Breast Cancer Detection Using Spectral Probable Feature on Thermography Images
Rozita Rastghalam
Department of Electrical Engineering, Najafabad Branch, Islamic Azad University,
Isfahan, Iran.
Hossein Pourghassem
Department of Electrical Engineering, Najafabad Branch, Islamic Azad University,
Isfahan, Iran.
Abstract—Thermography is a noninvasive, non-radiating, fast, and painless imaging technique that is able to detect breast tumors much earlier than the traditional mammography methods. In this paper, a novel breast cancer detection algorithm based on spectral probable features is proposed to separate healthy and pathological cases during breast cancer screening.
Gray level co-occurrence matrix is made from image spectrum to obtain spectral co-occurrence feature. However, this feature is not sufficient separately. To extract directional and probable features from image spectrum, this matrix is optimized and defined as a feature vector. By asymmetry analysis, left and right breast feature vectors are compared in which certainly, more similarity in these two vectors implies healthy breasts. Our method is implemented on various breast thermograms that are generated by different thermography centers. Our algorithm is evaluated on different similarity measures such as Euclidean distance, correlation and chi-square. The obtained results show effectiveness of our proposed algorithm.
Keywords— Breast thermogram, Breast cancer detection, Asymmetric analysis, Image spectrum, Spectral probable feacture
I. INTRODUCTION
Breast cancer is the most common disease amongst women in many countries. According to the National Cancer Institute of the United States of America, breast cancer is the second most common cancer among women and one of the ten leading causes of death among women in the United States [1] and as reported to the cancer registry of the Ministry of Health and Medical Education of Iran, breast cancer is the most common new cancer case (24/41% of all female cancers) allocated to the women [2]. There are several imaging techniques to detect breast cancer that mammography method is the most current one. Drawback of this technique is that it is invasive, is not proper for women with dense breasts, implants, fibrocystic breasts, or on hormone replacement therapy and experts believes that electromagnetic radiation can also be a triggering factor for cancerous growth. BRCA1/2 genes are responsible for correcting DNA mutations; mutations that may result from radiation. Therefore, use of mammography is not recommended for the patients who have frequent mutations in their genes. As compared to thermography, this method cannot detect tumor so early [3, 4, 5]. Digital Infrared Thermal Imaging (DITI) of the breast offers the opportunity of earlier detection of breast disease. DITI is a non-invasive test that
there is no contact with the body of any kind, no radiation and the procedure is painless. Research has shown that if tumor is detected earlier in patient breast (tumor size less than 10mm), the chance of cure will be 85% as opposed to 10% if the cancer is detected late. It should be noted that thermography detects tumor 8-10 years earlier than that mammography can detect a mass in the patient's body [6]. One of the popular ways for separating the normal from abnormal breast is the application of asymmetric analysis. In this way, the left and right breasts, or their features and any asymmetry in determining abnormalities of the breast can be compared.
Regarding, Qi et al. [7] proposed an automated asymmetry analysis technique for the separation of normal breast from abnormal. At the beginning, they detected the edges using Canny filter and identified the left and right body boundary curves and two lower boundaries of the breasts using Hough Transform. Thus, they could divide left and right breast and then draw Bezier histogram for each segment and compute the curvature from the two histograms. The difference in curvature is used as a measure for abnormality identification.
However, success of the proposed technique lie in effective edge detection and segmentation algorithms. In [8], the average temperature between the left and right breast is determined then the image is divided into the left and right of upper quadrant and lower quadrants. If the average temperature between each left and right quarter is smaller than one degree, the score will be 0.5 and the score will be 1for the difference up to one. An index is created through adding the scores for the four quadrants. In case the index is greater than 1, then it indicates the presence of the abnormality. In another paper [9], the temperature distribution of the breast and its histogram is drawn and asymmetry of the histograms represents that the breast is not normal. In another proposed method [10], abnormality was described using first order moments such as mean, skewness and kurtosis, and the second order statistical parameters namely co-occurrence matrix and parameters as like entropy, energy based on image transformation using wavelet analysis. Also it was expressed in this method that use of Artificial Neural Networks is useful.
