Experimental Study of the Effect of Different Scanning Patterns for Blood Smear Analysis

Transcription

1 Experimental Study of the Effect of Different Scanning Patterns for Blood Smear Analysis Khaled El Bannan Mostafa M.A. Mohamed Behrouz H. Far Christopher Naugler Department of Electrical and Computer Engineering Department of Biomedical Engineering Helwan University Cairo, Egypt Department of Electrical and Computer Engineering Department of Pathology and Laboratory Medicine Cumming School of Medicine Abstract Blood smear analysis is a very important and typical clinical exam performed using an optical microscope. Most clinical laboratories have a standard scanning pattern for cell finding under a microscope. The commonly used scan pattern for this examination is the Zigzag or the continued raster pattern either horizontal or vertical. In this research, various scanning patterns were investigated and compared to the standard vertical and horizontal continued raster pattern used in hematology smear analysis. The scan patterns employed are basic scans generated by the SCAN methodology to scan the workspace. In this paper we introduce alternate scan patterns that would reduce the slide scan time for increased workflow in a lab environment. However, due to the random nature of the scattered blood cells, the comparison of various scan patterns has manifested an improvement of 14% in the overall scan results. Keywords: Image scanning; Scanning patterns; SCAN methodology; Hematology testing; Peripheral blood smears. 1. INTRODUCTION Blood smear examination under the microscope can present an abundance of information to physicians. Abnormalities can be revealed using this method and consequently can help specialists in making a differential diagnosis [1]. Blood is smeared on a glass slide as a first step for examination under an optical microscope (illustrated in Fig. 1). To differentiate between various blood cells types, the blood smear has to be stained [2], in order to assess the morphology of white blood cells (WBC) (shown in Fig. 2a and 2b). The standard method for examining the WBC is to view the smear at 100x magnification (oil emersion). However, the smeared blood consists of several areas (1, 2 and 3) as shown in Fig. 1. Not all areas are appropriate for examination, as the distribution of cells varies in each particular area [2]. The red blood cells (RBC) are clumped together in area 3, and the WBC appears small and dark, obstructing analysis. For area 1 located at the edge of the slide feathering appears and RBC are too widely spread for any useful examination. Finally, area 2 is the optimal area for examination as the RBC are equally spaced and the WBC show optimal morphology [2]. Feather edge Area 1 Area 2 (Good working area) Area 3 Poor viewing area (Thick area) Fig.1. Different areas of peripheral blood smear The peripheral blood smear is first examined with a 10x magnification to pursue ranges where cells are evenly distributed, which is called the good working area (Fig. 2a) and excluding bad areas such as in (Fig. 2b) [2]. Some studies have been conducted to locate the good working area in a peripheral blood smear. As introduced in [3], the method was a quick detection of images scanned under low magnification, then applying algorithms to extract morphology, and number of connected components. Furthermore, other work was presented in [4-7] using image-processing methods to find the good working area. The WBC deferential count (part of the standard peripheral blood smear examination) requires counting 100 white blood cells to complete the exam [1]. The standard continued raster scanning pattern can be executed horizontally or vertically for this examination to scan the good working area of the smear (Fig. 3). It can be noted that using this type of scan pattern (called sometimes zigzag) is not the most efficient method for the task to be completed, so other scan patterns were explored in this experimental investigation. The selected patterns for this work are usually used in image encryption and decryption techniques as presented in [12-16] / copyright ISCA, CAINE 2016 September 26-28, 2016, Denver, Colorado, USA

