2012 IEEE EMBS International Conference on Biomedical Engineering and Sciences | Langkawi | 17th - 19th December 2012

Random Walkers Based Segmentation Method For Breast Thermography

Mehrdad Moghbel
Sensor Technology Laboratory, Institute of Advanced Technology, UPM Serdang, Malaysia E-mail: mehrdad2275@gmail.com

Dr.Syamsiah Mashohor
Dept. of Computer & Communication Systems, Faculty of Engineering, UPM Serdang, Malaysia E-mail: syamsiah@eng.upm.edu.my

Prof. Dr. Hajjah Rozi Mahmud

Dept. of Informatics and Imaging, Faculty of medicine and health sciences, UPM Serdang, Malaysia E-mail: rozi@medic.upm.edu.my

Assoc.Prof. dr. M. Iqbal Bin Saripan

Dept. of Computer & Communication Systems, Faculty of Engineering, UPM Serdang, Malaysia E-mail: iqbal@eng.upm.edu.my

Abstract In breast thermography diagnostic, proper detection and segmentation of the areola area as well as detection of breast boundaries present the biggest challenge. As the boundaries of breasts especially in the upper quadrants are usually not present, this produces a great deal of challenge to segment breasts automatically. Many approaches have been developed to segment the breast in the past such as Snakes, Active Contours and Circular Hough Transforms, but these methods fail to detect the boundaries of the breast with the required level of accuracy especially the upper boundaries of the breast. By utilizing most recent segmentation method which is Random Walkers, the breast can be segmented accurately which in turn will increase the accuracy and the reliability of computer aided detection/diagnosis systems. Keywords: image segmentation, random walkers, breast thermography, areola detection

achieve satisfactory results in presence of noise or incomplete image data. One of such methods which achieved a wide use in medical imaging is the random walkers algorithm. The use of random walkers for image segmentation was first suggested by [2], it is a supervised segmentation method which means the user must define a set of pixels belonging to each object for segmentation. This can be done automatically according to a predefined criterion or be done interactively by the user. The random walker segmentation is based on the probability of a random walker placed on a non-seed pixel reaching the seed first, in real life images with high pixel intensities it is practically impossible as the computational needs are beyond commercial computers. With thanks to the graph theory and the existing solution for the combinatorial Dirichlet problem, it was shown that the probability of a random walker reaching a seed point exactly equals the solution to the Dirichlet problem [2-4]. These probabilities can be calculated without the need to actually engage in the random walks with ease and speed. In image processing, random walkers is based on construction of an undirected graph, G = (V, E); with image pixels corresponding to nodes and connecting edges . Weights, are assigned to edges that connects nodes and , according to the following equation: (1) Where is a user adjustable parameter and is the image intensity at pixel . The weight is equivalent to the probability that the random walker crosses that particular edge, by adjusting we can adjust the edge crossing behavior of the random walker algorithm so that it may pass or stop at

I.

INTRODUCTION

Due to human eyes limitation to differentiate minute temperature differences produced by thermography imaging, computers are used to analyze the data. Operator selects the regions of the interest and the computers calculates the data, as this approach is difficult and times consuming, a number of computer aided detection/diagnosis (CAD) approaches based on artificial intelligence were developed. The problem with these CAD systems lies in the fact that although some of the designs are capable of achieving almost 90% accuracy [1], the breasts are segmented manually or semi-automatically before being introduced to the CAD system. Breast segmentation in thermal imaging is especially challenging as the image is actually a map of intensity variations. As these variations differ little from each other and are mostly uniform, it is very difficult to draw breast boundaries on the image. With the advancements in computational capacity and image processing techniques, new segmentation approaches are being developed which can

978-1-4673-1666-8 | IECBES 2012 |

600

2012 IEEE EMBS International Conference on Biomedical Engineering and Sciences | Langkawi | 17th - 19th December 2012

particular edges [2,5]. The random walkers have been successfully utilized by the image processing and the biomedical engineering practices for the purpose of segmenting medical images. Random walkers have shown great accuracy and noise tolerance usually present in medical images as seen in Figure 1. In a study comparing different segmentation methods for positron emission tomography (PET) images random walkers were found to provide the most accurate results [6], in case of CT images random walkers were more accurate than other segmentation methods such as level sets [7] and in case of brain MRI random walkers performed accurately [8]. Although thermography imaging differs significantly from MRI and CT scan imaging, high performance of random walkers in MRI and CT image segmentation where the edges between different regions are difficult to detect motivated the adaptation of random walkers segmentation of breasts in thermography.
1

image, the efficiency of these methods decreases dramatically. For medical evaluation breasts are divided to five categories based on their thermal properties namely TH1 (normalnonvascular), TH2 (normal-vascular), TH3 (equivocal), TH4 (Abnormal), or TH5 (severely abnormal), and these categories are illustrated in Figure 2. II.
PROPOSED SEGMENTATION