In a research by Ghayoomi zadeh and colleagues [11, 12], an algorithm consists of the following steps was suggested:
Conversion of Pseudo-color thermographs into gray scale, improving the brightness by thresholding, detecting the breast 2013 8th Iranian Conference on Machine Vision and Image Processing (MVIP)
boundaries using Hough transform, determining the parameters as like mean, variance skewness, and kurtosis for histograms of the left and right sides of the breasts and finally decision making based upon these parameters. They also found that for their set of thermographs from a specific IR camera, tumor cells correspond to pixels in Pseudo-color thermographs with Red intensity(RT)!100, Green intensity, and blue intensity(Gt&BT 20). However, thresholding based on this approach will result in few artifacts that belong to the same intensity level.
Our aim in this paper is to use asymmetry analysis screening automatic normal from abnormal images. The previous related works has focused more on the texture images and has paid attention to the separation of cancer images from normal ones through the extraction of different features from the texture of images. However, in this paper we compared the left and right breast spectrum information. For this purpose, we applied co- occurrence matrix to extract probable feature of image spectrum. At the end, our proposed method is evaluated with different measures as like Euclidean distance, correlation, chi- square[13] and those measures introduced the total number of changes in order to separate normal breasts from abnormal ones. The rest of this paper is organized as below. Section II, explains principles of our proposed breast cancer detection structure and describes this algorithm in details. The experimental results of the proposed breast cancer detection and conclusions are presented in sections III and IV, respectively.
II. PROPOSED BREAST CANCER DETECTION STRUCTURE
A. Breast Segmentation
The block diagram of the proposed structure is shown in Fig. 1. In Fig. 2 (a) is displayed a breast infrared thermogram of a patient in RGB and in Fig. 2 (b) is shown a breast thermogram of a cancer patient in gray scale. The red pixels in the RGB image have higher temperature values. As the metabolism and cellular consecutive dividing increase in the cancer cells, the temperature rises around the cancer cells, which it is an evident in the thermographic images as red spots. Breast segmentation is performed through cropping the left and right breasts from the main breast thermography images. Since our proposed algorithm is based on asymmetric analysis, use of a constant mask in all of images is very important. In fact, breast segmentation is obtained from cropping the left and right breasts in each thermography image with a unit mask. Left and right breasts segmentation is shown in Fig.3 [14].
B. Spectral Feature
Spectrum analyzer related to in any picture is an important feature in frequency domain. Suppose f(x,y) is an
N
M u image. Two-dimensional discrete Fourier transform of )
, (x y
f , is given with f(u,v) and is calculated as follows:
e N
vy M j ux M
x N Y
y x MN f
v u
f 1 2 ( )
0 1 0
) , 1 (
) ,
( ¦
¦ S (1)
Above equation can be converted from trigonometric to exponential. Frequency domain is simply the coordinate system and the frequency variables u and v as f(u,v)is also available. Most of the rectangular area is referred to as the frame frequency, where u and v are determined as
1 0,1,2,...M
u andv 0,1,2,...,N1. The frame
frequency has the same size as the input image [15]. The result of the Fourier transform of a breast segmentation image is shown in Fig. 4 (a). The value of Fourier transform in the origin of frequency domain f(0,0)is called dc component of the Fourier transform. Thus, spectrum center includes useful information and Spots away from the center covering noise and less important information. For this reason, we use the part of spectrum that is shown in Fig. 4 (b). The left half and hundred final pixels of spectrum changed to zero [16].
Fig. 1. Block diagram of the proposed breast cancer detection
(a) (b)
Fig. 2. thermography image (a). RGB, (b). gray scale
Fig. 3. Breast segmentation Convert RGB to grayscale Breast thermogram image
Segmentation based on crop
Spectral feature extraction
Spectral probable feature extraction
Asymmetric analysis evaluation
(a) (b)
Fig. 4. (a). Breast image spectrum, (b). The used reduced spectrum with discarding redundancy.
