Letters

A FULLY AUTOMATED SCHEME FOR BREAST DENSITY ESTIMATION AND ASYMMETRY DETECTION OF MAMMOGRAMS

Description
This paper presents a fully automated scheme for breast den- sity estimation and asymmetry detection on mammographic images. Image preprocessing and segmentation techniques are first applied to the image, in order to extract the feature s for the
Categories
Published
of 5
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  A FULLY AUTOMATEDSCHEME FOR BREASTDENSITY ESTIMATIONANDASYMMETRYDETECTION OF MAMMOGRAMS Stylianos Tzikopoulos, Harris Georgiou, Michael Mavroforakis, Sergios Theodoridis National and Kapodistrian University of Athens, Dept. of Informatics and Telecommunications,Panepistimiopolis, Ilissia, Athens 15784, Greece, { stzikop, xgeorgio, mmavrof, stheodor } @di.uoa.gr ABSTRACT This paper presents a fully automatedscheme for breast den-sity estimation and asymmetry detection on mammographicimages. Image preprocessing and segmentation techniquesare first applied to the image, in order to extract the features for the breast density categorization. Also a new fractal-related feature is proposed for the classification. The clas-sification to 3 classes is realized according to classificationand regression trees (CARTs). The same segmentation result is used to extract a set of new statistical features for eachbreast; thedifferenceof thesefeaturevalues,betweenthetwoimages of each pair of mammograms, are estimated and theasymmetric pairs are detected according to a modified ver-sion of k-nearest neighbor classifier. This composite method has been implemented and applied to miniMIAS database,consisting of 322 mediolateral oblique (MLO) view mammo-grams, obtained via a digitization procedure. The results arevery promising, showing equal or higher success rates com- pared to other related algorithms in the literature, despitethe fact that some of them use only small portions of the spe-cific database. In contrast our methodology is applied to thecomplete datatabase. 1. INTRODUCTION Breast cancer, i.e., a malignant tumor developed from breastcells, is considered to be one of the major causes for the in-crease in mortality among women, especially in developedcountries. Morespecifically, breastcanceris thesecondmostcommon type of cancer and the fifth most common cause of cancer death [13].While mammographyhas been proved to be the most ef-fective and reliable method for early breast cancer detection[15], the large number of mammograms, generated by popu-lation screening, must be interpreted and diagnosed by a rel-atively small number of radiologists. This is also one of thereasonswhyit is widelyacceptedtodaythatautomatedCom-puter Aided Diagnosis (CAD) systems are starting to play animportant role in modern medical practices.Most of the CAD systems try to detect abnormalitiesbased on a single mammographic image and on an objec-tive abstract model of the abnormalities. However, there isa high correlation between high breast parenchymal densityand high risk of breast cancer [20]. Thus, mammographicimages with high breast density value should be examinedmore carefully by the radiologists, creating a need for auto-matic breast parenchymal density estimation algorithms. In[11], suchalgorithmsintheliteraturearepresentedandanewtechnique, introducing a histogram distance metric, achievesgood results. Some existing algorithms, e.g., [2, 14], usethe texture information of mammograms, in order to extractmore features for the breast density estimation.Radiologists also pay attention to possible asymmetriesbetween the left and the right breast in a pair of mammo-grams, as they can provide clues about the presence of earlysigns of tumors [8]. In order to help the radiologists, manyCAD systems analyze the images of a mammogram pair anddetect asymmetric regions by applying some type of align-ment and direct comparison [21]. In [6], a new directionalanalysis method is proposed, using Gabor wavelets, in orderto detect possible asymmetries.In this work, a new breast parenchymal density estima-tion algorithm is proposed, using segmentation, first orderstatistics and fractal analysis of the mammographic imagefor the extraction of new statistical features, while the clas-sification task is performed using Classification