Detection of Suspicious Lesions by Adaptive

Thresholding in Mammograms
S.Allwin
#
, Y.Sobiya
*

*
Sobiya.Y
PG Student,
Infant Jesus College of Engineering,
Thoothukudi, India,

1
tsofi12@gmail.com
#
Allwin. S
Associate Professor,
Infant Jesus College of Engineering,
Thoothukudi, India,
2
allwinstephen@gmail.com


Abstract— Medical imaging is a very important branch of Image
processing. This field aims at creating tools with which doctors or
pathologist can detect a disease more easily. Over the years this
field has helped to bring down the number of deaths that occur
due to cancers and other deadly diseases. Mammography plays a
central part and is a more effective procedure for the early
detection of breast cancer and it is a reliable method for detection
of lesions in a women’s breast. The mammogram shows lumps
which is an indication of cancer. These lumps exhibit Texture
and shape features. In this paper, a novel detection algorithm,
which detects the lesions using texture and shape features, is
developed.

Index Terms- breast cancer, mammograms, masses, lesions,
thresholding.


I. INTRODUCTION

Cancer of the Breast or otherwise known as Breast cancer
is the most common form of cancer in women and it is the
second major cause of death after cervical cancer. Like all
other cancers it is caused by the uncontrollable growth of
cells. According to a medical survey 1 out of 8 women are
diagnosed with breast cancer yearly. To diagnosis this cancer,
doctors rely on imaging a womens breast using a tool called
mammogram. A mammogram fires low intensity x-rays into
the breast region by which lumps can be detected. According
to The International Agency for Research on Cancer, which is
part of the World Health Organization, there were
approximately 78,000 women per year affected by breast
cancer in India in 2001 and over 80,000 women in 2002.
Detection and diagnosis of breast cancer in its early stage
increases the chances for successful treatment and complete
recovery of the patient. Mammography is the most effective
and reliable detection method of breast cancer, and is applied
most widely in the clinic [1].
Breast cancer in mammograms, lesions can be classified as
micro calcification and space occupying. Space occupying
lesions are divided into three cases: masses, architectural
distortion (ARCH), and asymmetry (ASYM). Masses are
defined as space-occupying lesions that are described by their
shapes and margin properties. According to the shape and
boundary characteristics of masses, it can be further divided
into speculated masses (SPIC), circumscribed masses (CIRC),
and other masses (MISC). Micro calcifications are tiny
deposits of calcium that appear as small bright spots in the
mammogram. Fig. 1 shows the classification of lesions, and
Fig. 2 shows the typical examples of real space-occupying
lesions. In mammography, there are three main lesion
features: texture, shape, and gray level. SPIC and ARCH
lesions can be mainly characterized by lines radiating from the
central nucleus to their margins with oriented textural patterns.
Since CIRC lesions are mainly characterized by shape
features. For MISC and ASYM lesions, and other lesions that
are mainly characterized by gray level features, such as
brightness and gray value, gray-level feature- based detection
algorithms can obtain more comprehensive results and are
effective in mammographic mass detection, particularly the
adaptive Thresholding detection algorithms[5].



















Fig.1. Classification of lesions.

Lesion
s
Real
space
occupy
ing
Micro
calcific
ation
Asymm
etry
Archite
ctural
distorti
on
Other
masses
Speculat
ed
masses

Masse
s Circumsc
ribed
masses


Fig. 2. Types of lesions. (a) CIRC (b) SPIC (c) ARCH (d) MISC

Previously aims to detect the cancer include feature
detection using wavelet transforms. They have been used with
the combination of adaptive global thresholding segmentation
and adaptive local thresholding segmentation methods.
Histogram based adaptive global thresholding algorithm and
Window-based adaptive local thresholding method have been
employed previously for the detection part.
In this paper a novel approach is presented to detect
the lesions using adaptive tresholding with a blend other
algorithms.

II. DETECTION OF SUSPICIOUS LESIONS BY
ADAPTIVE THRESHOLDING IN MAMMOGRAMS

A. Preprocessing
The initial stages of lesions are transformed using
discrete wavelet transform. This transforms captures both
frequency and location of images. Harr wavelet transform may
be considered to simply pair up input values, storing the
difference and passing the sum.

