• Share
  • Email
  • Embed
  • Like
  • Save
  • Private Content
Kx2418461851
 

Kx2418461851

on

  • 171 views

IJERA (International journal of Engineering Research and Applications) is International online, ... peer reviewed journal. For more detail or submit your article, please visit www.ijera.com

IJERA (International journal of Engineering Research and Applications) is International online, ... peer reviewed journal. For more detail or submit your article, please visit www.ijera.com

Statistics

Views

Total Views
171
Views on SlideShare
171
Embed Views
0

Actions

Likes
0
Downloads
1
Comments
0

0 Embeds 0

No embeds

Accessibility

Categories

Upload Details

Uploaded via as Adobe PDF

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Processing…
Post Comment
Edit your comment

    Kx2418461851 Kx2418461851 Document Transcript

    • R.Poongothai, I.Laurence Aroquiaraj, D.Kalaivani / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1846-1851 A Novel Pectoral Removal Approach in Digital Mammogram Segmentation and USRR Feature Selection R.Poongothai*, I.Laurence Aroquiaraj**, D.Kalaivani*** * (Research Scholar,Periyar Universty,Salem) ** (Assistant Professor,Periyar Universty,Salem) *** (Research Scholar,Periyar Universty,Salem)Abstract Detection of micro calcification based on Another method such as a biopsy is used fortextural image segmentation and classification is detection of breast cancer, where the patientthe most effective early-diagnosis of breast cancer. undergoes a surgical procedure. Many CAD basedThe aim of segmentation is to extract a breast systems were used by the radiologist for theregion by estimation of a breast skin-line and a diagnosis of breast masses. The limitation of thesepectoral muscle as well as removing radiographic systems is mainly at breast mass detection. Therefore,artifacts and the background of the mammogram. accurate segmentation of a breast mass is anThe intensity value is taken using histogram important step for the diagnosis of breast cancer infunction. After detecting the region intensity value mammography. Breast mass is a localized swelling,from mammogram, the edge between pectoral which is described by its characteristics. Usuallyregion and breast region will be detected using these boundaries are of varying size and shape.Fuzzy Connected Component Labeling. Our Because of this, breast mass segmentation is aproposed method worked good for separating challenging task.pectoral region by eliminating cancer area. Theraster scan method is used for fixing the pectoral The rest of the paper is organized as follows.removal area in the original image. After removal Section 2 gives an introduction to the previous work.pectoral muscle from the mammogram, so that Section 3 present preprocessing using Medianfurther processing is confined to the breast region Filter.Section 4 presents pectoral muscle removalalone. Two error measures were used to compare using Fuzzy with CCL. Section 5 presentsthese three methods performance. One of the Segmentation using Watershed Transformation.measures is Mean absolute error (MAE) and Section 6 describes the GLCM for Featureanother one is hausdroff distance measure to find Extraction. Section 7 presents the proposedthe distance between the binary pectoral region unsupervised relative reduct (USRR) algorithm forand binary breast region alone. We applied feature selection. Section 8 describes the classifierunsupervised feature selection of rough set based tool. The experimental results are discussed inRelative Reduct algorithms and it demonstrates Section 9 and conclusion is presented in Section 10.effectively remove the redundant features. Thequality of the reduced data is measured by the II. Previous Workclassification performance and it is evaluated In the literature, several methods have beenusing WEKA classifier tool. proposed to identify the pectoral muscle in mammograms. T.S.Subashini et al. proposed aKeywords: Digital mammogram, morphological technique for pectoral muscle removal based onreconstruction, Fuzzy connected component labeling, connected component labeling. To extract