International Journal of Computer Applications (0975 – 8887) Volume 133 – No.10, January 2016 1 Automatic Segmentation of Acute Leukemia Cells A.H. Kandil Associate Professor Systems and Biomedical Engineering Department Faculty of Engineering, Cairo University. O. A. Hassan Systems and Biomedical Engineering Department. High Institutes of Engineering El-Shorouk city ABSTRACT The recognition of the acute Leukemia blast cells in colored microscopic images is a challenging task. Segmentation is the essential step for image analysis and image processing. In this paper, an algorithm is presented that consists of panel selection followed by segmentation using K-means clustering then a refinement process. This algorithm was applied on public dataset designed for testing segmentation techniques. The results were compared with two different segmentation techniques developed by other researchers on the same data set. Our algorithm results in a sensitivity of 97.4 % and specificity of 98.1%. The developed algorithm was tested to another dataset of samples extracted from patients in local hospitals. The algorithm results in sensitivity of 100%, Specificity of 99.747% and accuracy of 99.7617%. The results were approved by expert pathologists. General Terms Pattern Recognition, Image processing and segmentation. Keywords Leukemia, segmentation, image enhancement, K-means, and watershed method. 1. INTRODUCTION When the bone marrow generates abnormal white blood cells, they are known as the cancerous blood cells [1] which may be lymphocytic or myelogenous. The cancerous blood cells can be either acute or chronic. In the acute case, the patient gets worse very fast while in the chronic case, the patient gets worse slowly [2]. Leukimea can be classified as follows, Chronic Lymphocytic (CLL), Acute Lymphoblastic (ALL), Chronic Myelogenous (CML), and Acute Myelogenous (AML) [2, 3]. The French-American-British (FAB) classification [2, 4] And the WHO proposal [3, 5] are the two main medical classifiers for Leukimea. It was reported in FAB that the Blast cells in the peripheral blood smear classifies the morphology of Acute Myeloid (AMLs), in seven types (M-1 - M-7) and Acute Lymphoblastic (ALL) in ALL-L1, ALL-L2 and ALL-L3 [2] . Generally, microscopic investigation of the blood cells is performed manually by hematologists using with a light microscope. It is very tedious, time consuming. In case of analyzing a large number of cells, the visual identification is hard to achieve in reasonable time. Several algorithms and techniques have been developed for the blood cells recognition. Image processing is considered as an important tool for successful automatic diagnosis for both AML and ALL cases. In this work, bone marrow images are presented with heterogeneous staining and pixels features such as color and texture. The remainder of this paper is organized as follows. Section 2 includes the research background. In section 3, the proposed segmentation algorithm is described. In section 4, the results of the algorithm are presented and discussed. Finally, the conclusions are drawn in section 5. 2. BACKGROUND Segmentation is a basic step in the image analysis, its goal is to partition an image into a set of meaningful patterns easier to analyze. Similarity is a key step towards organizing the image pixels into regions that would correspond to semantically meaningful entities in the scene. There are two forms of segmentation, pixel based image segmentation and region based segmentation. According to Sabino DMU et al. [6], segmentation enables features to be extracted without the inclusion of extraneous material by defining the boundaries of the blood cells. Clustering has been widely used in segmentation of grey level images. Chen Q et al. [7] compared the performances of watershed segmentation for binary images with different distance transforms including Euclidean, City block and Chessboard. They pointed out that combining the watershed and Chessboard distance transform resulted in successful segmentation. Mohamed MMA et al. [8] studied the normal white blood cell nucleus segmentation algorithm. The proposed algorithm is based on Gram-Schmidt orthogonalization technique. Such a technique enhances the arbitrary selected color of RGB and diminishes the two other colors. They demonstrated the highest contrast of the nucleus. Morphological operations were used to enhance the segmentation. They presented an efficient technique for automatic normal blood cell nuclei segmentation. Assessment of the proposed technique on the blood image set gives 85.4% accuracy. Eosinophil was found to have the highest segmentation accuracy with 90.1%. Lymphocyte and Basophil have the lowest accuracy with 78.3% and 78.6% respectively. Their study was based on the image dataset of 367 color image used with the cell type’s distributions. Trivedi MM, Bezdek JC [9] showed that the color based clustering depends on the selection of color space. The used Fuzzy c-Means clustering algorithm that employs a Pyramid Data Structure (PDS) of Aerial images. This structure allows one to represent the original image at various levels with different resolution. According to Kim K. et al. [10] segmentation techniques as region growing, edge detection, threshold based and pixels clustering are considered as the procedure basis. Blood images are captured from colored CCD camera attached to the microscope. Kim K et al. rely on visual assessment to accomplish the segmentation step. Aimi, Salihah et al., pointed that based on the HSI (Hue, Saturation, and Intensity) colour space image. Colour segmentation based on S component image for both the Blast and nucleus of AML images. Then, they extracted the S component information from the enhanced RGB image. Develop the S component histogram (S-plot) from S component image to obtain