B. Kittler Minimum Error Thresholding
Kittler minimum error thresholding computes an optimal
image threshold and relative histogram. Every possible
threshold T is got by randomization technique namely pdf. It
is interpreted as the lower boundary of the second class of
pixels rather than the upper boundary of the first class. That is,
an intensity value T and is treated as being in the same class as
higher intensities rather than lower intensities. When the
values are split, the histogram at the threshold T, the number
of pixels in the two classes are obtained. The process is
continued only if both classes contain at least one pixel. Next
the standard deviations of the classes and the criterion function
are calculated.

C. Histogram-Based Adaptive Thresholding Method
The Histogram Based Adaptive Thresholding method is
used to find the coarse segmentation to get a rough
representation of the localization of suspicious lesions and
then uses the fine segmentation to improve the rough
representation. In this paper mammograms image are wavelet
transformed and the gray-level distribution of the target and
the background regions of the images approaches to
Probability density function. The target image has higher gray
level than the background. Using these equations 1 and 2 the
pdf is calculated [10] .

) (
)
`
¹
¹
´
¦
÷ ÷ =
[
o µ
o
1
2
2
1
2
exp ) (
1 2
1
x x pb
(1)


) (
)
`
¹
¹
´
¦
÷ ÷ =
[
o µ
o
2
2
2
2
2
exp ) (
1 2
1
x x pt (2)
µ2 >µ1


Here x is a pixel value, σ1 and σ2 are the standard deviations
of the background and the target of image, and μ1 and μ2 are
the means of the background and the target of image. Let PI
(x) be the PDF of original image I, and let p(B) and p(T) be
the a priori probabilities of the background and the target of
image I. Then the Probability density function of original
image is calculated using eqn 3[11].

) ( ) ( ) ( ) ( ) ( x pt T p x pb B P x pI + =
(3)

The Bays threshold λ1 is the intersection of two solid lines
that satisfy p(B)pb(λ1) = p(T)pt(λ1). The binary image R be
the segmentation result, when the overlap between pb(x) and
pt(x) is not significant, λ2 is often close to λ1. Hence, it is
reasonable to carry out segmentation according to λ2. Eqn 4 is
used to do this step[7].

¹
´
¦
>
<
=
1 ) , ( , 1
1 ) , ( , 0
) , (
ì
ì
j i SI
j i SI
j i R
(4)


Here ( i, j) denote the pixel coordinates and SI(i, j) denotes
the pixel value of (i, j). Usually, the Bays threshold λ1 cannot
be calculated because pb(x), pt(x), and the a priori probability
of each class are unknown. Assume that λ2 is the minimum
value in pI (x).

D. Window-Based Adaptive Thresholding Method
Local segmentation is expected to give more precise
results since the global segmentation finds a coarse
localization of the suspicious lesions. In for each pixel SI (i, j)
,

a decision is made to classify it into a potential suspicious
lesion pixel or a normal pixel using the following rule. If
SI (I, j) ≥ TH(i, j) and SI
dif
≥ M
voisiP
, then SI(i, j) belongs to the
suspicious area else, SI(i, j) belongs to the normal area. In this
rule, TH(i, j) is an adaptive threshold value calculated using
equation 5[3]

SIdiff
γ.
Mvoisip
j) TH(i, + = With
SI j) min(i, SI j) max(i, SIdiff
÷ =
(5)
M
voisiP
is an average of pixel intensity in a small
window around the pixel SI (i, j), is SImax (i, j) and SImin
(i, j) are the maximum and minimum intensity values in the
large window. γ is a thresholding bias coefficient. Its value
ranges from zero to one.











































