International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 1, January February 2013 ISSN 2278-6856

Detection of Sharp Contour element of the WBC and Segmentation of two elements like Nucleus and Cytoplasm
J.B.Nemane1, Prof. V. A. Chakkarwar2
1

Lecturer in IT Department, Govt. Poly. Ambad, (MS), India.

Lecturer in CSE Department, Govt. College of Engg. Aurangabad (MS), India.

Abstract: The quantity of WBC Cell is very important for

the doctor for doing diagnosis for various diseases. So, exact counting of cell plays very important role. Manual process of counting is a very time consuming task, and has less precision. So an automatic counting process is very important to get high precision results. For the counting of cell requires segmentation of the cell, for classifying purpose. In this paper, we propose a new algorithmto segment the image. Firstly, the proposed algorithm is used to remove noise Second algorithm used for the segmentation of image is by using watershed algorithm. The watersheld is just a method to determine separate regions can be counted efficiently.

The nucleus of each of the above types has a unique shape, and this is the most important feature used in cell classification. In addition to the shape of the nucleus, the phils category has granules with in the blood cell where as cytes category does not have granules. 2. MEAN SHIFT METHOD Let X = f {xi} ni=1 be a set of n data points in ddimensional space, Rd. The multivariate kernel density estimator with Gaussian kernel and a symmetric positive definite d& d bandwidth matrix H, computed at the point x is given by (1) Where (2) is the Mahalanobis distance from x to xi. By computing the gradient of (x) (3) After some algebra we have (4) Where (5) Assume now that the data points xi are extractedfrom an input image. Then the vector components ofxi contain both the spatial lattice information xsi = (xi; yi) T and range information xci = (ci1; ci2; ci3) T where ci1; ci2 and ci3 are three color components atposition (xi; yi). Then xi = (xTsi; xTri)T is a data pointin joint spatial-range domain. We assume that thebandwidth matrix H is diagonal having the diagonal terms equal to 2s for the spatial part and 2R for therange part. Then the Mahalanobis distance in Equ.(2) Can be rewritten as: Page 10

Keywords: WBC

Cytoplasm, Nucleus RBC, Segmentation,

1. INTRODUCTION
Cell segmentation is a challenging problem due to both the complex nature of the cells and the uncertainty present in video microscopy. Manual methods for this purpose are onerous, imprecise and highly subjective, thus requiring automated methods that perform this task in an objective and efficient way. Automated detection and classification of white blood cells is a major step in diagnosis of several diseases like Acute Lymphoblastic Leukaemia [1,2]. The traditional procedure requires a haematologist to manually countand classify the cells with the help of a microscope. An automated diagnosis system will alleviate the workload and the influence of subjective factors. Automated detection involves removal of red blood cells and platelets from the background. The main drawback of the existing methods is their inefficiency in handling cell images originating from different sources and environment [2]. There are three types of cells in normal human blood: red blood cells (RBCs), white blood cells (WBCs) [1, 7, and 8] and blood platelets. Generally, RBCs are simple and similar. While WBCs contain nucleus and cytoplasm and can be categorized into five classes: 1. Neutrophil 2. Eosinophil 3. Basophil 4. Monocyte 5. Lymphocyte Volume 2, Issue 1 January - February 2013

International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 1, January February 2013 ISSN 2278-6856
2. If zi is not labeled as 0, then return 3. Label zi as Idx 4. Recursively call a. Region Grow(zi , zi0, I dx), b. Region Grow(zi , zi1, I dx), c. Region Grow(zi , zi2, I dx), d. Region Grow(zi , zi3, I dx), 5. return (8) Let Y= {yi} i =1 and Z= {zi} i=1 be a set of data points in joint spatial-range domain of the processed data and the result filtered image respectively. The MS is formulated as follows. Algorithm 1: 1. Set t =0, and initial the error . 2. Set Y0i=xi for i=1 n. 3. Calculate the convergent value yi for i=1, ,n by 3.1Update Y it+1 3.2 If (yit , y it+1, H) > go to step 3. WATERSHED ALGORITHM For a segmentation purpose, the length of the gradient is interpreted as elevation information. Gradient Image Relief of the gradient
n n

(6) Where (7) The color distance in the RGB space is Defined as,

Let L = {1, 2...} be a set of labels of regions, and the unlabeled region is labeled as 0. According to the first assumption, the Region growing algorithm is designed as follows. Given a data point z which is labeled as Idx (I dx L) its neighbors are also labeled as, I dx if ||Y Tr || 2 2R.

4. Set Zi = (XsiT, Y Tr i), where Y r i is the range part of the convergent value Yi. Figure 1 (a) Gradient Image (b) Relief of the gradient Each data point yiis initialized at, xi and during calculating the convergent in step 3.1 and 3.2, y imoves iteratively along thegradient direction in both spatial and range domain. Finally, it converges to a local mode in the joint spatial-range domain. The advantage of the MS filter is that the image structure does not change during iterations. Then the MS will achieve better image structure preservation. The MS can remove noise while preserving edges or boundaries of the local structure by choosing the suitable 2sand 2R. 2.1 Region Merging Algorithm Given a data point of the filtered image z= ((x, y) T, ), its four neighbors are defined as follows: Z0 = ((x-1, y) T, ), Z1 = ((x+1, y) T, Z2 = ((x, y-1) T, Z3 = ((x, y+1) T, Figure 3 Watershed of the gradient (relief) ), ), ) Algorithm for Watershed Algorithm
(9)

Figure 2 Watershed of the gradient

On the other hand, z is called the center of its neighbors. Algorithm 2: 1. If ||Y Tr || > 2R then return.