thesegmentation, feature selection. pectoral muscles from this image, binary image should be obtained [4]. Sara Dehghani et alI. Introduction employed a technique for removal of dark Breast cancer is the second most major background parts which are not important inhealth problem in developed countries. According to processing of mammography images [3].Roshanthe latest study from NCI (U.S. National Cancer DharshanaYapa proposed a technique to compareInstitute) more than 207,090 new cases were reported performance of connected component labelingand about 39,840 were the death statistics in algorithms on grayscale digital mammograms [8].In2010.Nearly 1.4 million US women have a family Ali Cherif Chaabani et al. Proposed to extract thehistory of breast cancer. As there is no effective breast region,we used a method based on automaticmethod for its prevention, the diagnosis of breast thresholding(Otsu‟s) and connected componentcancer in its early stage of development has become labeling algorithm. Identifying the pectoral muscleVery crucial for the prevention of cancer. Computer- has been done using Hough transform and activeaided diagnosis (CAD) systems play an important contour [7]. Masek et al [9].employed both arole in earlier diagnosis of breast cancer [1]. threshold-based algorithm and a straight line fitting 1846 | P a g e
    • R.Poongothai, I.Laurence Aroquiaraj, D.Kalaivani / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1846-1851technique to represent the pectoral muscle. The scanning is done from the right or left side of thestraight line representation by this method is not an image to detect the intensity discontinuities. Theefficient way to identify the pectoral muscle due to resulting image contains pectoral muscles alone andthe existence of curved pectoral muscles. this region is completely removed from the original image.III. Preprocessing Mammograms are difficult images tointerpret, and a preprocessing phase is necessary toimprove the quality of the images and make thefeature extraction phase more reliable. In order tolimit the search for abnormalities by computer aideddiagnosis systems to the region of the breast withoutundue influence from the background of themammogram, removal of artifacts and removal ofpectoral muscle is necessary. Preprocessing stageconsists of two parts. The first part involves theremoval of unwanted parts from the image and thesecond part deals with reducing the high frequencycomponents present in the image. Artifacts areremoved by morphological open operation followedby reconstruction operation [2]. Median filtering issimilar to using an averaging filter, in that each pixelis set to an average of the pixel values in theneighborhood of the corresponding input pixelmedian filtering is better able to remove theseoutliers without reducing the sharpness of the image. Figure 1: (a) and (b): Different sizes and intensities of the pectoral muscle in breast mammograms.IV. Pectoral Muscle Detection and Removal Pectoral muscles are the regions in Procedure: merging Fuzzy with CCL algorithmmammograms that contain brightest pixels. These INPUT: over segmented imageregions must be removed before detecting the tumor OUTPUT: Pectoral muscle removal from thecells so that mass detection can be done efficiently. Original imagePectoral muscles lie on the left or right top corner 1. Iterate through each element of the data bydepending on the view of the image. We must detect column, then by row (Raster scanning).the position of the pectoral muscles (left top corner or 2. If the element is not the background. Get theright top corner) as is shown in the figure 1(a) and neighboring elements of the current element.