and Regres-sion Trees (CARTs). Furthermore, a new algorithm is pro-posed for breast asymmetry detection, using the feature val-ues already extracted from the breast parenchymal densityestimation step, using a modified version of k-nearest neigh-borclassifier. Both techniquesachievehighsuccess rates, of-ten higher than the corresponding values of other algorithmsin bibliography,while they use simpler and faster feature ex-traction methods.The rest of this paper is organized as follows. In sec-tion 2, the mammographicimage database used is presented.The breast parenchymal density estimation method and theasymmetry detection algorithm are described in section 3.Section 4 presents the results obtained by the two proposedalgorithms and, finally, the discussion and conclusions arepresented in section 5. 2. DATASET The new methodology presented in this work was appliedon miniMIAS database [16], available online freely for sci-entific purposes and consisting of 161 pairs of mediolat-eral oblique (MLO) view mammograms. The images of thedatabase srcinated as the product of a film-screen mammo-gramprocessintheUnitedKingdomNationalBreast Screen-ing Program. The films were digitized and the correspond-ing images were annotated according to their breast densityby expert radiologists, using three distinct classes: Fatty (F)(106 images), Fatty-Glandular (G) (104 images) and Dense-Glandular (D) (112 images), similar to [12]. Any abnormal-ities were also detected and described, including calcifica-tions, well-defined, spiculated or ill-defined masses, archi-tectural distortions or asymmetries. Each pair of images of the database is annotated as Symmetric (146 pairs) or Asym-metric (15 pairs). The severity of each abnormality is pro-vided, i.e., benignancy or malignancy. 17th European Signal Processing Conference (EUSIPCO 2009)Glasgow, Scotland, August 24-28, 2009 © EURASIP, 20091869  A typical mammographic image is shown in figure 1a.The presence of high noise is readily observed; this makesthe segmentation of the image a demanding task. More-over, speckle noise was added through the srcinal digitiza-tion processing of the film mammograms. The srcinal 0.2mm/pixel images were resized to 0.4 mm/pixel, as in [10]and [11], in order to reduce the required computational time.The initial bit depth of 8 bits was preserved. 3. METHODS3.1 Breast Density Estimation 3.1.1 Image Preprocessing The noise of the image, e.g., high intensity scanning labels,or tape artifacts, is detectedand excludedfromthe remainingprocessing, using the same concept as in [11]. Figure 1bshows an example of this process.In order to estimate the breast boundary, the algorithmin [11] was implemented. The key idea of the algorithm isthat the skin-air boundary is the smoothest section of identi-cal pixels near the breast edge, detected using a thresholdingtechnique. A result of the algorithm is presented in figure 1c.The pectoral muscle, which is the high-intensity trian-gular region across the upper posterior (left) margin of theimage, appeared only in MLO view of left-breast mammo-grams, is detected according to [10], with the modificationsand improvementsof [18]. An example is presented in figure1d.Besides the noise segmentation techniques already pre-sented, image processing techniques are also applied for theimprovement of the overall image quality. Specifically, agaussian smoothing filter [7] with variable kernel size  hsize and standard deviation  sigma  is applied on each image, inorder to remove the noise. Subsequently, an unsharp filter[7] withmask   h UNSHARP  =  11 + a  ·   − a a − 1  − aa − 1  a + 5  a − 1 − a a − 1  − a  of variable parameter  alpha  is applied for edge enhancement.The above parameters were automatically tuned accordingto the following scheme. The following values were given tothevariablesand, foreachcombinationofvalues,thesuccessrate of the breast density estimation technique was recorded: •  hsize : 3x3, 5x5, 7x7, 9x9, 11x11 (pixels x pixels) •  sigma : 0.1, 0.4, 0.7, 1.0 •  alpha : 0.1, 0.4, 0.7, 1.0The values that achieved the best success rate were the  hsize 7x7,  sigma  0.4 and  alpha  0.7; these best values were usedas the baseline for enhancing all the images in the databaseprior to any breast segmentation and parenchymal analysis. 