A Novel Method for Detection of Architectural Distortion in Mammogram
Upcoming SlideShare
Loading in...5

A Novel Method for Detection of Architectural Distortion in Mammogram



Among various breast abnormalities architectural ...

Among various breast abnormalities architectural
distortion is the most difficult type of tumor to detect. When
area of interest is medical image data, the major concern is to
develop methodologies which are faster in computation and
relatively noise free in processing. This paper is an extension
of our own work where we propose a hybrid methodology that
combines a Gabor filtration with directional filters over the
directional spectrum for digitized mammogram processing.
The most commendable thing in comparison to other
approaches is that complexity has been lowered as well as the
computation time has also been reduced to a large extent. On
the MIAS database we achieved a sensitivity of 89 %.



Total Views
Views on SlideShare
Embed Views



0 Embeds 0

No embeds


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.

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

    A Novel Method for Detection of Architectural Distortion in Mammogram A Novel Method for Detection of Architectural Distortion in Mammogram Document Transcript

    • ACEEE Int. J. on Information Technology, Vol. 02, No. 02, April 2012 A Novel Method for Detection of Architectural Distortion in Mammogram Amit Kamra1, Sukhwinder Singh2 and V K Jain3 1 Guru Nanak Dev Engineering College, Department of IT India, Ludhiana, India. 1 amit_malout@yahoo.com 2 UIET, Panjab University Department of CSE, Chandigarh, India 2 sukhdalip@pu.ac.in 3 SLIET,Deemed to be University,Sangrur, India vkjain27@yahoo.comAbstract--Among various breast abnormalities architectural stellate distortion. Karssemeijer and Brake [17] proposed adistortion is the most difficult type of tumor to detect. When method that is based on statistical analysis of a map of pixelarea of interest is medical image data, the major concern is to orientation. If increase of pixels pointing to a region is founddevelop methodologies which are faster in computation and then pixels are marked suspicious. The authors modified therelatively noise free in processing. This paper is an extension rejection procedure by sonka by suggesting that the gradientof our own work where we propose a hybrid methodology thatcombines a Gabor filtration with directional filters over the of mammographic image is obtained using the first derivativedirectional spectrum for digitized mammogram processing. of the Gaussian with a standard deviation of 1 mm. But specialThe most commendable thing in comparison to other care must be taken to choose appropriate values of standardapproaches is that complexity has been lowered as well as the deviation parameters in order to make the method goodcomputation time has also been reduced to a large extent. On enough to consider the strong gradient.the MIAS database we achieved a sensitivity of 89 %. Matsubara et al. [18] used notion of morphology to find architectural distortion around the skin. Line structure andIndex Terms--M ammogram images, automated system, concentration index were calculated to identify suspiciousarchitectural distortion estimation, Directional filter, Gabor regions. Sampat et al. [19] proposed a technique for thefilter. enhancement of spiculations, in which a linear filter is applied I. INTRODUCTION to the Radon transform of the image. The enhanced image is filtered with radial spiculation filters to detect spiculated Breast cancer is the leading and second most mortal masses and architectural distortion.disease in women. The growth of breast cancer in India is on Though these approaches were proposed for thethe rise and is rapidly becoming the number one cancer in estimation of architectural distortion, the estimationfemales pushing the cervical cancer to the second spot [1]. complexity for such approach is observed higher. Each of theDiagnosis of huge amount of images is a very time consuming method consists of lot of mathematical calculations, analyticalprocess which needs a lot of expertise. So there is a need for reasoning and lot of expertise. There was a need to developComputer-aided diagnosis for screening mammogram which a low complex algorithm for the estimation of architecturalmay help the radiologists in interpreting mammograms [2]. distortion. In this paper a hybrid method of Gabor filtrationLiterature survey reveals that different algorithms, for CAD with the incorporation of directional filter is proposed for thesystem have been suggested and developed using wavelets detection and extraction of orientation fields in the[3],[4],[5],[6] statistical analysis [7],[8] fuzzy set theory mammogram samples.[9],[10],[11] etc. for the automation of mammogram analysis.But fewer than half of the cases of architectural distortion II. GABOR FILTERING & ORIENTATION FIELD DETECTIONhave been detected by the existing methods as well as by theavailable CAD systems. Since Breast contains several piecewise linear structure Architectural distortion is a situation where the normal such as ligaments, ducts and blood vessels that causearchitecture of the breast is distorted with no definite mass oriented texture in mammograms. The presence ofvisible [12]. This includes spiculations radiating from a point architectural distortion is expected to change the normaland focal retraction or distortion at the edge of the oriented texture of the breast. Gabor filters are used mostly inparenchyma. It is best perceived at the interface between determining the oriented texture in image processing. Gaborbreast parenchyma and subcutaneous fat [14]. Ayres and filters have been presented in several works on imageRangayyan [13-16] used the concept of Gabor filters and phase processing [20], [21], [22], however, most of these works areportrait maps to characterize oriented texture patterns in related to segmentation and analysis of texture. A 2-D Gabormammograms to detect architectural distortion.The method function can be specified by the frequency of the sinusoidworked well to detect peaks related to architectural distortion. and the standard deviations and of the Gaussian envelope.However the procedure does not give good results in case of [23].© 2012 ACEEE 11DOI: 01.IJIT.02.02. 82
    • ACEEE Int. J. on Information Technology, Vol. 02, No. 02, April 2012 architectural variation from the obtained output. Buciu and Gascadi [24] have described an approach where mammogram sample is filtered using Gabor wavelets and directionalGabor filters are directly related to Gabor wavelets, since they features are extracted at different orientation and frequencies.can be designed for a number of dilations and rotations. Directional filter estimation (DFE) for image processingHowever, in general, expansion is not applied for Gabor to estimate directional flow was introduced by Bambergerwavelets, since this requires computation of bi-orthogonal and Smith [25].The filters has overcome the limited directionalwavelets, which may be very time-consuming. Therefore, selectivity of separable wavelets. It has been used in varioususually, a filter bank consisting of Gabor filters with various imaging applications such as texture classification [26], imagescales and rotations is created.Let be the mother denoising [27], fingerprint image enhancement andGabor wavelet, then this self-similar filter dictionary can be recognition [28].For the extracted average Gabor output theobtained by appropriate dilations and rotations of directional filter is applied to extract the 2D- directional through the generating function. variations on which the region of interest is extracted. The filtered input image usually has smooth regions with texture regions variation at distorted and non-distorted region. In order to achieve the directional variations, eight channel filter bank, with a tree-structure is proposed [29],Where (x0,y0) center of the filter in the spatial domain; which composes of three stages. At each stage, the image is total number of orientations desired; and scale filtered and decimated by a two channel filter bank. Theand orientation, respectively. The scale factor in (2) is meant downsampling matrices employed in the tree-structure areto ensure that the energy is independent. Gabor wavelets in Q1, D0, and D1. At the end of stage 3, eight sub bands with (approximately) equal numbers of samples are obtained fromthe frequency domain is defined as the decomposition. For a set of obtained filter output, S = {1, 2, 3, 4, 5, 6, 7, 8} defining the directional subband with modulation index, M = 8 indicating number of directional The design strategy used is to project the filters so as to bands.ensure that the half-peak magnitude supports of the filter The design problem can be formulated as an optimizationresponses in the frequency spectrum for the non-effected problem with an objective function [30]region as compared to the region with architectural distortion.By doing this, it can ensured that the filters will capture themaximum information with minimum redundancy. The obtained output for the gabor filter at differentorientation reveals the orientation in one particular direction,an average gabor output is derived given by, (4) (6) Where I(x,y) is the gabor output obtained at each Here for each direcion the i/p is multiplied with theorientation [27]. For the obtained average image for the Gabor corresponding filter coefficent ‘h’ and added to get the totaloutput, the orientation field is estimated by [24] cost function after processing in all orientaion. The filter structure for the directional filter is defined by [30]. The average estimated orientation field is then used forthe designing of a directional filter for the estimation of the Where k is the iteration for filtration and θ is the obtainedregion which is directionally different from the original region, orientation field for a given sample.as outlined in following section. The overall implementation could be summarized as shown in Fig 1. For a given sample, we apply the Gabor filter III. WHY DIRECTIONAL FILTRATION with different orientation angle up to 180 degrees. To the Basically the orientation can be extracted from the Gabor output obtained at each orientation we evaluate the averagefilter but to find the variation in this output, to distinguish orientation, this reveals us the dominant orientation for thebetween distorted and non-distorted region, directional filters given sample. To the obtained average output from Gaborare used. In direction filters, there are banks of filters with bank we find the orientation field given by equation 5.For thedifferent orientation and the obtained orientation image is obtained orientation field angle we take theta value as apassed from this bank, this filter filters the data as per the reference and create a directional filter. The obtained filtereddirection coefficient specified. Where the directions are output from the Gabor bank is the passed to this directionalmatched for a particular direction that is been filtered out. We filter and as a reference, orientation field is extracted, basedcan easily find a distinction in such case where we have an on this reference the orientation region is predicted.© 2012 ACEEE 12DOI: 01.IJIT.02.02.82
    • ACEEE Int. J. on Information Technology, Vol. 02, No. 02, April 2012The region which has got an architectural variation will be the Gabor based estimation obtained is as illustrated in figurediffering in its orientation and based on the applied directional 2(b).filter the region is predicted. Figure 2. (a) original sample (b) Obtained magnitude image for the Gabor output The sample is processed for 1800 orientation at a linear step and a mean average of all obtained Gabor result is carried Figure 1. Proposed system architecture out.Proposed Algorithm:1) Read the input mammogram sample.2) Apply Gabor filtration with 180 orientations at Uniformstep. Eq. (3)3) Compute the average Gabor output. Eq. (4)4) Compute the orientation field θ. Eq.(5)5) Define a directional filter h (k)j for the obtained θ. Eq.(6)6) Apply the directional filter on the extracted spectral map Figure 3. Extracted Non-distorted regionoutput. For the selected non-distorted region from the obtained7) Compute the cost function of the developed filter output. magnitude image the spectral density is observed shown inEq. (7) figure 4-8. The observation illustrates that the spectral density8) Repeat the iteration to converge the cost function. varies at average density of 3-4 units for the non-architectural9) Obtained direction map at the lowest cost function is one region, where as a spectral density variation with architecturaldirection region extracted. variation is observed to be at 1-2 variation. With this regard the obtained spectral density for a non-architectural region IV. RESULTS AND DISCUSSION is comparatively 2 times higher than the architectural region (see figure 4,7). Based on this spectral variation a region of The mammogram images used in our experiment was taken architectural distortion region is estimated. The extractedfrom the Mammographic image analysis society (MIAS).The region is then further processed for distortion regiondatabase contains 19 image sample showing cases of extraction based on the directional filter based approach.architectural distortion. For each abnormal case, thecoordinates of centre of abnormality are provided with theradius of a circle enclosing the abnormality. The widestabnormality corresponds to a radius of 117 pixels while theabnormality corresponding to smallest radius correspondsto 23 pixels. By knowing the location and proper size ofabnormality allows us to manually extract the sub imageswith proper size representing the tumor affected area. Themammogram samples are digitized to 1024 X 1024 pixels at8bit/pixel. The various parameters chosen for the Gabor filter Figure 4. Spectral plot for the extracted regionwere having the following values viz. bandwidth 1, phaseshift as 0, theta ranging between 0-pi, orientation of 1800 at –π/2 to π/2 Scale, the number of iteration considered is 12. Todiscard irrelevant information like breast contour, region ofinterest of size approximately 150x 150 pixels is extractedusing ‘imcrop’ function. The red circle in Fig 2.(a) shows thepatch depicting showing abnormality being cropped. Thespectrum realization for a mammogram sample using Gaborfilter for the distorted and non distorted region is illustratedin figure 4. For a Mini-MIAS image sample ‘mdb115’ shownin figure 2(a), Figure 5. Closed contour region plot for the extracted region© 2012 ACEEE 13DOI: 01.IJIT.02.02. 82
    • ACEEE Int. J. on Information Technology, Vol. 02, No. 02, April 2012 CONCLUSION Architectural distortion is the most difficult breast abnormality to be read due to its subtle characteristics. Gabor features obtained from the ROI are passed through the directional filters, are used for identifying distorted areas. One paper that was also working on the same applications is Matsurba et al [18]. It is difficult to have a direct and fair Figure 6. Extracted architectural distorted region comparison as methodology used differs. The developed system illustrates that the proposed approach is lower in complexity for architectural distortion estimation in comparison to other proposed methods. The approach of usage of directional filter estimation based on estimated orientation field reduces the estimation iteration as the orientation field estimation over a defined direction is carried out in such an approach. The overall system sensitivity is higher for the developed system, resulting in higher estimation accuracy. REFERENCES Figure 7. Spectral plot for the extracted architectural distortion [1] Breast cancer in India www.icmr.org region [2] Baker JA,Rosen EL, Lo,JY, Gimenez EI, Walsh R, Soo Ms (2003) Computer- aided detection in screening mammography: Sensitivity of commercial CAD systems for detecting architectural distortion. Am J Roentgenol 181: 1083-1088. [3] R. J. Ferrari, R. M. Rangayyan, J.E. L. Desautels, A. F. Frere, Analysis of asymmetry in mammograms via directional filtering with Gabor wavelets, IEEE Transactions on Medical Imaging, 20(9):953-964, 2001. [4] C.Chen and G. Lee. Image segmentation using multiresolution wavelet analysis and expectation maximization (em) algorithm for digital mammography. International Journal of Imaging Systems and Technology, 8(5):491–504, 1997.Figure 8. Contour region for the extracted architectural distortion [5] T.Wang and N.Karayiannis.Detection of microcalcification n region digital ammograms using wavelets. IEEE Trans. Medical The contour regions reveals the volume of distortion we Imaging, 17(4):498–509, 1998. [6] C. H. Wei, Y. Li C. T. Li , Effective Extraction of Gabor Featureshave in given region, in case of architectural distortion region for Adaptive Mammogram Retrieval, ICME, 2007, pp. 1503-1506.the image will suddenly changes its orientation, so we will [7] B.Patel and G .R.Sinha, Early detection of breast cancer usingnot be able to develop a close contour, this can give a density self similar fractal method IJCA Trans. Medical Imaging, 10(4):39-of closed contour higher in non-defective region than in 44, 2010.defective region. The computation time retrieved for the [8] L. Bocchi, G. Coppini, J. Nori, G. Valli, Detection of Single and“mdb115’ was 8.73 sec. The system was tested on all the 19 Clustered Microcalcifications in Mammograms Using Fractalsimages and on an average the computation time was nearly Models and Neural Networks, Medical Engineering &9.45 seconds. For the evaluation of the proposed approach Physics, 26:303-312, 2004.among all the 19 test samples the region isolated for the [9] Joseph Y.Lo, Marios Gvrielides,Mia K.Markey,Jonathan L.Jesneck , Computer aided classification of breastarchitectural distortion region is observed to be effectively microcalcifications clusters: Merging of features rom imagedetected. The test data are then mixed with 15 more false processing and radiologists.Proceedings of SPIE Vol.5032.pp 882-images and the recognition of the region is carried out. Among 890.2003.all 34 tests the region for the estimated architectural distortion [10] AP. Dhawan, Y. Chitre, C. Kaiser-Bonasso, M. Moskowita,correctly recognized is observed to be 17 samples. About 17/ Analysis of Mammographic microcalcifications Using Grey-level19 classifications are manually observed correctly recognized Image Structure Features, IEEE Trans, Med. Imag., 15(3):246-259,giving the system sensitivity of about 89%. 1996.© 2012 ACEEE 14DOI: 01.IJIT.02.02.82
    • ACEEE Int. J. on Information Technology, Vol. 02, No. 02, April 2012[11] D. Brazokovic and M. Neskovic. Mammogram screening using Biological Cybernetics, vol. 55, pp. 71–82, 1986.multiresolution-based image segmentation. International Journal of [21] A.K. Jain and F. Farrokhnia, “Unsupervised texturePattern Recognition and Artificial Intelligence, 7(6):1437– 1460, segmentation using Gabor filters,” Pattern Recognition., vol. 24,1993. no. 12, pp. 1167–1186, 1991.[12] American College of Radiology (ACR), Illustrated Breast [22] Ayres FJ, Rangayyan RM, “Characterization of architecturalImaging Reporting and Data System (BI-RADS), 3rd ed. Reston, distortion in mammograms”. IEEE Eng Med Biol Mag 24(1): 59-VA: Amer. Coll. Radiol., 1998. 67.[13] http://sprojects.mmi.mcgill.ca/mammography/masses5.htm [23] Ayres FJ, Rangayyan RM,”Reduction of false positive in the[14] F. J. Ayres and R. M. Rangayyan, “Characterization of detection of architectural distortion in mammogram by using aarchitectural distortion in mammograms,” IEEE Eng. Med. Biol. geometrically constrained phase portrait model” Springer 2007.Mag., vol. 24, no. 1, pp. 59–67, Jan./Feb. 2005. [24] Loan Buciu Alexandru Gascai “Directional features for[15] F. J. Ayres and R. M. Rangayyan, “Reduction of false positives automatic tumor classification of mammogram “Biomedical signalin the detection of architectural distortion in mammograms by using processing and control (2010), DO 10.1016/j.bspc 2010.10.003.a geometrically constrained phase portrait model,” Int. J. Comput. [25] Roberto H.Bamberger and Mark J.T.Smith. A filter bank forAssist. Radiol. Surg., vol. 1, no. 6, pp. 361–369, 2007. the directional decomposition of images: Theory and design. IEEE[16] R. M. Rangayyan and F. J. Ayres, “Gabor filters and phase Transactions on Signal Processing,40(7):882-893,1992.portraits for the detection of architectural distortion in [26] T. Chen and P. Vaidyanathan, “Considerations inmammograms,” Med. Biol. Eng. Comput., vol. 44, no. 10, pp. multidimensional filter bank design,” in Proceedings of IEEE883–894, 2006. International Symposium on Circuits and Systems, May 1993,[17] N. Karssemeijer and G. M. te Brake, Detection of Stellate pp. 643–645.Distortions in Mammograms, IEEE Trans, Med. Imag., 15(5):611- [27] Y. P. Lin and P. P.Vaidyanathan, “Theory and design of two-619, 1996. dimensional filter banks: a review,” Multidimensional Systems and[18] T. Matsubara, T. Ichikawa, T. Hara, H. Fujita, S. Kasai, T. Signal Processing, vol. 7, pp. 263–330, 1996.Endo, and T. Iwase, “Automated detection methods for architectural [28] T. T. Nguyen and S. Oraintara, “Multiresolution directiondistortions around skinline and within mammary gland on filter banks: theory, design and application,” Accepted formammograms,” in Proc. 17th Int. Congr. Exhib. Comput. Assist. publication in IEEE Transaction on Signal Processing, Oct 2004.Radiol. Surg. (International Congress Series), H. U. Lemke, M. W. [29] Suckling J, Parker J, Dance DR, Astley S, Hutt I, BoggisVannier, K. Inamura, A. G. Farman, K. Doi, and J. H. C. Reiber, CRM, Ricketts I, Stamakis E, Cerneaz N, Kok SL, Taylor P, BetalEds. London, U.K.: Elsevier, Jun. 20 D, Savage J (1994) The mammographic image analysis society digital[19] M. P. Sampat, G. J. Whitman,M. K.Markey, and A. C. Bovik, mammogram database. In: Gale AG, Astley SM, Dance DD, Cairns“Evidence based detection of spiculated masses and architectural AY (eds) Digital mammography: proceedings of the 2nddistortion,” in Proc. SPIEMed. Imag. 2005: Image Process., vol. international workshop on digital mammography, Elsevier, York,5747, J. M. Fitzpatrick and J. M. Reinhardt, Eds. San Diego, CA: pp 375–378SPIE, Apr. 2005, pp. 26–37. [30] Truong T. Nguyen and Soontorn Oraintara “ A mutliresolution[20] M. R. Turner, “Texture discrimination by Gabor functions,” directional filter bank for image applications” ICASSP, pp 37-40. 2004.© 2012 ACEEE 15DOI: 01.IJIT.02.02.82