(b), before removing it. For this searching for If there is no neighbors, uniquely label thenonzero pixels are simultaneously done from the left current element and continue.and right top corner. Width of the image in which the 3. Otherwise, find the neighbors with thenon zero pixel detected from both the corner were smallest label and assign it to the currentcounted and compared. If the left width is smaller element. Store the equivalence betweenthan the right width then it is assumed that pectoral is neighboring labels.on the left side of the image else it is on the right 4. Iterate through each element of the data byside. column, then by row. 5. If the element is not the background. Re- From the detected corner pixel the intensity label the element with the lowest equivalentdiscontinuity is detected on each and every column of label.the same row. Coordinates of the pixel in which the 6. To fuzzy construction, the input grays isintensity change is encountered is considered as ranged from 0-255 gray intensity, andwidth of the pectoral region. All the pixels, which lie according to the desired rules the gray levelinside pectoral width and half of the height of the is converted to the values of the membershipwhole image is segmented from the original image. functions.This rectangle shaped image contains the entire 7. Compare the cropped image and originalpectoral muscles. image. Then retrieving pixels from the original image. To extract the pectoral muscles [18] fromthis image binary image should be obtained by V. Segmentationsimple thresholding. This binary image contains Segmentation [19] refers to the process ofpectoral muscles and other tissues. To segment the partitioning a digital image into multiple segmentspectoral muscles alone from the binary image raster 1847 | P a g e
    • R.Poongothai, I.Laurence Aroquiaraj, D.Kalaivani / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1846-1851formed by the union of connected pixels. The goal of level spatial dependence matrix [13]. By default, thesegmentation is to simplify the representation of an spatial relationship is defined as the pixel of interestimage into something that is more meaningful and and the pixel to its immediate right (horizontallyeasier to analyze, regions. Image segmentation is adjacent), but you can specify other spatialtypically used to locate objects and boundaries. More relationships between the two pixels. Each elementprecisely, image segmentation [20] is the process of (I, J) in the resultant GLCM is simply the sum of theassigning a label to every pixel with the same visual number of times that the pixel with value I occurredcharacteristics such as color, intensity, or texture. The in the specified spatial relationship to a pixel withresult of image segmentation is a set of regions that value J in the input imagecollectively cover the entire image, or a set ofcontours extracted from the image. After median The Following GLCM features werefilter, Watershed Transformation Segmentation extracted in our research work [10]:Angular Second(WTS) is used to identify the suspicious region or Moment (F1), Contrast (F2), Correlation (F3), Summicro-calcification. An example of image Of Squares: Variance (F4), Inverse Differencesegmentation for the approach is given in Figure 2. Moment (F5), Sum Average(F6), Sum Variance(F7), Sum Entropy(F8), Entropy(F9), Difference Variance(F10), Difference Entropy(F11), Information Measures Of Correlation-I(F12), Information Measures Of Correlation-II(F13), The Maximal Correlation Coefficient(F14). VII. Feature Selection using Unsupervised Relative Reduct (USRR) In [11] [14], a FS method based on a relative dependency measure was presented. The technique was originally proposed to avoid the calculation of discernability functions or positive regions, which (a) (b) (c) can be computationally expensive withoutFigure 2.Example for Watershed Transformation optimizations. The authors replaced the traditionalSegmentation in the image mdb003.pgm. (a) Original rough set degree of dependency with an alternativeimage, (b) Preprocessing image, (c) Process of measure, the relative dependency. The USRRWatershed segmentation method. algorithm 1 is given below. Algorithm 1.Shows an Unsupervised Relative ReductVI. Feature Extraction The texture coarseness or fineness of animage can be interpreted as the distribution of the VIII.Classificationselements in the description matrix. The textural The classifier tool WEKA [12] is opendescription matrices GLCM are created for each source java based machine-learning workbench thatsegmented