2 by utilizing SCAN methodology in this procedure (different SCAN patterns are shown in Fig. 4). Fig. 2a. Equally distributed red blood cells using 10x magnification (area 2). Good area for analysis. Fig. 2b. Clumped red blood cells using 10x magnification (area 3). Bad area for analysis Fig. 3. Zigzag pattern for scanning the slide good area This paper is organized in 5 sections. Section 2 is a brief overview of the scanning algorithms and patterns. Section 3 describes the experimental setup, data processing, and methods used. Section 4 presents the experimental results. Section 5 discusses the conclusion. 2. Scanning Algorithms and Patterns Area # 2 Area # 3 Computer supervision can be utilized in the automation of scanning a peripheral blood smear using several methods. Search algorithms can be programmed to construct the scanning pattern path to control the automation for optimal search results. However, most search algorithms need a learning mechanism or available datasets to succeed in their search [8, 9]. No predefined dataset exists in this case, and furthermore cells are distributed randomly, so these types of search algorithms are expected to fail. One of the solutions can be using random search algorithms [10]. These algorithms take a step in one random direction, and reassess the next move depending on the previous searched results [10]. It is noted that these algorithms can take some extended time to find an answer due to its iterative nature. One of the main goals of automating the peripheral blood smear scan is to minimize the search time. Therefore, these types of algorithms cannot be implemented fully in this case because of poor performance. The solution we propose is to examine different predefined search patterns (i.e. space filling curves) and evaluate their overall performance empirically. Many research attempted to explore these efficient scanning curves as presented in [20-23]. There are several types of space filling curves, and the individual curve can only access any element of a two-dimensional array once [11], refer to Fig. 4. Two types of space-filling curves (SFC) are Peano and Hilbert curves as presented in [11]. First, Piano's is a twodimensional curve that splits a unit square into a 3x3 square grid cells, while concurrently partitioning the unit interval into nine subintervals. The formed subinterval is paired to a cell where the curve crosses the cells in a distinct order. The method is repeated to each subinterval cell pair, so that the path crossing is similar to the previous [11]. Hilbert's curve is comparable to Piano's in the cell traversal path but the unit square is subdivided into four parts [11]. Another type of SFC generator is the SCAN methodology. The SCAN is a two-dimensional space filling procedure that can produce an ample collection of scanning paths [16-20]. This type of algorithm can be employed as generic data accessing strategies as well as scan procedures for image processing [12]. The above mentioned research utilizes SFC either in image encryption or data accessing strategies. In this paper SFC various scan patterns are used to compare their performance to the standard scan pattern. Fig. 4 shows different simple SCAN patterns that construct the SCAN alphabet with the 15 SCAN letters [12]. The letters are classified to follow a scan patterns (orders) according to [12]: First, {R } is defined as raster scan pattern, {C } is continued raster scan pattern, {D } is diagonal scan pattern, {E } is diagonal with parallel returns, {A } is right orthogonal scan pattern, {I } is spiral in scan pattern, {O } is spiral out scan pattern, {L } is continued orthogonal scan pattern, {S } is vertical symmetric by row scan pattern, {H } is vertical symmetric by column scan pattern, {Y } is main diagonal symmetric scan pattern, {W } is diagonal symmetric by secondary lines scan pattern, {Z } is zeta scan pattern, {B

3 } is right bracket scan pattern, and finally {X } is xi scan pattern. Fig. 4 shows the graphical presentation of each scan letter. The SCAN language family consists of different versions (i.e. Simple SCAN, Extended SCAN, and Generalized SCAN). These versions are equipped with sets of rules and transformations that can compose a simple distinct scan patterns up to very complex ones. For detailed descriptions of SCAN languages refer to [12-16]. forming a matrix 10 x10 as shown in figure 5 and 6. This configuration simulates the condition of digitizing this area of the slide using a 100x objective lens (with total magnification of 1000x) and oil immersion. Actual processing