed from the B´zier smoothed histogram and
e
lesion characterization is achieved by means of fitness
between Gaussian and B´zier histograms of data
e
projected on principal components of the segmented Figure 7. System overview of the B´zier spline-based
e
lesions. The most interesting component of their thresholding and segmentation algorithm.
systems in the use of the B´zier splines as a basis of
e
thresholding of the mammographic image - the overall 6. Results
performance of their classification scheme is significantly The database used for this work comes from
worse than that seen from, for example, the ANN-based vina yaka mis sio ns hospital and is a set of 43 digitalized
scheme used by Chen et al[2], mammographic images, 25 of them corresponding to benign
B´zier splines are a spline approximation method,
e pathologies and 18 to breast cancer. Due to the good
developed by the French engineer Pierre B´zier for use
e performance of the detection stage, only few
in the design of Renault automobile bodies. Since a microcalcifications are not detected. Detection of the
B´zier curve lies within the convex hull of the control
e maximum possible number of microcalcificactions is very
points on which it is fitted, applying it to the important for the success of the system, being very critical
histogram of the original image produces a smoothed the correct adjustment of noise thresholds in the contourlet
histogram from which a threshold can be easily chosen pre-processing stage.
by simply finding the largest minimum or the
rightmost inflection point, which is where the highest
7. Conclusions
brightness level is located. As a rule, a B´zier curve
e
The proposed system combines several state of the
is a polynomial of degree one less than the number of
art image processing techniques, namely contourlet
control points used. Since a typicalgrayscale image
transforms for the noise removal of the mammographic
consists of 256 brightness levels, the histogram values of
images and border detection with the wavelet transform
these levels can be used as the control points for a B´zier
e
modulus maxima lines. The tested wavelet based
curve polynomial of degree 255. If the histogram levels
compression method proved to be an accurate approach
are denoted by pk = (xk , yk ), where both k and xk
for digitized mammography.
vary from 0 to 255, then these coordinate
points can be blended to produce a position vector P
(u) which describes the path of an approximating References
B´zier polynomial between p 0 and p255 :
e [1] Predrag Baki’c And. Application of neural
networks in computer aided diagnosis of breast
cancer. http://citeseer.nj.nec.com/404677.html.
[2] M.M. Anguh and A.C. Silva. Multiscale
segmentation and enhancement in mammo- grams.
Issn 2250-3005(online) September| 2012 Pag1321
5. International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5
http://www.visgraf.impa.br/sibgrapi97/anais/ART03/A Statistics, 8(3), September 1999.
RT03.ps.gz. [15] D. Hearn and M.P. Baker. Computer Graphics.
[3] L. Baker. Breast cancer demonstration project: Prentice Hall, Inc., 2nd edition, 1994.
five year summary report. Cancer, pages 194–225, [16] Y. Hirose, K. Yamashita, and S. Hijiva.
1982. Backpropagation algorithm which varies the
[4] Keir Bovis and Sameer Singh. Detection number of hidden units. Neural Networks, 4:61–
of masses in mammograms using texture features. 66, 1991.
http://www.dcs.ex.ac.uk/research/pann/master/cameo/p [17] International Commission on Radiological Units
aper80.ps. and Measurements. Medical Imaging - The
[5] Garra B.S., Krasner B.H., Horii S.C., Archer S., Assessment of Image Quality, 1996.
Muk S.K., and Zerman R.K. Improving the [18] Andrew Laine Iztok Koren and Fred Taylor.
distinction between benign and malignant breast Enhancement via fusion of mammographic
lesions: the value of sonographic texture analysis. features. http://citeseer.nj.nec.com/288745.htm.
Ultrasound Imaging, 15:267–285, 1993. [19] Barry Kalman, Stan C. Kwasny, and
[6] I. Christoyianni, E. Dermatas, and G. William R. Reinus.
Kokkinakis. [20] I. Koren, A. Laine, and F. of ab- normal tissue in digital
Neural classification Taylor. Image fusion
[7] Shane Dickson. Investigation of the use of Neural using steerable dyadic wavelet transform.
Networks for Computerised Medical Image Analysis. Proceedings of the IEEE International Conference
PhD thesis, Department of Computer Science, on Image Processing, 3:232–235, 1995.
University of Bristol, February 1998. [21] A. Laine, J. Fan, and W. Yang. Wavelets for
[8] Chen D.R., Chang R.F., and Huang Y.L. contrast enhancement of digital mam- mography.
Computer-aided diagnosis applied to us of solid IEEE Engineering in Medicine and Biology
breast nodules by using neural networks. Magazine, 14(5):536–550, 1995.
Radiology, 213(2):407–412, November http://citeseer.nj.nec.com/349213.html.