mammogram mage for defined distance can be run on any computer in which a java run timeand direction. The Haralick features are extractedfrom the textural description matrices. A feature Relative reduct (C, D)value [17] is a real number which encodes some C, the set of all conditionaldiscriminatory information about a property of an features;object. It may not always be obvious what type of D, the set of decision features;information or feature is useful for a particulardetection task. The texture analysis matrix itself does (1) R ← Cnot directly provide a single feature that may be used (2) ∀a ∈Cfor texture discrimination. Instead, the matrix can be (3) if (κR −{a} (D) 1)used as a representation scheme for the texture image (4) R ← R −{a}and the features are computed. The features based onthe distribution matrices should therefore capture (5) return Rsome characteristics of textures such as homogeneity,coarseness, periodicity and others. environment is installed. It brings together many machine learning algorithm and tools under aA. Gray-Level Co-occurrence Matrix common frame work. The WEKA is a well knownFeatures package of data mining tools which provides a It is a statistical method that considers the variety of known, well maintained classifyingspatial relationship of pixels is the gray-level co- algorithms. This allows us to do experiments withoccurrence matrix (GLCM), also known as the gray- several kinds of classifiers quickly and easily. The tool is used to perform benchmark experiment. 1848 | P a g e
    • R.Poongothai, I.Laurence Aroquiaraj, D.Kalaivani / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1846-1851IX.Results and Discussions The proposed algorithm is simulated inMATLAB and applied on several test images.Theresults of Mean Absolute Error (MAE) and HausdroffDistance Measures for all the proposed two methodswere compared with each other.The pectral muscleregion removal from the mammogram is had by ourthree proposed method.Two methods wereCCL(connected component labeling)+Fuzzy method,Straightline with Fuzzy.These two methods werecompared with each other by error measure anddistance measure.The Mean Absolute Error iscalculated for some 14 mammogram images and it isshown in given figure (3).The Hausdroff Distance iscalculated for same 14 Pectral removal mammogramimages and it is shown in figure(4). For the total Figure (4) shows Performance Analysis ofnumber of MIAS images,pectoral removed images Hausdroff Distance Measureswere detected with each method and the averagevalue is calculated for MAE and hausdroff distance Total Average for Average forwhich in the Table(1).Then segmentation can be number of Straight line CCL with Fuzzyperformed using watershed segmentation Mammogr with Fuzzytransformation and 14 Haralick features are extracted am images MAE H.D MAE H.Dusing gray-level co-occurrence matrix (GLCM). The takenbest features can be selected based on relative (MIASdependency measure that is Unsupervised Relative Database)Reduct (RR) method. The selected features areclassified by using WEKA classification. It is found 322 0.7198 2652. 0.5164 1670.that the classification accuracy increased and 57 30 71 528decreased mean absolute error, which is shown infigure (5) and (6) .The classification results, clearlyindicates that the method selects useful features Table (1) shows Performance Analysis for thewhich are of comparable quality. average of MAE and Hausdroff Distance Measures Figure (5) shows Accuracy of Classification.Figure (3) shows Performance Analysis of MeanAbsolute Error Measures X.Conclusion This paper presents a method to detect pectoral muscle and segment the masses from the mammography. The proposed work was done using Mini-Mias mammogram database. Here we have presented several aspects of image processing techniques that can be applied for detection of pectoral muscle in digital mammography.To removing background noise from the original image 1849 | P a g e