of cell counting is done on image samples as presented in figure 5 and 6 which gives a total image database number of 756 images that were processed. Only basic SCAN patterns were employed in this research to acquire the WBC count in peripheral blood smears. The fixed simple SCAN pattern scans the slide good working area and the result is compared to the performance of the standard continued raster (zigzag) pattern. Fig. 5. Sample images digitized using 10x object lens with spiral out {O } curve drawn over it. Fig. 4. Different basic SCAN patterns 3. Experimental Setup 3.1 DATASET A number of slides were digitized (21 slides) using an optical microscope and a digital camera. The good working area of the slide was targeted using a 10x objective lens (with overall magnification of 100x) and the slides were digitized in a form of a 2 x 18 matrix of images for each slide. Every matrix contained 36 images of 710 x 480 pixels resolution as shown in Fig. 2a. Moreover, in figure 5 and 6 it can be noted that purple dots are randomly scattered over the image area. These dots are the stained white blood cells that will be counted using different scanning patterns to obtain the maximum number of cell in minimum step count. Each image is divided into smaller squares. The total number of squares constructing the image is 100 squares Fig. 6. A {D } curve applied to the sample image. 3.2 SCAN PATTERNS USED Thirteen different Simple SCAN patterns namely: R, C, D, E, A, I, O, L, S, H, Z, B and X (refer to Fig. 4) were used on the 10 x 10 blocks of each image. Furthermore, an extra four patterns were devised to scan the 10 x 10 matrix diagonally to segregate any potential patterns in the WBC population grouped in one place. Fig. 7 illustrates the shapes of the four extra patterns used. First, a diagonal raster path with a different start point at the corners was used to segregate any potential WBC population in the corners. Second, was a center starting diagonal raster propagating starting upwards to the top and ending downwards to the bottom segregating population at the center in one direction. Third, a center diagonal starting raster with symmetrical propagation segregated any population of cells at the center. Finally, the

4 Performance (%) Performance (%) above mentioned three curves were repeated using basic scan letters X, Z, and B as building blocks for the raster scan. 3.3 IMAGE AND DATA PROCESSING A MATLAB program was developed to import the image samples and to divide the image area into a 10 x 10 square blocks. Each block can be considered as a step in the path of the scanning pattern. One step (block) may contain a number of white blood cells or it may be empty. In order to achieve this counting scheme using the different scans mentioned before, the digitized image has to be processed further to segment the white blood cells from each image. Image processing filtering techniques were used on the images as presented in [24-29] which automate the counting of white blood cells for a large set of data. The output of this MATLAB program is a 10 x10 matrix populated with the number of the WBCs in each element of the matrix. All the SCAN patterns and the extended ones mentioned earlier were coded in another MATLAB program. This program imported the populated matrix previously prepared in the previous step and processes it with all scan patterns to acquire a predetermined number of WBC. This predetermined number of WBC is a variable input for the program to acquire. The output of this program is the number of the acquired WBC and their corresponding steps count. To evaluate the performance of each scan curve, the number of steps needed to reach to the required WBCs count was calculated, then the steps of every scan curve were compared to the standard zigzag curve in a percentage form. First propagation Symmetric propagation inclined, X, and spiral in scan patterns performed better than the other ones, with overall performance improvement of 14%, 11.9%, and 11.5% respectively from the standard vertical continued raster. Fig. 9 is a comparison between the 12 best performing scan patterns and the horizontal continued raster. The graph indicates a maximum of 6.4% increased performance of the raster diagonal pattern over the standard horizontal pattern at low WBC counts. The performance of all curves declines when reaching the average number of