Fig 3: Structure of proposed system
IV. EXPERIMENTAL RESULTS
The data used in this work is obtained from the mini- MIAS
database of mammograms and various data was collected from
Nellai Scan center. The test has been done on more than 80
images; all images are digitized at the resolution of
1024×1024 pixels and 8- Bit accuracy (gray level). Image
(a),(d),(g) are original images, image (b), (e) and image (h) are
manual images and images (c), (f) and (i) are detected
cancerous images. The wavelet transform applied on the
original mammograms images and the singularities are
removed and generated the lesion gray-scale images.
Subsequently, the wavelet transform on the histograms (PDF
curves) removed the fluctuations.
Hence, the global local minima can be found as the
adaptive global threshold to implement the coarse
segmentation. After a convolution image the coarse
segmentation on gray-level image, it is easier to select the
adaptive local threshold to perform the fine segmentation.
Finally get the adaptive threshold value and detected the
lesions from the images.





Fig 4 Example of result showing: (a) Original image (b) Manual image




(c) Detected image

Start

Input image
Normalized
image
DWT of the image
Generate Probability
density
Find the
Minimum
threshold value
using PDF
Find the
segmented
image on
above image
Convolute
the image
Using Adaptive
threshold find the
value
Show the Final image and
threshold value
Stop


Fig 5 Example of Figure showing: (d) Original image; (e) manual image



(f) Detected image




Fig 6 Example of figure showing: (g) Original image (h) manual image



(i) Detected image


The lesions can be detected using the features like gray level,
shape, and intensity. By using these features the percentage of
detection lesions is 92% as shown in the table1

Property
name
No of
Images
Successful
detection
Failed
detection
SPIC 14 14 0
ASYM 10 9 1
CIRC 11 10 1
ARCH 12 10 2
MISC 10 9 1
CALC 13 13 0
Total 70 65 5
Percentage 92%

Table 1: Detection results for different lesions

V. CONCLUSION

In this paper a new algorithm is presented for the
detection of suspicious lesions in mammograms. Adaptive
Thresholding method was used to detect lesions. Images taken
from the database and various other sources have shown that
the proposed detection system is capable of detecting
suspicious lesions of different types lesions classifying get the
accuracy is 92%. The lesion with small pixel values is not
detected properly. Furthermore, the detection results for some
types of lesions mainly characterized by texture feature can be
improved if other combinations of lesion features are taken
into account in the proposed method.
REFERENCES

