Document Sample

International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING & 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME TECHNOLOGY (IJCET) ISSN 0976 – 6367(Print) ISSN 0976 – 6375(Online) IJCET Volume 5, Issue 1, January (2014), pp. 52-61 © IAEME: www.iaeme.com/ijcet.asp ©IAEME Journal Impact Factor (2013): 6.1302 (Calculated by GISI) www.jifactor.com A NOVEL METHOD FOR CLUMPED PARTICLES SEPARATION IN MICROSCOPIC IMAGES A. AL-Marakeby Systems and Computers Engineering Dept., Faculty of Engineering, Al-AzharUniversity, Cairo, Egypt ABSTRACT Microscopic image processing has been recently applied to many fields such as blood cell counting, tissue analysis and material microstructures analysis. Cell counting is an important process which helps the diagnosis of many diseases. A main problem in cell counting and also in other microscopic image analysis is the overlap between objects. This overlap and connection between particles reduces the accuracy of counting or gives classifications errors. Many techniques have been used to isolate the objects but the segmentation process still has many errors. In this research,a novel method for separation of clumped particle is developed. This method depends on iterative hypothesis and verification technique. Extracted features are used to generate a set of hypotheses, depending on particles boundary and colors. These hypotheses are verified using specific measures and distances, and then the best hypothesis is chosen. This method is efficient for generic shape analysis and matchinginstead of the assumption of circular or elliptical particles shapes.In addition to that, the proposed technique overcomes the problems of noisy and cut boundary, and the problems of computational complexity in some other techniques. This method is compared to circle and ellipse detection methods and higher accuracy is achieved. Keywords: Particles Separation- Segmentation-Microscopic Images – Blood Cell Counting. 1. INTRODUCTION Segmentation is very important stage for most successful image processing and computer vision applications. The errors in the segmentation are diffused to the next stages and cause the low performance of the final results. Microscopic image processing is a field concerns with the analysis and processing of images obtained from a microscope. Microstructures analysis of material, complete blood count (CBC), and tissue analysis are some examples from many applications of microscopic image processing. The automation of blood cells counting has the advantages of 52 International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME reducing costs and increasing speed and precision. The correct segmentation of blood cells represents a challenge when the cells are overlapped. Connected component analysis is a simple technique used to detect cells, but it gives correct results only when the cells are far and separated from each other's. Fig.1 shows some of connected cells which are in interpreted as a single cell. Venkatalakshmi et.al. used Hough transform to detect and count red blood cell[13]. Hough Transform has better performance than connected component analysis but still suffer from many problems. Cell shape can be non-circular , and some overlaps structures are difficultto be detected by Hough transform. Fig.1 Overlapped Cells Sharif et. al. used masking and watershed algorithm to segment RBC[6]. Huang used watershed segmentation combined with morphological operators to separate overlapping blood cells[5]. Theerapattanakul et al. used active contours for the segmentation process of blood cells[7]. Other techniques based on SVM, neural networks, and Gabor filter, can be found in [8][9][11][14].In this research a novel method is developed which solves many problems found in other techniques. This method depends on iterative hypothesis and verification technique. Extracted features are used to generate a set of hypotheses, depending on particles boundary and colors. These hypotheses are verified using specific measures and distances, and then the best hypothesis is chosen. This method is efficient for generic shape analysis and matchinginstead of the assumption of circular or elliptical particles shapes. This paper is organized as follow: section 2 explains the circle and ellipse detection, section 3 explains the proposed technique working with generic shape detection, section 4 illustrates the color model used in the separation process , section 5 gives the results and discussions, and finally section 6 gives the conclusion. 2. CIRCLE AND ELLIPSE DETECTION An edge detection stage is required before running the shape detection algorithms. Edge detection converts the color or gray image into binary image, where the boundaries of objects are detected. There are many edge detection algorithms such as Sobel, Prewitt, Roberts and Laplacain, Canny edge detection. In this research Sobel edge detection is used as shown in fig.2.Compared to other edge operator, Sobel has two main advantages: 1- Since the introduction of the average factor,it has some smoothing effect to the random noise of the image. 2-Because it is the differential of two rows or two columns, so the elements of the edge on both sides has been enhanced,so that the edge seems thick and bright[15]. 