All images used for this study were captured by state of the art FLIR thermal cameras and have good resolution and sensitivity. The developed algorithm needs no human intervention until the final result is displayed to the user, if the user is not satisfied with the segmentation results she/he can appoint new seeds interactively to fine tune the segmentation. Although the algorithm showed great success for all images from TH1 till TH5, this interactive mode will further increase the accuracy of the results. Main steps for the breast segmentation with random walkers include the removal of the background which does not fall into the body boundaries, dynamic intensity level stretching and filtering, areola detection and seed implementation for segmentation A. Edge detection and background removal The removal of the background in many image segmentation algorithms provides better accuracy. For this study an adaptive method was developed which selects the best parameters for the edge detection automatically, it should be noted that the images were smoothen by a Gaussian smoothing function before this process. After the edges were detected the image is then abstracted from the background, the process work successfully in all images in our data base, the example can be seen in figure 2. B. Dynamic intensity level stretching and filtering As the intensity properties of the images were different in respect to the machine that they were acquired with, three different approaches were developed to modify the image based on the intensity values of the original image. For increasing the level of detail in the image before applying the areola detection, a series of Top-Hat and Bottom-Hat enhancements are applied on the image, this is done to emphasize the edges present in the image. A circular structuring element was used. The results can be seen in Figure 3.

Figure 1: Random walkers were able to segment the image in presence of grate noise (1), PET segmentation (2) CT segmentation of tumor from lung (3)

Image segmentation approaches are usually based on the identification of characteristics or features of the object and leveraging on them to achieve a proper segmentation. In breast thermography the lack of defined edges on the upper boundaries of the breast and the fact that breasts have different shape and characteristics even between breasts of a single individual, makes segmentation of breasts a difficult task for all segmentation methods. Many methods have been developed such as Snakes [15], Hough transform based segmentation [9], Morphological Image Segmentation [10] and curvature based segmentation [9][11]. Although almost half of all breast cancers occur in the upper region of the breast known as the tail of the breast [12,13], most of these methods cannot segment the upper region of the breast with required accuracy.

In case of TH4 and TH5 patients, due to intensive disfigurations in the heat distribution and shapes present in the

Figure 2: TH levels from TH1 to TH5 in order from left to right; the increase in thermal distortion with increase in TH level can be seen clearly

601

2012 IEEE EMBS International Conference on Biomedical Engineering and Sciences | Langkawi | 17th - 19th December 2012

D. Seed implementation and Random Walker algorithm After the nipple area was detected, the seeds are implemented dynamically according to the position of the areola. The main seeds for the breast are shaped as an ellipse as the breasts are usually shaped like ellipses. The random walkers show incredible ability in detecting the missing boundaries or where the boundaries are very weak and almost undetectable by other methods. In case of breasts the upper boundaries represents the greatest challenge to any segmentation method. Random walker showed promising results as seen in Figure 5,the red area shows the segmented breasts, it should be noted that the seeds were not put dynamically and the same seeding position were used for all images and only covered a portion of the breast area, this was done to show the robustness of algorithm.

a
Figure 3: (a) Image areola detection without dynamic intensity stretching in an image with poor intensity levels, (b) areola detection with intensity stretching, (c) top-hat image, (d) bottom-hat image and (e) structuring element.

C. Detecting the areola area A series of modified circular Hough transforms [14] are applied to the image to find the most likely candidates for the areola area, after the calculation a matrix of possible candidates are produced. As the areola area shows significant intensity variations from the surrounding tissue, an algorithm which was developed to analyze the intensity variations is then applied to each point analyzing radios of 4 to 25 pixels around each point. After analyzing these variations the algorithm returns the most likely area of the areola with its calculated area and diameter. Although the algorithm functioned well in most of the images, it struggled to produce accurate results in non-standard images with poor intensity levels as seen in Figure 4.

Figure 5: Image at the left shows the seeds, a) red seeds represent breast and the green seeds represent the background, b) segmented image for TH1 Patient, c) segmented image for TH2 and d) segmented image for TH1 patient with small breast.

Figure 4: The algorithm is still accurate even in images with high variation in intensity of the area surrounding the areola, but in case of images with very poor intensity destributaion ( the image at right hand) the algorithm looses accuracy. The (+) sign shows the likely candidates of the nipple and the white circle shows the setected areola region and its outline .