[1] Liu, C. F. Babbs, and E. J. Delp, ―Multiresolution detection of
speculated lesions in digital mammograms,‖ IEEE Trans. Image
Process.,vol. 10, no. 6, pp. 874–884, Jun. 2001.
[2] K. Bovis and S. Singh, ―Detection of masses in mammograms
using texture features,‖ in Proc. 15th Int. Conf. Pattern Recog.,
2000, vol. 2,pp. 267–270.
[3] G. Cardenosa, ―Mammography: An overview,‖ in Proc. 3rd Int.
Workshop Digital Mammography, Chicago, IL, Jun. 9–12, 1996,
pp. 3–10.
[4] J. Suckling, S. Astley, D. Betal, N. Cerneaz, D. R. Dance, S.-L.
Kok, J. Parker, I. Ricketts, J. Savage, E. Stamatakis, and P. Taylor,
Mammographic Image Analysis Society MiniMammographic
Database, 2005. [Online]. Available:
http://peipa.essex.ac.uk/ipa/pix/mias/
[5] M. Zhang, M. L. Giger, C. J. Vyborny, and K. Doi,
―Mammographic texture analysis for the detection of spiculated
lesions,‖ in Proc. 3rd Int.Workshop Digital Mammography, K.
Doi, M. L. Giger, R. M. Nishikawaand R. A. Schmidt, Eds.,
Chicago, IL, Jun. 9–12, 1996, pp. 347–350.
[6] F. J. Ayres and R. M. Rangayyan, ―Characterization of
architectural distortion in mammograms,‖ IEEE Eng. Med. Biol.
Mag., vol. 24, no. 1, pp. 59–67, Jan./Feb. 2005.
[7] N. Karssemeijer and G. M. te Brake, ―Detection of stellate
distortions in mammogram,‖ IEEE Trans. Med. Imag., vol. 15, no.
1, pp. 611–619, Oct. 1996.
[8] D. Guliato, R. M. Rangayyan, J. D. Carvalho, and S. A. Santiago,
―Polygonal modeling of contours of breast tumors with the
preservation of spicules,‖ IEEE Trans. Biomed. Eng., vol. 55, no.
1, pp. 14–20, Jan. 2008.
[9] H. Kobatake, M. Murakami, H. Takeo, and S. Nawano,
―Computerized detection of malignant tumors on digital
mammograms,‖ IEEE Transmit. Imag, vol. 18, no. 5, pp. 369–378,
May 1999.
[10] H. Li, Y. Wang, K. J. Ray Liu, S.-C. B. Lo, and M. T. Freedman,
―Computerized radiographic mass detection—Part I: Lesion site
selection by morphological enhancement and contextual
segmentation,‖ IEEE Trans. Med. Imag., vol. 20, no. 4, pp. 289–
301, Apr. 2001.
[11] B. R. Groshong and W. P. Kegelmeyer, Evaluation of a Hough
Transform Method for Circumscribed Lesion Detection, K. Doi,
M. L. Giger, R. M. Nishikawa, and R. A. Schmidt, Eds.
Amsterdam, The Netherlands: Elsevier, 1996, pp. 361–366.
[12] A. Mencattini, M. Salmeri, R. Lojacono, M. Frigerio, and F.
Caselli,m ―Mammographic images enhancement and denoising for
breast cancer detection using dyadic wavelet processing,‖ IEEE
Trans. Instrum. Meas., vol. 57, no. 7, pp. 1422–1430, Jul. 2008.
[13] G. M. te Brake and N. Karssemeijer, ―Segmentation of suspicious
densitiesin digital mammograms,‖ Med. Phys., vol. 28, no. 2, pp.
259–266,Feb. 2001.
[14] S. Singh and K. Bovis, ―An evaluation of contrast enhancement
techniques for mammographic breast masses,‖ IEEE Trans. Inf.
Technol. Biomed., vol. 9, no. 1, pp. 109–119, Mar. 2005.

Types of lesions. When the values are split. Harr wavelet transform may be considered to simply pair up input values. D. (3) The Bays threshold λ1 is the intersection of two solid lines that satisfy p(B)pb(λ1) = p(T)pt(λ1). Every possible threshold T is got by randomization technique namely pdf. j). In this rule. TH(i. In for each pixel SI (i. Window-Based Adaptive Thresholding Method Local segmentation is expected to give more precise results since the global segmentation finds a coarse localization of the suspicious lesions.pb( x)  1 2  1 2   exp  x  1 2  12    (1) (2) Fig. Using these equations 1 and 2 the pdf is calculated [10] . C. If SI (I. and let p(B) and p(T) be the a priori probabilities of the background and the target of image I. j)  Mvoisip  γ. j) (5) MvoisiP is an average of pixel intensity in a small window around the pixel SI (i. SI (i. Usually. 