3.1.2 Feature Estimation The previously proposed methodology was applied to eachmammogram of the miniMIAS database and the results areillustrated in figure 1, showing: •  The initial  I   image (figure 1a). •  The background area, labels and artifacts have been ex-cluded, to obtain the  Back   area (figure 1b). •  The human-tissue  HuT   area (figure 1c), which has beenobtainedafterextractingbackground,labels,artifacts andnoise from the initial image. (a)  I   area (b)  Back   area (c)  HuT   area (d)  BrT   area Figure 1: a) Initial Image  I  , b) background  Back  , c) tissue-rich area  HuT   and d) breast tissue area  BrT  •  The segmented breast tissue  BrT   area (figure 1d), whichhas been obtained after extracting the pectoral musclefrom the human-tissue  HuT   area.In order to analyze and model the overall noise levels inthe image,the meanandvarianceof thepixelintensityvaluesare estimated in the  Back   area (no tissue or artifacts), usingequations (1)-(2): F  1  =  µ   Back   = ∑ ( i ,  j ) ∈  Back   I  ( i ,  j )  N  (  Back  )  (1) F  2  = σ  2  Back   = ∑ ( i ,  j ) ∈  Back  (  I  ( i ,  j ) − µ   Back  ) 2  N  (  Back  )  (2)where  N  (  R )  is the number of pixels in region  R .Then estimate the synthetic features  F  3  and  F  4  for thebreast tissue (  BrT  ) area, using equations (3)-(4): F  3  =  S   BrT   N  (  BrT  )  (3) F  4  =  P  BrT  µ  2  BrT  (4)where  S   BrT   is the surface and  P  BrT   the power of the  BrT   areaand can be found according to equations (5)-(6) S   BrT   =  ∑ (  x ,  y ) ∈  BrT   I  (  x ,  y )+ 1 + |  I  (  x + 1 ,  y ) −  I  (  x ,  y ) | +  |  I  (  x ,  y + 1 ) −  I  (  x ,  y ) |  (5) P  BrT   =  ∑ (  x ,  y ) ∈  BrT  |  I  (  x ,  y ) | 2 (6)Next, an algorithm for the computation of the fractal-related feature,based on the powerspectrum[7] of the imageis provided. The initial image is resized to the lower resolu-tion of 1.6mm/pixel, after placing black (zero-valued) pix-els to the non-  HuT   area. The absolute values of the Fouriertransform of the derived image are estimated and averagedover the four quarters. The estimated image is cropped tobecome square and the logarithmic values over the main di-agonal of the image are extracted. An exponential function  f   (  x ) =  A exp (  Bx ) + C   is fitted to the extracted data and thefeature  F  5  =  B  is obtained,as the feature related to the fractalexponent of the texture of the human tissue [9].Next, the human tissue  HuT   is used to perform the min-imum cross entropy (MCE) thresholding [4] three times, ac-cording to the following scheme: •  T   is the (baseline) threshold derived from MCE at graylevel range  1 , 2 8 − 1  1870  (a) Mask (b) x-axis histogram (c) y-axishistogram Figure 3: a) Initial Mask, b) x-axis and c) y-axis histogramvalues corresponding to the left and the right mammogramsdiffer significantly. Suppose that for the left breast mam-mogram we have estimated the feature vector  f   and for thecorrespondingright breast mammogramthe feature vector  g .Then, construct the following differential features of equa-tions (9)-(11) that can be used to detect possible asymmetrybetween a pair of mammographic images: F   ASYMMD 1 − 38  =  |  f  i  − g i | max (  f  i , g i )  (9) F   ASYMMD 39 − 76  =  |  f  i  − g i |  (10) F   ASYMMD 77 − 114  =  |  f  i − g i | 3 (11)where 1  ≤  i  ≤  38, resulting to a feature space of 114 featuresin total. 3.2.2 Classification For the classification of a pair of mammograms accordingto a possible asymmetry, a modified version of the typicalk-nearest neighbor classifier is used. The classifier imple-mented is described below. Consider a two class problemwith classes  C  1  and  C  2 , containing  N  1  and  N  2  samples re-spectively. For an unknown input pattern,  w , estimate the  k  nearest neighbors  n i , 1  ≤  i  ≤  k  , according to Euclidean Dis-tanceandthencalculatethevaluesofthefollowingvariables: sum 1  ( w ) =  N  2  N  1 · k  ∑ i = 1 ni ∈ C  1 1 d  i (12) sum 2  ( w ) =  N  1  N  2 · k  ∑ i = 1 ni ∈ C  2 1 d  i (13)where  d  i  is the Euclidean Distance of the unknown pattern  w to the  n i  nearest neighbor.Then the unknown pattern is classified as: w  ∈  C  1  , if sum 1  ( w ) > sum 2  ( w ) C  2  , if sum 1  ( w ) < sum 2  ( w )  (14)The