1. Use the red channel for a better clearance of particles 2. Apply quality enhancement functions such as deblurring, denoising. 3. Use Granulometry as RBC size estimator 4. Estimate WBC size upon medical references and RBC size Page 11

Volume 2, Issue 1 January - February 2013

International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 1, January February 2013 ISSN 2278-6856
5. Proper binary conversion (Otsu method) 6. Use Proper edge detection method (Canny solution). 7. Merging binary and edge detection outputs (Otsu and Canny outputs) 8. Filling current image and make a solid objects 9. Separate WBCs from RBCs (Mask function) Improving the quality of the results of Watershed Algorithms. The quality of the best can be further improved in watershed algorithm by using colormap technique, some image processing filters and the Granulometric Analysis. 3.1 Color Map Technique For performing the segmentation of image in watershed algorithm the basic criteria is to get border of the cell distinctly so that we can detect the cell more accurately. To obtain the accuracy we can perform the channel split of the image i.e. R, B and G channel. The detect the border in R channel and then further respective steps can be perform on other channels of the image. A. Image Processing Filters The quality of the picture contents can be further improved by applying image processing filters like DE blurring (Enhancing Contrast) DE noising [6] (eliminating Blood platelets, and other unwanted noise) Smoothing (reducing the number of connected components). B. Granulatiy Analysis In Images RBCs are found a swell as WBCs. Hence we need two separated images for RBCs and WBCs. In this paper, we used size characteristics as an effective factor to distinguish two main objects. Granulometry algorithm [8], [9], [10] is divided into following Steps: a. morphological openings b. Structure Elements (SE model referring to object shapes) c. Maximum Size of SE (which is calculated according to primary estimation based on haematology science, and consideration of a noticeable point to enhance algorithm accuracy and speed) d. Increasing size SE up to max (Loop Function) e. Ps (k) = Sum of all the pixels in the image, opened with k-th SE. f. Normalized function N (k) =1- Ps (k)/Ps (0), where Ps (0) is the Ps of the original image, (SE Size=1 pixel). g. The Granulometric density function can be defined as G (k) = N (k+1) -N (k).

Figure 4 WBC Types in Blood Smear image Separation of WBC and RBC So far, a solid binary image is formed; everything is acceptable to apply the Watershed algorithm. Different objects including WBC, RBC are separated intolabelled regions for each object [12] . It is possible to enhance Watershed algorithm by employing watersnake. But in separated labelled regions there is not any discrimination between WBCs and RBCs, they cannot be distinguished from another. Thus before applying the Watershed method on the whole image we'd better form a new image only with WBCs based on size characteristics. 3.2 Proper Binary Image Watershed procedure needs binary image as input.Because of intrinsic characteristics of staining blood smear, particles color and their intensity can not be easily distinguished from the background [14]. All binary conversions are not 100% practical for blood smear images as shown in Fig. 6

Figure 5: (Left to Right): average binary, Otsu binary method 3.3 Proper Edge Detection Otsu algorithm has better efficiency in comparison withaverage algorithm, but it is not satisfactory. To achieve a desirable image, membrane detection should be followed [16]. Experimental results proved that canny filters demonstrate borders of WBC. 3.4 Circle Mask We notice that morphological operators such as Erosion or Dilation are not the best choices and always need some restrictions which are not available in all blood smear image such as agglutination between RBCs which make problems in using these simple morphological operations. A reliable method by our proposed method can consist of [17]: Applying Granulometry over blood smear image and saving density value pixels referred to RBC size (thats called intensity_area_prime_ max)

Volume 2, Issue 1 January - February 2013

Page 12

International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 1, January February 2013 ISSN 2278-6856
Estimating WBC size from RBC size calculated from Granulometry, medical references [18] and an acceptable marginal range: C1=RBC size*2 5 (margin) Moving surrounded mask over blood smear image and detecting the exact matching objects of the same size and then changing all its intensity into 0. Applying circular mask function and Granulometry in a closed loop by an initial radius value (C1=RBCsize*25(margin)) and then increasing integer radius (C1) value until we get similar density value pixels output as we had in first running Granulometry over main image.Saving WBC indicator in a new image variable [20].Omitting possible noises by applying median filter 5*5.

Figure 7 The Segmentation Experiments of Eosinophil.