2. j) belongs to the normal area. σ1 and σ2 are the standard deviations of the background and the target of image. a decision is made to classify it into a potential suspicious lesion pixel or a normal pixel using the following rule. Kittler Minimum Error Thresholding Kittler minimum error thresholding computes an optimal image threshold and relative histogram. j) and SImin (i. it is reasonable to carry out segmentation according to λ2. Hence. The process is continued only if both classes contain at least one pixel. Then the Probability density function of original image is calculated using eqn 3[11]. That is. j) is an adaptive threshold value calculated using equation 5[3] With TH(i. j) ≥ TH(i. Here x is a pixel value. and the a priori probability of each class are unknown. when the overlap between pb(x) and pt(x) is not significant. j) denote the pixel coordinates and SI(i. pI ( x)  P( B) pb( x)  p(T ) pt ( x) 0. . is SImax (i. They have been used with the combination of adaptive global thresholding segmentation and adaptive local thresholding segmentation methods. In this paper a novel approach is presented to detect the lesions using adaptive tresholding with a blend other algorithms. j)  SI min(i. j )   1. DETECTION OF SUSPICIOUS LESIONS BY ADAPTIVE THRESHOLDING IN MAMMOGRAMS A. j) are the maximum and minimum intensity values in the large window. storing the difference and passing the sum. Its value ranges from zero to one. the histogram at the threshold T. It is interpreted as the lower boundary of the second class of pixels rather than the upper boundary of the first class. an intensity value T and is treated as being in the same class as higher intensities rather than lower intensities. j) denotes the pixel value of (i. In this paper mammograms image are wavelet transformed and the gray-level distribution of the target and the background regions of the images approaches to Probability density function. (a) CIRC (b) SPIC (c) ARCH (d) MISC pt ( x)  1 2  1 2   exp  x   2 2 2 2   µ2 >µ1  Previously aims to detect the cancer include feature detection using wavelet transforms. Assume that λ2 is the minimum value in pI (x). The target image has higher gray level than the background. The binary image R be the segmentation result. then SI(i. the Bays threshold λ1 cannot be calculated because pb(x). This transforms captures both frequency and location of images. the number of pixels in the two classes are obtained. Preprocessing The initial stages of lesions are transformed using discrete wavelet transform. Let PI (x) be the PDF of original image I. SI(i. j) and SIdif ≥ MvoisiP. Eqn 4 is used to do this step[7]. SI diff SI diff  SI max(i. j) belongs to the suspicious area else. j )  1 R(i. j). pt(x). Next the standard deviations of the classes and the criterion function are calculated. SI (i. Histogram based adaptive global thresholding algorithm and Window-based adaptive local thresholding method have been employed previously for the detection part. II. λ2 is often close to λ1. j )  1 (4) Here ( i. Histogram-Based Adaptive Thresholding Method The Histogram Based Adaptive Thresholding method is used to find the coarse segmentation to get a rough representation of the localization of suspicious lesions and then uses the fine segmentation to improve the rough representation. γ is a thresholding bias coefficient. and μ1 and μ2 are the means of the background and the target of image. j). B.