previous classifier is similar to the a standard k-nnclassifier. The difference lies in that the confidence value of each class is multipliedwith a constant term, in order to copewith the class imbalance problem [17]. If, for example, class C  1  is oversampled, the constant  N  2  N  1 ( < 1 )  is multiplied with sum 1 , resulting to a different weight of the patterns of eachclass.BREASTDENSITYTRUE CLASSF G DPREDICTEDCLASSF 88 (88) 21 (12) 7 (3)G 10 (12) 52 (59) 28 (37)D 8 (6) 31 (33) 77 (72)Table 3: Results of the proposed breast density estimationalgorithm. Values inside parentheses are the results obtainedwhen using the manual segmentation method. 4. EXPERIMENTS ANDRESULTS The method for evaluating the algorithms is the leave-one-out, which is one of the most commoncross-validationmeth-ods [17]. 4.1 Breast Density Estimation Theproposedmammographicbreastdensityestimationalgo-rithmwas testedonalltheimagesoftheminiMIASdatabase,fully annotated according to 3 breast density classes. Notethat masks capable of extracting the background, obtainedby manual segmentation of the tissue-related areas [19] havebeen used. Thus, it was possible to compare the results de-rived by the fully automated and the manually segmentatedtechniques, as it is presented in table 3. Values inside paren-theses were the corresponding values, using the manual seg-mentation. For the evaluation of the algorithm the work in[11] was used, where the Closest Point Distance algorithmproposed achieved 66.15% success rate, while a previouswork [1] reported 65%, when applied to a subset of the min-iMIAS database. The results of our method achieved a suc-cess rate of 67.32%, using the fully automatic segmentationmethod. As expected, when using the manual segmentationthe results were better (68.01%) since the feature extractionprocedureused slightlybetter segmentationdata of the mam-mographic breast area. 4.2 Asymmetry Detection The proposed asymmetry detection algorithm was applied toall the images of the database, fully annotated as symmet-ric (SYMM) or asymmetric (ASYMM). The features wereprocessed through univariate significance analysis, specifi-cally T-test [5], resulting to a feature vector of predefinedlength 18. The results of the algorithm are shown in table 4.Similarly to breast density estimation algorithm, the resultsderived by the fully automated and the manual segmentationtechniquesarepresented. Valuesinparenthesesarethecorre-sponding results when using the manual segmentation tech-nique. Forthe evaluationofthe algorithmthe workpresentedat [6] was used, where an asymmetrydetection technique us-ing Gabor wavelets was presented and tested on 80 imagesof the miniMIAS database, achieving an average classifica-tion accuracy of 74.4%. The results obtained were for themanual segmentation 73.91% and for the automatic method70.19% However, note that our method is computationallysimplerandmoreimportantlyit is basedonquantitiesusedin4.1. Thus our method addresses the tasks of mammographicbreast density estimation and asymmetry detection in an uni-fying context. 1872  BREASTPAIRTRUE CLASSSYMM ASYMMPREDICTEDCLASSSYMM 208 (218) 12 (10)ASYMM 84 (74) 18 (20)Table 4: Results of the proposed asymmetry detection al-gorithm. Values inside parentheses are the results obtainedwhen using the manual segmentation method. 5. DISCUSSION AND CONCLUSION The results of this new method for mammographic breastdensity estimation and asymmetry detection were analyzedand evaluated against all the images of the miniMIASdatabase. The high level of noise of the images, due to thedigitization process, has made the segmentation process aneven harder classification task; however the success rate re-mains high when using the manual ground truth segmenta-tion technique and close to the results produced when usingthe fully automated segmentation technique [18].The proposed algorithm for  mammographic breast den-sity estimation  achieves better results than the work at [11]and similar results as the work at [14], which uses only asmall portion of the miniMIAS. The work at [2] achieveshigher values of success, but it uses textural features, whichare computationally very expensive. The work we proposeuses simple first order statistics features and a new techniquefor the power