4. EXPERIMENTS
A qualitative analysis of the output of the segmentation method proposed in Figure. From the specified results, we can say that the results of watershed algorithms are better than the mean shift algorithms. As the mean shift algorithm basically depends on the way the seed points are picked. The seed points are used for the cluster calculation. The formation of the cluster are basically depends upon the seed points. So, any fluctuation or some kind of the errors in proper selection of the seed points results in wrong clusters. In the watershed algorithms, the formation of the regions basically deals with the pattern analysis and it does not depend on the random picking up of seed points The final white-cell segmentation capability of the Watershed method has been evaluated by a human supervisor which evaluated the output of the Watershed algorithm checking 100 white cells. The obtained accuracy value is 93%.

Figure 8 The Segmentation Experiments of Lymphocyte.

5. CONCLUSION
The experimental results show that the MS filter can successfully remove noise from the WBC images while preserving the edges. As the mean shift algorithm basically depends on the way the seed points are picked.

Figure 9 The Segmentation Experiments of Monocyte.

Figure 6 The Segmentation Experiments of Basophil.

Volume 2, Issue 1 January - February 2013

Page 13

International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 1, January February 2013 ISSN 2278-6856
[9] Laurent Najman and Michel Couprie,Watershed algorithms and contrast preservation, France, 2003 [10] Cecilia Di Ruberto, Andrew Dempster, Shahid Khan, Bill Jarra, segmentation of blood image using morphological,IEEE,2000 [11] B.Ravi Kumar, Danny K Joseph, T.V Sreenivas, Teager energy based energy based blood cell segmentation, IEEE, 2002 [12] N. Theera-Umpon, White Blood Cell Segmentation and Classification in Microscopic Bone Mar row Images, Lecture Notes in Computer Science, vol. 3614, pp. 787796, August 2005. [13] B.R. Kumar, D.K. Joseph, and T.V. Sreenivas, Teager Energy Based Blood Cell Segmentation, International Conference on Digital Signal Processing, vol. 2, pp. 619-622, 2002. [14] E. Montseny, P. Sobrevilla, and S. Romani, A Fuzzy Approach to White Blood Cells Segmentation in Color Bone Marrow Images, IEEE Conference on Fuzzy Systems, 2004, vol. 1, pp. 173-178, 2004. [15] A. Sanpanich, W. Iampa, C. Pintavirooj, and P. Tosranon, White Blood Cell Segmentation by Distance Mapping Active Contour, ISCIT 2008, October 2008. [16] D. Barash and D Comaniciu, A Common Framework for Nonlinear Di_usion, Adaptive Smoothing, Bilateral Filtering, and Mean Shift, Image, vol. 22, no. 1, pp. 73-81, Jan.,2004. [17] Recommendations on Uniform Color Spaces, Color Di_erence Equations, Psychometric Color Terms, C.I.E. Supplement No. 2 to CIE publication no. 15(E-131) 1971/(TC-1.3) 1978. [18] W. Shitong and W. Min, A New detection Algorithm (NDA) Based on Fuzzy Cellular Neural Networks for White Blood Cell Detection, IEEE Trans.on Information Technology in Biomedicine, vol. 10,iss. 1, pp. 5-10, 2006. [19] Sobel I (1978) Neighborhood coding of binary images for fast contour following and general array binary processing. Comput Graph Image Process 8:127 135 [20] Prewitt JMS (1970) Object enhancement and extraction. Picture Processing and Psychopictorics, pp. 75 149

Figure 10 The Segmentation Experiments of Neutrophil. The sharpness of the boundaries of the filtered image is enhanced, but the segmented images are highly over segmented in the cytoplasm region .Finally, it can be seen that the effectiveness of the segmented images highly depends on the selected seed points. This paper also presented a new literature based on Granulometry, mask function and pre processing considerations to achieve a fully automated detection and segmentation of blood cells in order to separate labeled regions for Watershed algorithm. We have presented a method for WBC segmentation into nucleus and cytoplasm which helps in the diagnosis of several diseases. Our method successfully segments WBC images into nucleus, cytoplasm and background. Experimental results of Watershed algorithm indicate that the analysis of blood cells is achievable and it offers remarkable segmentation accuracy as compared to the mean shift algorithm. REFERENCES [1] http://en.wikipedia.org/wiki/Blood [2] JumaAl-Muhairy, Youssef Al-Assaf, Automatic white blood cell segmentation based on image processing, IFAC, 2005 [3] http://en.wikipedia.org/wiki/Anemia [4] http://en.wikipedia.org/wiki/Agglutination (biology) [5] R. Safabakhsh, F. Zamani, A Robust MultiOrientation Gabor Based System for Discriminating Touching White and Red Cells in Microscopic Blood Image , Computer Engineering and IT Dept., AmirKabirUniversity of Tech, IEEE, 2003 [6] HieuTatNguyen, Marcel Worring, Rein van Boomgaard, Watersnakes: Energy-Driven , IEEE TRANSACTIONS, 2003 [7] Couprie M., Bertrand G., Topological Greyscale Watershed Transformation, proceedings of the V SPIE USA, July 1997, Vol.3168, pp 136-146 [8] Jaesang Park and James M. Keller, Snakes on the Watershed, IEEE TRANSACTIONS, 2001 Volume 2, Issue 1 January - February 2013

Page 14