53 International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME Fig. 2 Sobel edge detection From many circle and ellipse detection techniques, Hough Transform (HT) is used to detect particles. The main advantage of the HT in extracting circle and ellipse, is its robustness against discontinuous or missing data points. This is becausethe HT does not require the connectivity of all the contour pointsof circle or ellipse [1]. 2.1 Circle detection HT converts features points in the 2Dimage into parameter space. HT for circle detection depends on the detection of 3 parameters: the circle center coordinates a,b and the radius r. The circle equation is given by: (1) This equation can be represented in parametric space as: x = a + r * cos(θ), y = b + r*sin(θ) (2) The main problem with HT for circle detection is the computational complexity and memory requirements. Many techniques are used to reduce the complexity and memory requirements of HT. Randomized Hough Transform (RHT)keeps standard Hough Transform (SHT)'s robustness,but costs much less memory space and computing time than SHT.[4] [12]. RHT each time samples 3 edge points of the image randomly and calculates 3 parameters (center position and radius) of the circle to be detected instead of mapping all theedge points to parameter space in SHT [4].Gradient methods are used to restrict the center of the circle and also the radius to some value extract from the gradient of the curves. Xing et. al. used one-dimensional circle Hough transform depending on that the circle center is on the gradient line of circle edge points[16]. Fig.3 shows using HT to separate particles in microscopic images Fig.3 Separation of particles based on HT circle detection. 54 International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME 2.2 Ellipse detection Ellipse is a more general shape model than circle model. For this system many particles have elongated shapes which make the circle HT is not accurate. The problems in circle detections are increased when moving to ellipse detection. While circle detection depends on 3 parameters, ellipse detection has 5 parameters. Fig.4 shows the different parameters required for specifying an ellipse. These parameters are the center of the ellipse, the major axes, the minor axes and the angle of the ellipse with X axes. The following equation represents the parametric equation of the ellipse: (3) Evidently this approach would require solving the equation for five different pointson the ellipse, thus mapping xy points to a five dimensional space, (hence having to manage a fivedimensional accumulator). This approach is not only memory expensive but also computationally intensive, asthe algorithm in its brute search form would have complexity[2]. As in circle detection, many algorithms are used also in ellipse detection to reduce its complexity. Randomized Hough Transform (RHT) is used in ellipse detection, where an n-tuple ofpixels in the image is mapped to a single point in the parameterspace[10].Han et al. simplified ellipse detection by dividing the image into several sub-images by the properties of ellipse, andthen point pairs are chosen from the sub-images to calculate the parameters of ellipses[3]. Alex et al. considered every pair of edge pixels as possible end points of the major axis of a hypothetical ellipse. After that, all other edge pixels will be used to vote on the half-length of the minor axis of this hypothetical ellipse[1]. Fig.5 illustrates the separation of particles using HT for ellipse detection. L2 O L1 Fig.4 ellipse parameters Fig.5 Separation of particles based on HT ellipse detection. 3. GENERIC SHAPE DETECTION SYSTEM The circular and elliptical assumption for particles shapes is not valid for all particles. Many particles have deformed shapes especially when they are touched or connected with other particles.These non-circular and non-elliptical shapes reduce the accumulator value for this object. Object detection depends on thresholding the accumulator values where value less than kis not considered as an object.Hence, to detect these objects, a threshold for parameter space should be 55 International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME decreased, causing the detection of false objects. Fig.6 shows these problems where the shapes of the particles are not exactly fit to the ellipse or circle assumption. Fig.6 non-circular and non-elliptical shapes The technique used in this research depends on generic shape analysis and matching. The boundaries of these shapes are extracted from different images and stored in a database. The distances between the particle under inspection and all shapes in the database are measured. An iteration process starts with the generation of assumptions and verification of these assumptions. The assumptions covers different shapes exists in the database, different dimensions and coordinates, and the overlap of particles (single particle, two particles, three particles …….etc). The verification of assumptions depends on an evaluation function which has terms for many factors. 