    • R.Poongothai, I.Laurence Aroquiaraj, D.Kalaivani / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1846-1851we used median filter,then pectoral muscle removal Cancer Screening.CA Cnacer J. Clin. 60,using Fuzzy with connected component labeling. 99–119, 2010. [6] Lee, C.-S., Kanmaz, T.J.: “Recent Controversies in Cancer Screening Recommendations” US Pharm., vol. 35(11), pp. 3–8 (2010) (oncology Supply). [7] In 2010,Ali Cherif Chaabani, Atef Boujelben,Adel Mahfoudhi,Mohamed Abid, International Journal of Digital Content Technology and its Applications, Volume 4,Number 3,June 2010. [8] Roshan Dharshana Yapa and Koichi Harada “Connected Component Labeling Algorithms for Gray-Scale Images and Evaluation of Performance using Digital Mammograms”, Department of InformationFigure (6) shows Mean Absolute Error Value. Engineering,Hiroshima University,japan. International Journal of Computer Science In this paper, we have considered the and Network Security, Vol.8 No.6, Juneproblem of detecting cancer masses by the 2008.application of simple thresholding followed by [9]. Suckling J, Parker J, Dance D, Astley S,connected component labeling with fuzzy and an Hutt I, Boggis C, et al. “The mammographicalgorithm to remove artifacts in digital mammograms images analysis society digital mammogramusing morphological open operation followed by database”, 2nd International Workshop onreconstruction. Further the pectorial muscle was Digital Mammography , York, Englandremoved successfully using simple thresholding and ,Exerpta Med 1994;1069:pp375-378.raster scan methods. Then features selected based on [10]. R. M. Haralick, K. Shanmugan, I. Dinstein,relative reduct method and classification using “Textural features for image classification”,WEKA classifier. Finally, as a future work we will IEEE Trans. Syst., Man, Cybern., Vol. 3(6),consider more samples and mass shape spectral 1973, pp. 610–621.identifications in mammograms using near sets. [11]. J. Han, X. Hu, and T.-Y. Lin, “Feature subset selection based on relativeReferences dependency between attributes,” in Proc. of [1] S. Don, Eumin Choi, and Dugki Min,Breast the 4th International Conf. on Rough Sets Mass “Segmentation in Digital and Current Trends in Computing, Uppsala, Mammography Using Graph Cuts” ICHIT pp. 176–185, 2004. 2011, CCIS 206, pp. 88–96, Springer-Verlag [12]. Holmes, G., Donkin, A., Witten, I.H.: Berlin Heidelberg 2011. “WEKA: a machine learning workbench.” [2] Aswini Kumar Mohanty, Mrs. Swasati In: Proceedings Second Australia and New Sahoo ,Mrs. Arati Pradhan, Saroj Kumar Zealand Conference on Intelligent Lenka, “Detection of Masses from Information Systems, Brisbane, Australia, Mammograms Using Mass shape pp. 357-361 (1994). Pattern”,Vol 2 (4), pp.1131-1139 [13]. M. Vasantha “Medical Image Feature, ,ISSN:2229-6093, IJCTA July-August 2011. Extraction, Selection And Classification” [3] Sara Dehghani and Mashallah Abbasi ,European Journal of Scientific Research Dezfooli, “A Method for Improve Vol. 2(6), 2071-2076, ISSN 1450-216X, Preprocessing Images Mammography” Vol.50 No.4 (2011), pp.543-553,2010. International Journal of Information and [14]. R. Roselin, K. Thangavel, C. Velayutham, Education technology, Vol.1, No.1, April “Fuzzy Rough Feature Selection for 2011 ISSN: 2010-3689. Mammogram Classification”, International [4] T.S.Subashini, V.Ramalingam, S.Palanivel, Journal of Electronic Science and “Pectoral muscle removal and Detection of Technology (JEST), Vol. 9, No. 2, pp 124- masses in Digital Mammogram using CCL”, 132 ,June 2011. International Journal of computer [15]. K.Thangavel,M.Karnan, R.Sivakumar, A. Applications (0975-8887) Volume 1 – KajaMohideen “Ant Colony System for No.6,2010. Segmentation and Classification of [5] Smith, R.A., Cokkinides, V., Brawley, Microcalcification in Mammograms”, O.W.: Cancer Screening in United State, International Journal on Graphics, Vision 2010: “A review of current American and Image Processing, GVIP,P1150619001, Cancer Society”, Guidelines and Issues in 07, Special Issues on Mammograms, ISSN : 1850 | P a g e