WBC per image. What is observed from the graph is that the diagonally inclined, X, and spiral in scan patterns performed better than the other patterns, with overall an performance improvement of 6.4%, 4.3%, and 3.9% respectively from the standard horizontal continued raster No. of Cells C vert. C horz. Cen R Diag Sym Cen R Diag Cen diag B sym Cen diag X sym Cen diag Z sym E I R Diag 10_10 X Fig. 8. A comparison between the standard scan pattern (Vertical) and the best performing scan pattern set. The X-axis shows the WBC count, and the Y-axis show the percentage performance of each curve Start point Start point Propagation direction propagatio Second propagation Start point propagatio Fig.7. Extra diagonal scan patterns with different propagation directions. 4. Results The 756 sample images from 21 blood slides were processed by all scan patterns counting curves. The average WBC count for each image was found to be around 13 WBCs from the available data. The required number of WBCs to count in each image was varied from 1 to 13 and the results were retrieved and graphed. Fig. 8 is a comparison between the 12 best performing scan patterns and the standard vertical continued raster (zigzag). The graph indicates a max of 14% higher overall performance for the raster diagonal pattern over the standard one at lower WBC count. The performance of all curves declines when reaching the average number of WBC per image. What is observed from the graph is the diagonally No. of Cells C Horz. Cen R Diag Sym Cen R Diag Cen diag B sym Cen diag X sym Cen diag Z sym E I R Diag 10_10 Fig. 9. A comparison between the standard scan pattern (Horizontal) and the best performed scan pattern set. The X-axis shows the WBC count, and the Y- axis show the percentage performance of each curve. The above results yield a better overall performance of the horizontal C raster scan pattern over the vertical C raster one by maximum of 7.6% at low WBC counts. X

5 5. Conclusion Various scanning patterns were investigated and compared to the standard vertical and horizontal continued raster pattern used in hematology smear analysis. The scan patterns employed were scans generated by the SCAN methodology and the diagonally inclined raster scan patterns. The comparison results revealed a maximum of 14% and 6.4% outperformance of the diagonally inclined raster scan over the vertical standard scan pattern, and the horizontal standard scan pattern respectively at low WBC counts. Furthermore, SCAN language generated scan patterns namely the I pattern outperformed the vertical standard scan pattern a by maximum of 11.5 %, and 3.9% over the horizontal standard scan pattern respectively at low WBC counts. This result presents the efficiency of the introduced scans patterns that can reduce the slide scan time and increase workflow in a lab environment. However, further investigation indicated an overall performance of 7.6 % in favor of the horizontal standard over the vertical standard one at low WBC acquired count, which indicates that the horizontal standard scan pattern is more efficient over the vertical one. This work was performed on parts of the good working area on a per image basis. Further investigation is needed with larger sample dataset to confirm the acquired results over the whole good working area and this will be is the basis of the future work. ACKNOWLEDGMENT This work is supported by Mitacs, and Smart Labs Limited. REFERENCES [1] Kathy W. Jones," Evaluation of Cell Morpholgy and Introduction to Platelet and White Blood Cell Morphology", 1993,Chapter 5 p [2] "Hematology sequence, Blood Labs", files/bockenstedt - Lab Info.pdf. [3] Angulo J, Flandrin G, " Automated detection of working area of peripheral blood smears using mathematical morphology", Anal Cell Pathol. 2003;25(1): [4] W. Xiong et al., "Automatic working area classification in peripheral blood smears using spatial distribution features across scales," Pattern Recognition, ICPR th International Conference on, Tampa, FL, 2008, pp [5] W. Xiong et al., "Cell Clumping Quantification and Automatic Area Classification in Peripheral Blood Smear Images," Biomedical Engineering and Informatics, BMEI '09. 2nd International Conference on, Tianjin, 2009, pp [6] Wei Xiong, et al.,"automatic Area Classification in Peripheral Blood Smears", IEEE Transaction on Biomedical engineering. 2010; 57(8). [7] E. A. Mohammed, B. H. Far, M. M. A. Mohamed and C. Naugler, "Automatic working area localization in blood smear microscopic images using machine learning algorithms," Bioinformatics and Biomedicine (BIBM), 2013 IEEE International Conference on, Shanghai, 2013, pp [8] A.A. EL-Harby, " New Square Scan Algorithm", GVIP Journal 2005; 5(9): [9] Er. Waghoo Parvez, Er. Sonal Dhar, " Path Planning Optimization Using Genetic Algorithm A Literature Review", International Journal of Computational Engineering Research 2013; 3(4): [10] Zelda B. Zabinsky," Random Search Algorithms", Department of Industrial and Systems Engineering, University of Washington, online course, April5, [11] Herman Haverkort, "An inventory of three-dimensional Hilbert spacefilling curves ", Sept. 13, 2011, [12] Nikolaos G. Bourbakis, Christos Alexopoulos, and Allen Klinger, " A Parallel Implementation of the SCAN Language", Comput. Lang. 1989; 14(4): [13] S.S. Maniccam, N.G. Bourbakis,"Image and video encryption using SCAN patterns", Pattern Recognition 2004; 37: [14] S.S. Maniccam, N.G. Bourbakis, " Lossless image compression and encryption using SCAN ", Pattern Recognition 2001; 34: [15] Bourbakis, N.G., "Image data compression-encryption using G-scan patterns," in Systems, Man, and Cybernetics, Computational Cybernetics and Simulation., 1997 IEEE International Conference 1997; 2: [16] S.S. Maniccam, N.G. Bourbakis," Lossless compression and information hiding in images", Pattern Recognition 2004; 37: [17] Chao-Shen Chen; Rong-Jian Chen, "Image Encryption and Decryption Using SCAN Methodology," in Parallel and Distributed Computing, Applications and Technologies, PDCAT '06. Seventh International Conference 2006; 0: [18] S.S. Maniccam, N.G. Bourbakis," Image and video encryption using SCAN patterns", Pattern Recognition 2004; 37: [19] Smila Mohandhas and Sankar. S, " NOVEL ALGORITHMS FOR FINDING AN OPTIMAL SCANNING PATH FOR JPEG IMAGE COMPRESSION ", International Journal of Emerging Technology in Computer Science & Electronics (IJETCSE) 2014; 8(10): [20] Reza Moradi Rad, Abdolrahman Attar, and Reza Ebrahimi Atani, " A New Fast and Simple Image Encryption Algorithm Using ScanPatterns and XOR", International Journal of Signal Processing, Image Processing and Pattern Recognition 2013; 6(5): [21] Megha Kashyap, Dipesh Kumar Sharma, Love Verma, " A novel combined approach of cryptography and digital watermarking for enhanced security", International Journal of Engineering Sciences Research-IJESR 2013; 4(7274): [22] Ouni, T.; Lassoued, A.; Ktari, J.; Abid, M., "Adapted Scan Based Lossless Image Compression Scheme," in Signal-Image Technology and Internet-Based Systems (SITIS), 2011 Seventh International Conference 2011: (0): [23] Ouni, T.; Lassoued, A.; Abid, M., " Lossless image compression using gradient based space filling curves (G-SFC)", Signal, Image and Video Processing 2015; 9(2): [24] M. Mohamed, B. Far and A. Guaily, "An efficient technique for white blood cells nuclei automatic segmentation," Systems, Man, and Cybernetics (SMC), 2012 IEEE International Conference on, Seoul, 2012, pp [25] M. Mohamed and B. Far, "An enhanced threshold based technique for white blood cells nuclei automatic segmentation," e-health Networking, Applications and Services (Healthcom), 2012 IEEE 14th International Conference on, Beijing, 2012, pp [26] E. A. Mohammed, M. M. A. Mohamed, C. Naugler and B. H. Far, "Chronic lymphocytic leukemia cell segmentation from microscopic blood images using watershed algorithm and optimal thresholding," Electrical and Computer Engineering (CCECE), th Annual IEEE Canadian Conference on, Regina, SK, 2013, pp [27] Mohammed Emad A, Mohamed Mostafa M. A., Far Behrouz H, Naugler Christopher,"Peripheral blood smear image analysis: A comprehensive review", Journal of Pathology Informatics, 2014; 5(1). [28] E. A. Mohammed, B. H. Far, M. M. A. Mohamed and C. Naugler, "Application of Support Vector Machine and k-means clustering algorithms for robust chronic lymphocytic leukemia color cell segmentation," e-health Networking, Applications & Services (Healthcom), 2013 IEEE 15th International Conference on, Lisbon, 2013, pp [29] M. M. A. Mohamed and B. Far, "A Fast Technique for White Blood Cells Nuclei Automatic Segmentation Based on Gram-Schmidt Orthogonalization," Tools with Artificial Intelligence (ICTAI), 2012 IEEE 24th International Conference on, Athens, 2012, pp