3.1 Boundary Extractionand Shapes database Many shape descriptors exists such as Fourier descriptors, wavelet descriptors, signatures, moments,…etc. Most of these descriptors depend on the closed contours. Noisy contours with cut regions or concatenated objects cause problems for these techniques. In other words the Fourier descriptor of a special boundary is completely different with the Fourier descriptor of the half of this boundary. Hence, in this research the spatial domain is better and easier in working with it. We can conclude that this shape consists of two concatenated particles by a simple analysis. The boundaries of different particles with different shapes are extracted as shown in fig.7. The shapes are selected to avoid similar shapes. These boundaries are stored in a database. The database contains the full resolution boundaries and different resolutions or modified boundaries can be obtained after that by sampling or filters. Fig.7 boundary extraction 56 International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME 3.2 Boundary Distance The distance measure between the boundary of the particle under inspection and a stored template in the database is important for obtaining good performance. The small value of this value indicates better interpretation and robust results. The large value indicates unknown or uncertain interpretation and hence increases segmentation and analysis errors. The distance between template and boundary is measured by the summation of all distance from each point in the boundary to the template. The template is approximated by a set of line segments. Each segment linear equation is extracted and the distance between a point and a line is given by equation 4. The solid line in fig.8 represents the template while the dotted line represents the boundary of the particle. (4) Fig.8 boundary distance 3.3 Hypothesize-and-test framework A process of generating assumptions and testing if these assumptions are valid or not, is used to separate the particles and determine the location, shape, and dimensions of these particles. The search space of this process is very large and many techniques are used to find the solution efficiently with high accuracy. The search space include finding the overlap type (no overlap – two particle – three -….etc), the location of centers and the shape of particles. After each assumption generation the verification process is applied and the distance is measured as discussed in the previous section. To reduce the search space with affecting the accuracy, the boundary color analysis, gradient analysis, and maximum curvature analysis are used. Color Analysis is detailed and illustrated in section 4. The gradient analysis of the boundary is very important and reduces the search space too much. The center of the particle is deduced from the curvature of the boundary, and hence no need to test other assumptions. The center may be not accurate but searching around the determined location gives very accurate results. The maximum curvature analysis detects some points which are candidate to be the connection of two or more particles. Some noise and deviated boundaries may generate false points but the verification stage will discard any false assumptions. Fig. 9 illustrates the assumption generation for one and two particles. Fig.10 shows the maximum curvature points (blue dots). 57 International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME Fig.9 assumptions and verifications Fig.10 maximum curvature points (blue dots) The verification stage depends on many factors. The main factor is the boundary distance discussed in section 3.2. This distance has a clear discrimination between single or multiple particles. The other factors used in generating hypotheses are used again to measure and verify the assumption. The color model and the maximum curvature points are used but with a smaller weight while these factors are not dominant and may give false results. Some other heuristics are used to accelerate the process of hypotheses generation and also in the verification of these assumptions. Fig.11 shows the segmentation of three clumped particles. Fig. 11 segmentation of three overlapped particles 58 International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME 4. COLOR MODEL Working with edges and boundaries only, discards a lot of useful information which can be important for solving the overlap problem. The particles color model has a significant impact on the performance of the system. As shown in fig.6 the color and intensity are varied overthe area of the particles. A light spot can be found near the center of the particle, and a dark region exists at the overlapping area. In Fig.12the color analysis for blood cell image is illustrated. Figure 12: Color and intensity variations over particles This analysis is based on different thresholds for colors and each level or cluster is assigned a specific color. The result shows that two connected particles has two spots while each single particle has a unique spot. This analysis is not usually true and it is used for generating hypotheses. The iterative process for hypothesize and test can verify or discard the assumptions generated by the color model as discussed in section 3. 5. RESULTS AND DISCUSSIONS A set of blood cell images is used to test the developed system [17]. Four techniques are used: 1- connected component analysis, HT for circles, HT for ellipse, and the generic shape detection discussed in section 3. Table.1 reports the accuracies of these techniques. Table.1 Accuracy of different techniques for counting blood cells Technique Accuracy Connected component Analysis 88% Hough Transform for circle detection 92% Hough Transform for Ellipse detection 94% Generic Shape 98% The connected component analysis has no separation at all, so any clumped particles are interpreted as a single particle. HT for circle can separate particles but it is restricted to the circular shapes and sometimes generates false object detection. The HT elliptical system is similar to circle detection but with more generalization to elliptical shapes. The proposed technique of generic shape detection system has many improvements and robustness. It can detect more shapes, deals with noise and cut boundary, and has higher accuracy. 59 International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME 6. CONCLUSION Separation of clumped particles represents a problem for obtaining better results for microscopic image analysis. The assumption of particles shapes to be standard shapes such as circles and ellipses is not valid assumption and causes errors and reduced performance. The proposed technique in this research has better performance and deals with different shapes effectively. Using color information to separate overlapped particles increases the accuracy and improves the performance. The hypothesize and test framework used in this research reduces the search space based on bottom up features interpretation and some other heuristics. REFERENCES [1] Alex Yong, Sang Chia, Maylor K. H. Leung, How-Lung Eng, SusantoRahardja, ELLIPSE DETECTION WITH HOUGH TRANSFORM IN ONE DIMENSIONAL PARAMETRIC SPACE. IEEE International Conference on Image Processing, 2007. [2] Cosmin A. Basca, MihaiTalos, and Remus Brad, Randomized Hough Transform for Ellipse Detection with Result Clustering, Serbia & Montenegro, Belgrade, November 22-24, 2005. [3] Han Fei, GuoYanling, Wang Lili, A New Ellipse Detector Based on Hough Transform, Second International Conference on Information and Computing Science2009. [4] Huadong Sun, Yaobin Mao, Ningjian Yang, Dian Zhu,A Real-Time and Robust Multi-Circle Detection Method Based on Randomized Hough Transform, International Conference on Computer Science and Information Processing (CSIP) 2012. [5] HUANG Jiandeng, An Improved Algorithm of Overlapping Cell Division, International Conference on Intelligent Computing and Integrated Systems (ICISS), 2010 [6] J. M. Sharif, M. F. Miswan, M. A. Ngadi, MdSahHj, Salam, MuhammadMahadi bin Abdul Jamil,Red Blood Cell Segmentation Using Masking and Watershed Algorithm: A Preliminary Study,2012 International Conference on Biomedical Engineering (ICoBE), 2012, [7] J. Theerapattanakul, J. Plodpai, C. Pintavirooj, AN EFFICIENT METHOD FOR SEGMENTATION STEP OF AUTOMATED WHITE BLOOD CELL CLASSIFICATIONS,. IEEE Conference TENCON 2004. [8] JiyeQian, Bin Fang, Chunyan Li and Lin Chen, Coarse-to-Fine Particle Segmentation in Microscopic Urinary Images, 3rd International Conference onBioinformatics and Biomedical Engineering , 2009. [9] Nasution, AMT, EK Sury`aningtyas, Comparison of Red Blood Cells Counting using two Algorithms: Connected Component Labeling and Back projection of Artificial Neural Network, IEEE IPGC 2008. [10] Robert A.McLaughlin, Randomized Hough Transform: Better Ellipse Detection , IEEE TENCON - Digital Signal Processing Applications, 1996 [11] R. Safabakhsh, F. Zamani,A Robust Multi-Orientation Gabor Based System for Discriminating Touching White and Red Cells in Microscopic Blood Image, Information and Communication Technologies, 2006. [12] SI-YU GUO, XU-FANG ZHANG, FAN ZHANG, ADAPTIVE RANDOMIZED HOUGH TRANSFORM FOR CIRCLE DETECTION USING MOVING WINDOW, Proceedings of the Fifth International Conference on Machine Learning and Cybernetics, Dalian, 13-16 August 2006 [13] Venkatalakshmi.B1, Thilagavathi.K, Automatic Red Blood Cell Counting Using Hough Transform, Proceedings of 2013 IEEE Conference on Information and Communication Technologies 2013. 60 International Journal of Computer Engineering and Technology (IJCET), ISSN 0976- 6367(Print), ISSN 0976 - 6375(Online), Volume 5, Issue 1, January (2014), © IAEME [14] Wei-Liang Tai1, Rouh-Mei Hu2,5, Han C.W. Hsiao, Rong-Ming Chen, and Jeffrey J. P. Tsai, Blood Cell Image Classification Based on Hierarchical SVM, IEEE International Symposium on Multimedia, 2011. [15] WenshuoGao, Lei Yang, An Improved Sobel Edge Detection, 3rd IEEE International Conference on Computer Science and Information Technology (ICCSIT), 2010 [16] Xing Chen, Ling Lu, Yang Gao, A New Concentric Circle Detection Method Based onHough Transform, The 7th International Conference on Computer Science & Education (ICCSE) 2012. [17] http://library.med.utah.edu/ (2013). 61

DOCUMENT INFO

Shared By:

Categories:

Tags:

Stats:

views: | 2 |

posted: | 2/1/2014 |

language: | Unknown |

pages: | 10 |

OTHER DOCS BY iaemedu

How are you planning on using Docstoc?
BUSINESS
PERSONAL

By registering with docstoc.com you agree to our
privacy policy and
terms of service, and to receive content and offer notifications.

Docstoc is the premier online destination to start and grow small businesses. It hosts the best quality and widest selection of professional documents (over 20 million) and resources including expert videos, articles and productivity tools to make every small business better.

Search or Browse for any specific document or resource you need for your business. Or explore our curated resources for Starting a Business, Growing a Business or for Professional Development.

Feel free to Contact Us with any questions you might have.