(d).Bit accuracy (gray level). Hence. all images are digitized at the resolution of 1024×1024 pixels and 8. The wavelet transform applied on the original mammograms images and the singularities are removed and generated the lesion gray-scale images. (f) and (i) are detected cancerous images. Image (a). After a convolution image the coarse segmentation on gray-level image. (e) and image (h) are manual images and images (c).MIAS database of mammograms and various data was collected from Nellai Scan center. Finally get the adaptive threshold value and detected the lesions from the images. image (b).(g) are original images. Subsequently. EXPERIMENTAL RESULTS Start Input image Normalized image DWT of the image The data used in this work is obtained from the mini. Find the Minimum threshold value using PDF Generate Probability density Convolute the image Find the segmented image on above image Fig 4 Example of result showing: (a) Original image (b) Manual image Using Adaptive threshold find the value Show the Final image and threshold value Stop (c) Detected image Fig 3: Structure of proposed system . the wavelet transform on the histograms (PDF curves) removed the fluctuations. the global local minima can be found as the adaptive global threshold to implement the coarse segmentation.IV. it is easier to select the adaptive local threshold to perform the fine segmentation. The test has been done on more than 80 images.

pp. E. Available: http://peipa. ―Detection of masses in mammograms using texture features. Jun. Delp. S. The lesion with small pixel values is not detected properly. Eds. Oct. Med. no. pp. Jan. F. L. J. and intensity. Giger. K. Bovis and S. and K. te Brake. Kok. Cerneaz. A. J. 267–270.pp. 3–10. Giger. Stamatakis. Images taken from the database and various other sources have shown that the proposed detection system is capable of detecting suspicious lesions of different types lesions classifying get the accuracy is 92%. Mammographic Image Analysis Society MiniMammographic Database. Schmidt.‖ IEEE Eng... 1. M.‖ in Proc. ―Characterization of architectural distortion in mammograms.. Vyborny. Pattern Recog. J. ―Mammography: An overview. 59–67. Dance. CONCLUSION (f) Detected image In this paper a new algorithm is presented for the detection of suspicious lesions in mammograms. 6. S. Babbs. pp.essex. 2005.‖ in Proc. J. Taylor. Chicago. shape. and E. vol. 611–619. 24. R. Karssemeijer and G./Feb. no. F. the detection results for some types of lesions mainly characterized by texture feature can be improved if other combinations of lesion features are taken into account in the proposed method. Parker. 2001. 874–884. M. Singh.uk/ipa/pix/mias/ M. G. Ayres and R.ac. K. D. N. N. IL.‖ in Proc. vol. Biol. Mag. Nishikawaand R. 2. I.-L. Astley. By using these features the percentage of detection lesions is 92% as shown in the table1 [7] Liu. Savage. 9–12. 10. 3rd Int. Imag. Rangayyan. 15th Int. pp.. 1996. pp. M. 9–12. J. D. Suckling. ―Detection of stellate distortions in mammogram.. 1996. IL. 1996. M. Doi. M. L.‖ IEEE Trans. 3rd Int. Chicago. Conf. Adaptive Thresholding method was used to detect lesions. 2005. 1. Betal. R. Image Process. J. ―Mammographic texture analysis for the detection of spiculated lesions. Cardenosa. Ricketts. 15. Zhang. and P. C.Property name SPIC ASYM CIRC Fig 5 Example of Figure showing: (d) Original image. no. C.‖ IEEE Trans.Workshop Digital Mammography. Doi.vol. . 347–350. Jun. Workshop Digital Mammography. vol. Furthermore. [Online]. Med. (e) manual image No of Images 14 10 11 12 10 13 70 Successful detection 14 9 10 10 9 13 65 92% Failed detection 0 1 1 2 1 0 5 ARCH MISC CALC Total Percentage Table 1: Detection results for different lesions V. REFERENCES Fig 6 Example of figure showing: (g) Original image (h) manual image [1] [2] [3] [4] [5] (i) Detected image [6] The lesions can be detected using the features like gray level. 2000. ―Multiresolution detection of speculated lesions in digital mammograms. Jun.

―Segmentation of suspicious densitiesin digital mammograms. no. 289– 301. vol. 1. H. A. Phys. Amsterdam. 1422–1430.. . M. 361–366. pp. Freedman. H. A. pp.[8] [9] [10] [11] [12] [13] [14] D. Wang. Jan. J. 1996. vol. and S.‖ IEEE Transmit. Santiago. Y.. pp. 259–266. 109–119. Nawano. and R. Imag. Kegelmeyer. K. Eng. H. Technol. 55. 20. Guliato. May 1999. Singh and K. Schmidt. pp. vol. R. 4. 2001.‖ IEEE Trans.m ―Mammographic images enhancement and denoising for breast cancer detection using dyadic wavelet processing. pp. G. 18. 28. D. The Netherlands: Elsevier. Frigerio.‖ IEEE Trans. vol. Eds. 57. Evaluation of a Hough Transform Method for Circumscribed Lesion Detection. no. no. Mar. A. T. no. Murakami. K. pp.-C. 14–20. pp. and F. ―Computerized radiographic mass detection—Part I: Lesion site selection by morphological enhancement and contextual segmentation. no. Rangayyan.‖ Med. Takeo. Li. M. 2008. Mencattini. Med.Feb. Giger. Lo. Salmeri.‖ IEEE Trans. ―Polygonal modeling of contours of breast tumors with the preservation of spicules. P.‖ IEEE Trans. and M. Ray Liu. Apr. 5. M. Karssemeijer. Meas. no. 2001. te Brake and N. 9. M. Jul. L. Groshong and W. 2. 1. B. S. M. Caselli. J. and S. vol. Biomed. Inf. R. Kobatake. Lojacono. Bovis. Nishikawa. Imag. S. Instrum. R. ―An evaluation of contrast enhancement techniques for mammographic breast masses. M. B. Biomed. vol. Carvalho. 7.. Doi. R. M. 369–378. 2005.. 2008. ―Computerized detection of malignant tumors on digital mammograms..

Sign up to vote on this title
UsefulNot useful