C. Spectral co-occurrence feature
Gray level co occurrence matrix (GLCM), one of the most known texture analysis methods, estimates image properties related to second-order statistics based on neighboring pixels.
The GLCM is a two dimensional array which takes into account the specific position of a pixel relative to other pixels [17]. The GLCM is a tabulation of how often different combination of pixel brightness values occur in an image.
GLCM matrices are produced at a distance of d 1,2,3,4 and for direction of data given as0q,45q,90q,180q. So far, GLCM has been used to feature extraction from image texture, but extraction spectral feature, directional information of spectrum and spectral probable feature from GLCM are proposed for the first time in this paper [18–21].
D. Optimized spectral co occurrence features
Spectrum texture of healthy breast image is very similar to spectrum texture of abnormal breast image. Definitely, use of GLCM to extract the texture features from image spectrum for intercept normal image of cancer is not effective. Actually, images spectrum do not have acceptable information in their texture. Our target is to extract the changes in directional feature spectrum and also probable feature spectrum of each breast. In the image spectrum, there are not several gray levels, and approximately all image spectrum pixels in terms of color have medium gray levels. Thus, our proposal is the extract feature vector of GLCM. The previous step is produced the gray level co-occurrence matrix with (8u8u4)size from image spectrum. This matrix is converted to a matrix with
) 7 1
( u elements in our approach. This (1u7) matrix shows the number of occurrences of our optimum features that is extracted from GLCM matrices. This 7 optimum features is defined as difference between co-occurrence levels (DCL). As the result, the above optimized matrix shows the occurrence number of differences between co-occurrence levels (ONDCL). Fig. 5 displays the ONDCL extraction procedure from a (8u8) GLCM matrices. The first element of DCL vector is generated from the sum of all GLCM matrix entries, which their subtraction of row
(i )
and column (j) is one. The Second element is derived from total GLCM matrix entries, for which the difference between i, j values is 2. According to Fig.5, the sixth element of DCL matrix is made with the sum of GLCM(1,7) , GLCM(2,8) , GLCM(7,1) and GLCM(8,2). This
method will continue to produce the seventh element of DCL vector. In equation (2), formation Method of DCL vector in mathematical terms is expressed. In fact, DCL vector is made based on rate of difference between row and column of GLCM.
Indeed, this optimized matrix is captured from compress GLCM information. In GLCM, paying attention to the details is extreme and this is due to the GLCM failure in image spectrum. The value of each pixel and its neighbors is not considered in this new approach. However, the changing number of each pixel value with its neighbors is important.
This method is not only suitable for image spectrum, but also this is useful for the image itself. The DCL is defined as,
7 2 1 , ,..
k and k j
i , if i,j GLCM(i.j)
DCL(k) ¦
(2)where GLCM is the co-occurrence matrix with 8 levels.
Fig. 5: DCL optimal vector extraction from GLCM
III. EXPERIMENTAL RESULTS
We studied 28 breast thermograms available from Ann Arbor Thermography Center [22], Thermal imaging lab in San Francisco Bay Area [23], American College of Clinical Thermology [24], Thermography of Iowa [25], Sunstate Thermal Imaging Center in Australia [26], and thermography of Chicago [27]. Although the images from these sources are varied in their resolutions and generally did not follow a unified protocol, our algorithm could separate cancer cases from normal ones with acceptable accuracy. For choice a reasonable decision boundary between normal from abnormal breast, half of this images is used as training cases.
Abundance of each DCL vector entries is shown in Fig. 6 (a) and (b), which are related to normal and cancer cases, respectively. These graphs act as a histogram indicating the abundance in each DCL vector entries. If the number of each DCL element for the left and right breasts are closer to each other (in other words, the graph for left and right breasts have more overlapping), the person will be considered to be healthy. In addition, the more difference between left and right breast graphs, cancer or fibrocystic risks will be higher. In fact, with this vector, the changes in directional feature and probable feature of left and right breast spectrum are compared with each other. DCL histogram of the left and right breasts for two normal cases are displayed in Fig. 6 (a), for
one of which the performance of our algorithm is true and predicted as normal that is shown with TN (true negative). In another one, forecast of algorithm has not been true since the normal was diagnosed as abnormal incorrectly that is displayed with FP (false positive). Fig. 6 (b) includes DCL vector histogram of the left and right breasts for two cancer cases that Left.TP (left true positive) and Right.TP (right true positive) graphs are related to the left and right breasts in a cancer case that have predicted cancer in a correct way.
Left.FN (left false negative) and Right.FN (right false negative) diagrams are indicating the left and right breasts in a cancer case that are intended as normal. Thermography image related to a FP case is provided in Fig. 2. (a). This normal image has some segments with red color and certainly our algorithm identifies these segments as cancer, while the patient is not suffering from cancer. Sometimes pulmonary and respiratory
1 2 3 4 5 6 7
0.5 1 1.5 2 2.5
3x 105
DCL
number of occurrence
Left.FP Right.FP Left.TN Right.TN
(a)
1 2 3 4 5 6 7
0.5 1 1.5 2 2.5 3 3.5
4x 105
DCL
Number of occurrences
Left.TP Right.TP Left.FN Right.FN
(b)
Fig. 6. Two examples of asymmetric analysis. (a) Two samples of extracted DCL optimal feature vector for normal cases (true negative and false positive), (b) Two samples of extracted DCL optimal feature vector for cancer cases (true
positive and false negative)
problems are considered to be the reasons of the aforesaid red segments in thermography images, and this is the weakness point of thermography imaging technique, not our proposed algorithm. Our algorithm follows with little errors and high accuracy. However, this flaw and low resolution in certain thermography images reduces the accuracy in our algorithm and other methods. Of course this drawback is negligible versus thermography major advantages, such as tumor detection about 8 years earlier than mammography. Our
algorithm is evaluated with different similarity measures such as correlation, Euclidean, Chi-square, and also the sum of feature vector entries of left and right breast is considered as the total number of changes in image spectrum of each breast.
Difference between the total number of changes (DTNC) for left and right breast is selected as another measure. As shown in Fig. 6 (a) and (b), major alterations occurred in the first and second levels of the rate of change. As such, the introduction of DTNC measure in image spectrum for which the main changes are limited to one or two level can be true. It should be noted that use of the above-mentioned measure is not suitable to compare texture images or the image itself in which alterations occur in various levels. Of course, DTNC measure with correct weighting to feature vector entries can be appropriate. For chi-square, Bhattacharyya distance, Euclidean distance and DTNC measure, a low score represents a better match in two histograms than a high score. However, for correlation a high score represents a better match than a low score. The result of this evaluation is shown in TABLE I, in which the incorrectly prediction is displayed with red color in each measure. For example, evaluation with DTNC measures has two errors, one of which is related to FP and another one is corresponding to FN. And also prediction with correlation measure has three FN errors and four FP errors. Definitely;
correlation is not a suitable measure in our work. In this method, these measures are appropriate that compare the corresponding entries of DCL vectors. As shown in TABLE II, if un-optimized GLCM is applied to the image spectrum, the accuracy in separation of normal from abnormal cases will be reduced, since special attention has paid to all the changes and even the details of image texture for the generation of GLCM matrix and this much accuracy has minimized the data effects and useful features. Incorrect predictions are displayed with red color in TABLE II, same as TABLE I. As it is evident, the number of errors in this method is much greater than our proposed algorithm. Efficient features of spectrum are extracted by means of optimal GLCM that has been indicated in the present study by DCL, and also this optimal matrix can be more effective than GLCM for feature extraction from texture.
IV. CONCLUSION
On time diagnosis of breast cancer is one of the fundamental issues for the researchers. Taking into account the huge costs for treatment and development of the aforesaid disease amongst women, early identification of breast cancer is considered to an optimal step towards the reduction of bad social and health consequences. Previous studies provided various influential methods for the classification of cancer templates. Till date, none of the above could be able to separate cancer templates accurately. In our proposed method, normal and abnormal patterns have been separated from each other through a new procedure in which all the image spectrum features are indicated. Definitely, using of these spectrum features besides the tissue features will optimize classification accuracy. In this paper, we analyzed thermal breast images using probable feature of spectrum to determine difference between normal and abnormal cases. The results of
this study indicate that useful feature of spectrum can be extracted with the modification of gray level co-occurrence matrix and use of this optimal matrix. In addition, with asymmetric analysis and comparison of these effective features of left and right breast, breast cancer screening is fulfilled. This optimal matrix with little changes can be used for extracted beneficial feature of texture images.
TABLE I. The cancer detection results of our algorithm by GLCM and DCL
Similarity measures
Correlation DTNC
Chi square Euclidean
Cancer Normal Cancer Normal
Cancer Normal Cancer Normal
Image class
3 (FN) 5 (TP) 1 (FN) 7 (TP) 2 (FN) 6 (TP) 1 (FN) 7 (TP) Cancer
2 (TN) 4 (FP) 5 (TN) 1 (FP) 5 (TN) 1 (FP) 5 (TN) 1
(FP) Normal
TABLE II. The cancer detection results by GLCM without DCL.
similarity measures
Correlation Bhattacharyya
Chi square Euclidean
Cancer Normal Cancer Normal
Cancer Normal Cancer
Normal Image
class
6 (FN) 2 (TP) 2 (FN) 6
(TP) 3 (FN) 5 (TP) 2 (FN) 6 (TP) Cancer
1 (TN) 5 (FP) 3 (TN) 3
(FP) 4 (TN) 2 (FP) 4 (TN) 2 (FP) Normal
REFERENCES
[1] B.F. Jones, “A reappraisal of the use of infrared thermal image analysis in medicine,” IEEE Trans. Med. Imaging, vol. 17, pp.
1019-1027, 1998.
[2] Sh. Babazadeh, A. Andalib, A. Amuheidari, M. tabataaeian, H.
Emami, A. Adibi, F. Taleghani, M. Alamsamimi, M. Roayaei, M.
Hoseini, “A study of long term trend in breast cancer epidemiological factors and clinical parameters in Esfahan province,” Journal of Isfahan Medical School, vol. 29, pp. 61-84, January 2012.
[3] J. Head, F. Wang, C. Lipari, R. Elliott, “The Important Role of Infrared Imaging in Breast Cancer,” IEEE Engineering in Medicine and Biology, vol. 19, pp. 52-57, 2000.
[4] N. Arora, D. Martins, D. Ruggerio, E. Tousimis, A. Swistel, M.
Osborne, R. Simmons, “Effectiveness of a noninvasive digital infrared thermal imaging system in the detection of breast cancer,”
The American Society of Breast Surgeons, vol. 196, pp. 523-526, 2008.
[5] D. Kennedy, T. Lee, D. Seely, “A Comparative Review of Thermography as a Breast Screening Technique,” Integrative Cancer Therapies, vol. 08, pp. 9-16, 2009.
[6] E. Y. K. Ng, L. N. Ung, F. C. Ng, L. S. J. Sim, “Statistical analysis of healthy and malignant breast thermography,” Journal of Medical Engineering and Technology, vol. 25, pp. 253-263, 2001.
[7] H. Qi, W. Snyder, J.F. Head, R.L. Elliott, “Detecting breast cancer from infrared images by asymmetry analysis,” Proc. 22nd Annu.
Conf. IEEE Engineering in Medicine and Biology Society, Chicago, pp. 23–28, July 2000.