In case of TH4 and TH5 some artifacts might be present in the segmentation of the image which are identified and removed automatically. The ability of the random walkers to differentiate different areas with different parameters marked as belonging to the same region proves invaluable in segmentation of TH4 and TH5 type breast thermography where there are abnormalities within the breast which must be included in the segmentation. The developed algorithm

602

2012 IEEE EMBS International Conference on Biomedical Engineering and Sciences | Langkawi | 17th - 19th December 2012

showed great accuracy in fully dynamic and automatic mode for the detection of the breast boundaries as seen in Figure 6. a b

more clinics for increased sensitivity in evaluation of the breast cancer. ACKNOWLEDGMENT Authors would like to thank Dr. Edward B. Jay director and founder of thermography assessment services whose help with preparation of the patient database and over 30 years of thermography experience proved invaluable. REFERENCES
[1] T.Z. Tan, C Quek, G.S. Ng, E.Y.K. Ng, A novel cognitive interpretation of breast cancer thermography with complementary learning fuzzy neural memory structure. Expert Systems with Applications, 2007. 33(3): p. 652-666. Grady, L. Multilabel random walker image segmenConference on Computer Vision and Pattern Recognition, IEEE 2005. CVPR 2005: p. 768-770. Grady, L., Random walks for image segmentation. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2006. 28(11): p. 17681783. Grady, L. and A.K. Sinop. Fast approximate random walker segmentation using eigenvector precomputation. Conference onComputer Vision and Pattern Recognition, IEEE 2008. CVPR 2008: p. 1-8. Andrews, S., G. Hamarneh, and A. Saad, Fast random walker with priors using precomputation for interactive medical image segmentation. Medical Image Computing and Computer-Assisted InterventionMICCAI 2010, 2010: p. 9-16. Bagci, U., et al. A graph-theoretic approach for segmentation of PET images. 2011: Engineering in Medicine and Biology Society,EMBC, 2011 Annual International Conference of the IEEE:p.8479-8482 Chen, M., et al. Random walk-based automated segmentation for the prognosis of malignant pleural mesothelioma. 2011 IEEE International Symposium on Biomedical Imaging: From Nano to Macro , p. 19781981 Choubey, M. and S. Agrawal, An implementation of random walk algorithm to detect brain cancer in 2-d MRI. Magnetic resonance imaging, 2012. 2(1): p. 998-1001. Qi, H. and J. Head. Asymmetry analysis using automatic segmentation and classification for breast cancer detection in thermograms. Engineering in Medicine and Biology Society, 2001. Proceedings of the 23rd Annual International Conference of the IEEE: vol 3 . p.2866-2869 Zeng-qiang;, C.B.-p.M. Automated Image Segmentation and Asymmetry Analysis for Breast Using Infrared Images. in Education Technology and Training, 2008. and 2008 International Workshop on Geoscience and Remote Sensing. ETT and GRS 2008. International Workshop on 2008. Scales, N.K., C.; Prize, M., Automated image segmentation for breast analysis using infrared images in Engineering in Medicine and Biology Society, 2004. IEMBS '04. 26th Annual International Conference of the IEEE 2004. p. 1737-1740. [2012/6/28]; Available from: http://www.imaginis.com/breast-cancerscreening-prevention/how-to-perform-breast-self-exam-bse. [2012/6/26]; Available from: http://www.cancerconcerns.com/TEST/test2.htm. Peng, T., et al. Algorithms for on-line monitoring of components in an optical tweezers-based assembly cell. Proceedings of IDETC/CIE 2006 ASME 2006 International Design Engineering Technical Conferences & Computers and Information in Engineering Conference September 10-13, 2006, Philadelphia, USA Ng, E. and Y. Chen, Segmentation of breast thermogram: improved boundary detection with modified snake algorithm. Journal Of Mechanics in Medicine and Biology, vol. 6, no. 2, pp. 123-136, 2006.

[2]

[3]

[4]

[5]

[6]

[7]

[8]

Figure 6: a) TH5 patient, b) binary mask with artifacts resulted of running algorithm on TH5 thermography case, c) final segmented breast in image (a) and d) another final segmentation for a TH5 patient. with (+) representing the nipple and the blue circle representing the areola outline, Dynamic areola detection and seed placement were utilized without any user input.

[9]

[10]

III.

CONCLUSION [11]

The segmentation of breasts is needed for proper evaluation of the TH levels. Although many attempts have been tried, the absence of clear boundaries in the upper region of the breasts provides a great challenge to any segmentation approach while approximately half of the breast cancers develop in that region which makes the proper segmentation of upper regions of the breast mandatory. With help of random walkers ability to overcome this problem breasts can now be accurately extracted from the image with good accuracy and reliability. This in turn contributes significantly to the accuracy and reliability of any CAD system used for thermography evaluation along with reducing the human intervention as much as possible thus enabling the use of CAD systems in

[12] [13] [14]

[15]

603