    • R.Poongothai, I.Laurence Aroquiaraj, D.Kalaivani / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1846-1851 1687-398X-1687-3998, pp. 1-12, 2005.[16]. M. A. Duarte, A. V. Alvarenga, C. M. [25]. Jie Wei,”Shape Indexing and Recognition Azevedo, A. F. C. Infantosi, W. C. A. Based on Regional Analysis”, IEEE Pereira “Automatic Microcalcifications TRANSACTIONS ON MULTIMEDIA, VOL. Segmentation Procedure Based onOtsu‟s 9, NO. 5, pp.1-13, AUGUST 2007. Method and Morphological Filters”, RIO [26]. http://www.imaginis.com/breast DE JANEIRO, BRAZIL IEEE 102 ISBN: health/breast-cancer-statisticson- 978-1-61284-918-8 Vol. 11,pp. 103- incidence-survival-and-screening-2#breast- 108,IEEECatalog Number: CFP1118G-ART cancer-casesdeaths-per-year-u-s-and-world). 2011. [27]. Breast Cancer in Men. A complete patient‟s[17]. Shu-Ting Luo&Bor-WenCheng“Diagnosing guide.http://www.breastdoctor. Breast Masses in Digital Mammography com/breast/men/ cancer.htm. Using Feature Selection and Ensemble [28]. Breast Cancer in Men. Male breast cancer Methods” J Med Syst DOI 10.1007/s10916- information center.http://interact.withus. 010-9518-8 © Springer Science&Business com/ interact/ mbc/. Media, LLC, Vol .34, 2010. [29]. M. Karnan, K. Thangavel, “Automatic[18]. Dhirendra Pal Singh and Ashish Khare,” detection of the breast border and nipple Automatic Image Segmentation using position on digital mammograms using Threshold Based Parameter and Genetic genetic algorithm for asymmetry approach Algorithm”, International Journal of to detection of microcalcifications”. Research and Reviews in Computer Science Computer Methods and Programs in (IJRRCS) ,Vol. 3, No. 3, pp. 1672-1677, Biomedicine Vol. 87, no.1, pp 12-20, 2007. ISSN: 2079-2557 © Science Academy [30]. T.O.Gulsrud, “Texture analysis of Digital Publisher, United Kingdom, June 2012. Mammograms,” Ph.D. Thesis, Aalborg[19]. Guillaume Kom, Alain Tiedeu, Martin Kom, University, Stavanger, USA, pp.30-32, John Ngundam,” a comparative evaluation 2000. of two algorithms of detection of masses on [31]. D. C. He, L. Wang, „Texture classification mammograms”, Signal & Image using texture spectrum,” Pattern Processing: An International Journal Recognition, vol. 23, pp 905-910, 1990. (SIPIJ), Vol.3, No.1, February 2012. [32]. M. H. Horng, “Texture feature coding[20]. C. Karthikeyan , Dr. B. Ramadoss.,” Image method for texture classification,” Opt. Registration and Segmentation for the Eng., Vol 42 (1), pp. 228-238, 2003. Quantitative Evaluation of Complex [33]. J.S. Suri, R.M. Rangayyan: Recent Medical Imaging and Diagnosis over Head Advances in Breast – A Systematic Review “ ,International Imaging,“Mammography, and Computer- Journal Comp. Tech. Appl., Vol 2 (4), 1058- Aided Diagnosis of Breast Cancer”, SPIE, 1061 IJCTA ,pp.1 1058-1061 ,ISSN:2229- Bellingham, Washington, USA, 2006. 6093, JULY-AUGUST 2011. [34]. American College of Radiology (ACR):[21]. JelenaBozek, Mario Mustra, KresimirDelac, ACR Breast Imaging Reporting and Data and MislavGrgic “A Survey of Image System, 4th Edition, Reston, USA, 2003 Processing Algorithms in Digital [35]. E.S. de Paredes: Atlas of Mammography, Mammography” M. Grgic et al. (Eds.): Rec. 3rd Edition, Lippincott Williams & Wilkins, Advan. inMult. Sig. Process. andCommun., Philadelphia, USA, 2007. SCI 231, pp. 631–657,© Springer-Verlag [36]. I. H. Witten and E. Frank, Data Mining, Berlin Heidelberg 2009. Practical Machine Learning Tools and[22]. Vibha S. Vyas and Priti Techniques, 2nd ed. San Francisco: Morgan P.Rege,“Malignancy Texture Classification Kaufmann, 2005. in Digital Mammograms based on [37]. M. Heath, "The Digital Database for Chebyshev Moments and Logpolar Screening Mammography",Proceedings of Transformation”, ICGST-BIME Journal, the 5th International Workshop on Digital Volume 7, Issue 1, pp.29-35, December Mammography, Toronto, Canada, pp. 212- 2007. 218, 2000.[23]. J.G. Zhang, T. Tan, “Brief Review of [38]. L.M. Bruce, R.R. Adhami: "Classifying Invariant Texture Analysis Methods, in Mammographic Mass Shapes Using the Pattern Recognition”, Vol. 35, No.3, Wavelet Transform Modulus-Maxima pp.735-747, 2002. Method", IEEE Transactions on Medical[24]. Jan flusser, “moment invariants in image Imaging, Vol. 18, Issue 12, pp. 1170-1177, analysis,transactions on engineering, 1999. computing and technology”, Vol.11, pp. 196-201, February, 2006. 1851 | P a g e