spectrum estimation, making it suitable forreal-time applications.The  asymmetry detection  scheme uses the segmentationalready obtained via the breast density estimation procedure.It achieves a success rate similar of the bibliography, al-though it uses all the images of the miniMIAS database, in-stead of a small subset, as in [6]. Therefore our experimentalresults are considered more reliable. Furthermore, the use of the modified version of the k-nn algorithm has been proveda simple yet effective way to overcome the problem of theimbalanced classes. REFERENCES [1] L. Blot and R. Zwiggelaar. Background texture extrac-tion for the classification of mammographic parenchy-mal patterns. In  Medical Image Understanding and  Analysis , pages 145–148, 2001.[2] A. Bosch, X. Munoz, A. Oliver, and J. Martı. Mod-eling and Classifying Breast Tissue Density in Mam-mograms. In  Proceedings of the 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition-Volume 2 , pages 1552–1558. IEEE Com-puter Society Washington, DC, USA, 2006.[3] L.Breiman.  ClassificationandRegressionTrees . Chap-man & Hall/CRC, 1998.[4] A. D. Brink and N. E. Pendock. Minimum cross-entropy threshold selection.  Pattern Recognition ,29(1):179–188,1996.[5] W.W. CooleyandP.R. Lohnes.  MultivariateDataAnal- ysis . John Wiley & Sons Inc, 1971.[6] RJ Ferrari, RM Rangayyan, JEL Desautels, andAF Frere. Analysis of asymmetry in mammogramsvia directional filtering withGabor wavelets.  Medical Imaging, IEEE Transactions on , 20(9):953–964,2001.[7] R.C. Gonzalez and R.E. Woods.  Digital Image Pro-cessing . Prentice Hall, 2007.[8] M.J. Homer.  Mammographic Interpretation: A Practi-cal Approach . McGraw-Hill Companies, 1991.[9] LM Kaplan. Extended fractal analysis for texture clas-sification andsegmentation.  Image Processing, IEEE Transactions on , 8(11):1572–1585,1999.[10] S.M. Kwok, R. Chandrasekhar, Y. Attikiouzel, andMT Rickard. Automatic pectoral muscle segmenta-tion on mediolateral oblique view mammograms.  Med-ical Imaging, IEEE Transactions on , 23(9):1129–1140,2004.[11] M. Masek, Electronic University of Western AustraliaSchool of Electrical, Computer Engineering, and Uni-versity of Western Australia Centre for Intelligent In-formation Processing Systems.  Hierarchical Segmen-tation of Mammograms Based on Pixel Intensity . PhDthesis, 2004.[12] M. Mavroforakis, H. Georgiou, N. Dimitropoulos,D. Cavouras, and S. Theodoridis. Significance anal-ysis of qualitative mammographic features, using lin-ear classifiers, neural networks and support vector ma-chines.  European Journal of Radiology , 54(1):80–89,2005.[13] R.M. Nishikawa. Current status and future directionsof computer-aided diagnosis in mammography.  Com- puterizedMedicalImagingandGraphics ,31(4-5):224–235, 2007.[14] A. Oliver, J. Freixenet, A. Bosch, D. Raba, andR. Zwiggelaar. Automatic classification of breast tis-sue. In  Iberian Conference on Pattern Recognition and  Image Analysis , pages 431–438. Springer, 2005.[15] MKJ Siddiqui, M. Anand, PK Mehrotra, R. Sarangi,and N. Mathur. Biomonitoring of organochlorines inwomen with benign and malignant breast disease.  En-vironmental Research , 98(2):250–257,2005.[16] J. Suckling, J. Parker, D. Dance, S. Astley, I. Hutt,C. Boggis, I. Ricketts, E. Stamatakis, N. Cerneaz,S. Kok, et al. The Mammographic Image AnalysisSociety Digital Mammogram Database. In  Exerpta Medica.InternationalCongressSeries , pages375–378,1994.[17] S. Theodoridis and K. Koutroumbas.  Pattern Recogni-tion . Academic Press, 4th edition, 2009.[18] St. Tzikopoulos, H. Georgiou, M. Mavroforakis,N. Dimitropoulos, and S. Theodoridis. A fully au-tomated complete segmentation scheme for mammo-grams. In 16thInternationalConferenceonDigitalSig-nal Processing, Santorini, Greece , July 2009.[19] M.A. Wirth.  MIAS Mask Database . University of Guelph, Canada, 2005.[20] JN Wolfe. Risk for breast cancer development deter-mined by mammographic parenchymal pattern.  Can-cer  , 37(5):2486–92,1976.[21] F.F. Yin, M.L. Giger, K. Doi, C.J. Vyborny, and R.A.Schmidt. Computerized detection of masses in digitalmammograms: Automated alignment of breast imagesand its effect on bilateral-subtraction technique.  Medi-cal Physics , 21:445–452, 1994. 1873
Search
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks