Weitere ähnliche Inhalte Ähnlich wie A Novel Method for Detection of Architectural Distortion in Mammogram (20) Mehr von IDES Editor (20) Kürzlich hochgeladen (20) A Novel Method for Detection of Architectural Distortion in Mammogram1. 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.com
Abstract--Among various breast abnormalities architectural stellate distortion. Karssemeijer and Brake [17] proposed a
distortion is the most difficult type of tumor to detect. When method that is based on statistical analysis of a map of pixel
area of interest is medical image data, the major concern is to orientation. If increase of pixels pointing to a region is found
develop methodologies which are faster in computation and then pixels are marked suspicious. The authors modified the
relatively noise free in processing. This paper is an extension
rejection procedure by sonka by suggesting that the gradient
of our own work where we propose a hybrid methodology that
combines a Gabor filtration with directional filters over the of mammographic image is obtained using the first derivative
directional spectrum for digitized mammogram processing. of the Gaussian with a standard deviation of 1 mm. But special
The most commendable thing in comparison to other care must be taken to choose appropriate values of standard
approaches is that complexity has been lowered as well as the deviation parameters in order to make the method good
computation 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 and
Index Terms--M ammogram images, automated system, concentration index were calculated to identify suspicious
architectural distortion estimation, Directional filter, Gabor
regions. Sampat et al. [19] proposed a technique for the
filter.
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 the
the rise and is rapidly becoming the number one cancer in estimation of architectural distortion, the estimation
females pushing the cervical cancer to the second spot [1]. complexity for such approach is observed higher. Each of the
Diagnosis of huge amount of images is a very time consuming method consists of lot of mathematical calculations, analytical
process which needs a lot of expertise. So there is a need for reasoning and lot of expertise. There was a need to develop
Computer-aided diagnosis for screening mammogram which a low complex algorithm for the estimation of architectural
may help the radiologists in interpreting mammograms [2]. distortion. In this paper a hybrid method of Gabor filtration
Literature survey reveals that different algorithms, for CAD with the incorporation of directional filter is proposed for the
system 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 DETECTION
have been detected by the existing methods as well as by the
available 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 cause
architecture of the breast is distorted with no definite mass oriented texture in mammograms. The presence of
visible [12]. This includes spiculations radiating from a point architectural distortion is expected to change the normal
and focal retraction or distortion at the edge of the oriented texture of the breast. Gabor filters are used mostly in
parenchyma. It is best perceived at the interface between determining the oriented texture in image processing. Gabor
breast parenchyma and subcutaneous fat [14]. Ayres and filters have been presented in several works on image
Rangayyan [13-16] used the concept of Gabor filters and phase processing [20], [21], [22], however, most of these works are
portrait maps to characterize oriented texture patterns in related to segmentation and analysis of texture. A 2-D Gabor
mammograms to detect architectural distortion.The method function can be specified by the frequency of the sinusoid
worked 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 11
DOI: 01.IJIT.02.02. 82
2. 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 directional
Gabor 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 processing
However, in general, expansion is not applied for Gabor to estimate directional flow was introduced by Bamberger
wavelets, since this requires computation of bi-orthogonal and Smith [25].The filters has overcome the limited directional
wavelets, which may be very time-consuming. Therefore, selectivity of separable wavelets. It has been used in various
usually, a filter bank consisting of Gabor filters with various imaging applications such as texture classification [26], image
scales and rotations is created.Let be the mother denoising [27], fingerprint image enhancement and
Gabor wavelet, then this self-similar filter dictionary can be recognition [28].For the extracted average Gabor output the
obtained 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. The
and orientation, respectively. The scale factor in (2) is meant downsampling matrices employed in the tree-structure are
to 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 from
the 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 optimization
responses 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 the
maximum information with minimum redundancy.
The obtained output for the gabor filter at different
orientation 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 the
orientation [27]. For the obtained average image for the Gabor
corresponding filter coefficent ‘h’ and added to get the total
output, 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 for
the designing of a directional filter for the estimation of the Where k is the iteration for filtration and θ is the obtained
region 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 average
filter but to find the variation in this output, to distinguish orientation, this reveals us the dominant orientation for the
between distorted and non-distorted region, directional filters given sample. To the obtained average output from Gabor
are used. In direction filters, there are banks of filters with bank we find the orientation field given by equation 5.For the
different orientation and the obtained orientation image is obtained orientation field angle we take theta value as a
passed from this bank, this filter filters the data as per the reference and create a directional filter. The obtained filtered
direction coefficient specified. Where the directions are output from the Gabor bank is the passed to this directional
matched for a particular direction that is been filtered out. We filter and as a reference, orientation field is extracted, based
can easily find a distinction in such case where we have an on this reference the orientation region is predicted.
© 2012 ACEEE 12
DOI: 01.IJIT.02.02.82
3. ACEEE Int. J. on Information Technology, Vol. 02, No. 02, April 2012
The region which has got an architectural variation will be the Gabor based estimation obtained is as illustrated in figure
differing 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 Uniform
step. 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 region
output. For the selected non-distorted region from the obtained
7) Compute the cost function of the developed filter output. magnitude image the spectral density is observed shown in
Eq. (7) figure 4-8. The observation illustrates that the spectral density
8) Repeat the iteration to converge the cost function. varies at average density of 3-4 units for the non-architectural
9) Obtained direction map at the lowest cost function is one region, where as a spectral density variation with architectural
direction 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 extracted
from the Mammographic image analysis society (MIAS).The
region is then further processed for distortion region
database contains 19 image sample showing cases of
extraction based on the directional filter based approach.
architectural distortion. For each abnormal case, the
coordinates of centre of abnormality are provided with the
radius of a circle enclosing the abnormality. The widest
abnormality corresponds to a radius of 117 pixels while the
abnormality corresponding to smallest radius corresponds
to 23 pixels. By knowing the location and proper size of
abnormality allows us to manually extract the sub images
with proper size representing the tumor affected area. The
mammogram samples are digitized to 1024 X 1024 pixels at
8bit/pixel. The various parameters chosen for the Gabor filter
Figure 4. Spectral plot for the extracted region
were having the following values viz. bandwidth 1, phase
shift as 0, theta ranging between 0-pi, orientation of 1800 at –
π/2 to π/2 Scale, the number of iteration considered is 12. To
discard irrelevant information like breast contour, region of
interest of size approximately 150x 150 pixels is extracted
using ‘imcrop’ function. The red circle in Fig 2.(a) shows the
patch depicting showing abnormality being cropped. The
spectrum realization for a mammogram sample using Gabor
filter for the distorted and non distorted region is illustrated
in figure 4. For a Mini-MIAS image sample ‘mdb115’ shown
in figure 2(a), Figure 5. Closed contour region plot for the extracted region
© 2012 ACEEE 13
DOI: 01.IJIT.02.02. 82
4. 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 Features
have 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 using
not 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 Fractals
images 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 breast
architectural distortion region is observed to be effectively
microcalcifications clusters: Merging of features rom image
detected. 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-level
19 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 14
DOI: 01.IJIT.02.02.82
5. 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 texture
Pattern 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 architectural
Imaging 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 a
architectural 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 signal
in 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 for
Assist. 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 in
mammograms,” Med. Biol. Eng. Comput., vol. 44, no. 10, pp. multidimensional filter bank design,” in Proceedings of IEEE
883–894, 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 direction
distortions around skinline and within mammary gland on filter banks: theory, design and application,” Accepted for
mammograms,” 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, Boggis
Vannier, K. Inamura, A. G. Farman, K. Doi, and J. H. C. Reiber, CRM, Ricketts I, Stamakis E, Cerneaz N, Kok SL, Taylor P, Betal
Eds. 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 2nd
distortion,” 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 375–378
SPIE, 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 15
DOI: 01.IJIT.02.02.82