[8] M. Frize, C. Herry, R. Roberge, “Processing of thermal images to detect breast cancer,” 4th International IEEE EMBS Special Topic Conference on Information Technology Applications in Biomedicine, vol. 680, pp. 234-237, 2003.
[9] H. Yang, SH. Xie, Q. Lin, SH. Ye, SH. Chen, H. Li, “A new infrared thermal imaging and its preliminary investigation of Breast Disease Assessment,” Proc. Of IEEE/ICME, International Conference on Complex Medical Engineering, pp. 1071-1074, 2007.
[10] M. Wiecek, R. Strakowski, T. Jakubowska, B. Wiecek, “Software for classification of thermal imaging for medical applications,” 9th International Conference on Quantitative Infrared Thermography, vol. 14, N0. 2, 2008.
[11] H. Ghayoumi Zadeh, A. Kazerouni, Haddadin, “Distinguish breast cancer based on thermal Features in Infrared Images,” Canadian Journal on Image processing and computer vision, Vol. 2, pp. 54- 58, 2011.
[12] N. Selvarasu , “Image Processing Techniques and Neural Networks for Automated Cancer Analysis from Breast Thermographs-A Review,” Indian Journal of Computer Science and Engineering (IJCSE), vol. 3, No. 1, pp. 133-137, 2012.
[13] X. Liu, D. Wang, “Texture Classification Using Spectral Histograms,” IEEE Transactions on Image Processing, vol. 12, pp.
661-670, No. 6, June 2003.
[14] O. D. Nurhayati, A. Susanto, T. Sri Widodo, “Detection of the Breast Cancer from Thermal Infrared Images based on Statistical Characteristics,” International Journal of Science Engineering and Technology, Vol. 2, No. 2, pp. 65-69, 2009.
[15] S. Wang, “Applications of Fourier Transform to Imaging Analysis,”
Journal of the Royal Statistical Society, vol. 171, may 2007.
[16] A. Gharekhan, A. N. Oza, M. B. Sureshkumar, P. Panigrahi,
“Polarized spectral features of human breast tissues through wavelet transform and principal component analysis,” Pramana journal of physics, vol. 75, No. 6, pp. 1281-1286, December 2010.
[17] A. MohdKhuzi, R.Besar, W. Zaki , NN. Ahmad,”Identification of masses in digital mammogram using gray level co occurrences matrices,” Biomedical Imaging and Intervention Journal, vol. 5, No. 3, June2009.
[18] R. Haralick, K. Shanmugam, I. Dinstein, “Textural Features for Image Classification,” IEEE Trans. on Systems, Man, and Cybernetics, vol. SMC-3, pp. 610-621, November 1973.
[19] R. Nithya, B. Santhi, “Comparative Study on Feature Extraction Method for Breast Cancer Classification,” Journal of Theoretical and Applied Information Technology, vol. 33, pp. 220-226, November 2011.
[20] M. Hauta-Kasari, J. Parkkinen, T. Jaaskelainen, R. Lenz, “Multi spectral texture segmentation based on the spectral co occurrence matrix,” Pattern Analysis & Applications, pp. 275–284, 1999.
[21] R. Nithya, B. Santhi, “Classification of Normal and Abnormal Patterns in Digital Mammograms for Diagnosis of Breast Cancer,”
International Journal of Computer Applications, vol. 28, No. 6, pp.
21-25, August 2011.
[22] AAT: http://aathermography.com (last accessed March 2013).
[23] MII: http://www.breastthermography.com/case_studies.htm (last accessed March 2013).
[24] ACCT:www.thermologyonline.org/Breast/breast_thermography_w hat.htm (last accessed July 2011).
[25] http://www.thermographyofiowa.com/casestudies.htm (last accessed March 2013).
[26] STImaging: http://www.stimaging.com.au/page2.html (last accessed March 2013).
[27] http://www.drnick.net/index.php?p=224034 (last accessed January 2013)