1999. http://www.cs.ccu.edu.tw/ [22] A. Laine, S. Schuler, J. Fan, and W. Huda.
rfchang/radio9911.pdf. Mammographic feature enhancement by
[9] V. Goldberg and A. Manduca. Improvement in multiscale analysis. IEEE Transactions in
specificity of ultrasonography for diag- nosis of Medical Imaging, MI-13:725–740, 1994.
breast tumors by means of artificial intelligence. http://citeseer.nj.nec.com/laine94mammographic.ht
Medical Physics 1992, 19:1475– ml.
1481, 1992. [23] H. Li, K. Liu, and S. Lo. Fractal modeling and
[10] R.C. Gonzalez and R.C. Woods. Digital Image segmentation for the enhancement of
Processing. Addison Wesley., 1992. microcalcifications in digital mammograms.
[11] R. Gupta and P.E. Undrill. The use of texture IEEE Trans. Med. Imaging, 16:785–798,1997.
analysis to identify suspicious masses in http://citeseer.nj.nec.com/li97fractal.html.
mammography. [24] S. Liu and E. Delp. Multiresolution detection
http://www.biomed.abdn.ac.uk/Abstracts/A00337/ of stellate lesions in mammograms. In
[12] R. Hanka, T. Harte, A. Dixon, D. Lomas, and P. Proceedings of the IEEE International
Britton. Neural networks in the inter- pretation of Conference on Image Processing, pages 109–112,
contrast-enhanced magnetic resonance images of October 1997.
the breast. In Proceedings Current Perspectives in http://citeseer.nj.nec.com/article/liu97multiresolutio
Healthcare Computing Conference, Harrogate, UK, n.html.
pages 275– [25] Sheng Liu. Multiresolution detection of
283, March 1996. spiculated lesions in digital mammograms.
[13] Trevor Hastie, Debra Ikeda, and Robert http://citeseer.nj.nec.com/179226.html.
Tibshirani. Computer-aided diagnosis of mam- [26] Laura N. Mascio, Jose M. Hernandez, and
mographic masses. Technical report, Clinton M. Logan.
Departments of Preventive Medicine & Biostatis- [27] A. Mendez, P. Tahoces, M. Lado, M. Souto, and
tics, University of Toronto, June 1996. J. Vidal. Computer aided diagnosis: Automatic
ftp://utstat.toronto.edu/pub/tibs/mammo.ps.Z. detection of malignant masses in digitized
[14] Trevor Hastie, Debra Ikeda, and Robert mammograms. Medical Physics, 25(6):957–964,
Tibshirani. Statistical measures for the 1998.
computer-aided diagnosis of mammographic [28] A. Ossen, T. Zamzow, H. Oswald, and E. Fleck.
masses. Journal of Computational and Graphical
Issn 2250-3005(online) September| 2012 Pag1322
6. International Journal Of Computational Engineering Research (ijceronline.com) Vol. 2 Issue. 5
Segmentation of medical images using neural-
network classifiers. In Proceedings of the
International Conference on Neural Networks
and Expert Systems in Medicine and Healthcare
(NNESMED’94), pages 427– 432,1994.
http://citeseer.nj.nec.com/ossen94segmentation.html
[29] Riccardo Poli and Guido Valli. Optimum
segmentation of medical images with hopfield
neural networks. Technical Report CSRP-95-12,
University of Birmingham School of Computer
Science, October 1995.
http://citeseer.nj.nec.com/156628.html.
[30] Qi and Snyder. Lesion detection and
characterization i n digital mammography by
b´zier histograms.
e
http://citeseer.nj.nec.com/348097.html.
[31] Gonzalez R.C. and Woods R.E. Image
Compression, pages 312–315. Reading, Mass.:
Wesley, 1991.
[32] Guido Valli Riccardo. Neural networks and
prior knowledge help the segmentation of
medical images.
http://citeseer.nj.nec.com/336457.html.
[33] D. E. Rumelhart, G. E. Hinton, and R. J.
Williams. Learning representations by back-
propagating errors. Nature, 323:533–536, 1986.
[34] H. Sari-Sarraf, S. S. Gleason, and R. M.
Nishikawa. Front-end data reduction in
computer-aided diagnosis of mammograms: A
pilot study. In SPIE’s Medical Imag- ing
Conference, February 1999. http://www-
ismv.ic.ornl.gov/publications/spie99.pdf.
[35] Sameer Singh and Reem Al-Mansoori.
Identification of regions of interest in
digital mammograms. Journal of Intelligent
Systems, 10(2):183–217, 2000.
http://www.dcs.ex.ac.uk/research/pann/pdf/pa
nn SS 005.pdf.
[36] P. Undrill and R. Gupta. Texture analysis and
boundary refinement to outline mam- mography
masses. IEEE Colloquium (Digest), 072:5/1–5/6,
1996.
[37] Kevin S. Woods. Automated image
analysis techniques for digital mammography.
http://citeseer.nj.nec.com/woods94automated.html.
Issn 2250-3005(online) September| 2012 Pag1323