142 1-4244-1525-X/07/$25.00 © 2007 IEEE High Throughput Algorithm for Leukemia Cell Population Statistics on a Hemocytometer Brinda Prasad Dept. of Electrical and Computer Engg University of Calgary Alberta, Canada Email: bprasad@ucalgary.ca Wael Badawy Dept. of Electrical and Computer Engg University of Calgary Alberta, Canada Email: badawy@ucalgary.ca Abstract— This paper presents a high throughput cell count and cluster classification algorithm to quantify population statis- tics of leukemia cell lines on a conventional hemocytometer. The algorithm has been designed, implemented and tested on test images that vary in image quality. The proposed algorithm uses a recursively segmented, median filtered and a boosted Prewitt gradient mask to generate a boundary box that encloses all the identified cells. Intensity profile plots acting as signature plots further assist in classifying a single isolated cell from a cell cluster. Processed results compared manually by a biological expert resulted in an accuracy of 95% for even low quality images with a computational time ranging between 8-12sec. Improved performance from the proposed algorithm could be observed when compared with other conventional image analysis tools. I. I NTRODUCTION Hemocytometer is a device used to count blood cells, corpuscles, organelles within cells and many other types of microscopic particles. This device resembles a conventional microscope slide with etched rectangular indentation that is carefully crafted to create a chamber area and depth of known dimensions. The etched chamber has a 9mm 2 area that is divided into nine squares each with 1mm 2 area. One of the square (1mm 2 area) represents a volume of 0.1mm 3 or 10 -4 ml. The square with area of 1mm 2 is further divided and sub-divided into eight squares each with area of 1/16mm 2 and 1/256mm 2 respectively. The etched chamber resembles a grid that assists in extracting the cell count and computing the cell concentration or cell density in a specific volume of fluid as in Eq. 1, C =  n v  (1) where, C = cell concentration in cells/ml; n = avg. number of cells/mm 2 area and v = volume counted = 10 -4 . Thus, c = n × 10 +4 The hemocytometer is often used in a laboratory setup to quantify leukemia cell count and thereby determine the viability of leukemia cell samples for experiments [1]. As seen in Eq. 1, it is crucial to obtain reliable leukemia cell count for accurate cell concentration and viability assessment for correct interpretation and diagnosis on a given cell sample [1]. The leukemia cell count is conventionally done manually by researchers using a standard hemocytometer [2]. The manual counting process using a standard hemocy- tometer tends to be error-prone, tedious, inaccurate and time- consuming. Additionally, to ease the manual counting process the cell sample is often diluted if the cells are too crowded or clustered. The presence of clusters in pathological samples with high leukemia concentrations, such as leukemia, is very common [3]. Hence, cell clusters can often add additional complexity to the manual counting process. There have been several algorithms in the past for cell detection that were either based on region-finding algorithms [4] [5] or based on contour- detection algorithms [6] [7]. The region finding algorithms are computationally intensive since they divide the grey-level histogram to compute thresholds for segmentations. Contour- detection based algorithms perform poorly in noisy images. Recently, a method for processing cell images included iden- tifying cells by morphological operations and sorting them based on size [8]. However, the assumption that the authors make based on uniform cell size might not always be true. This paper presents a high-throughput automatic leukemia cell count and cluster classification algorithm to automatically quantify the distribution of leukemia cells on a conventional hemocytometer grid. The proposed algorithm provides an automated cell count that includes complex clusters to yield a more consistent and representative data. The algorithm is tested on hemocytometer slides using daudi leukemia cell lines that largely vary in image quality as shown in zoomed in portion of Fig. 1 (a)-(c). The area of interest for processing these images are indicated with a black boundary as seen in Fig. 1. It can be observed that daudi cells in the test images appear largely elliptical and translucent. Processing and selectively extracting daudi cells from other dust particles and noise can be challenging owing to the variability in the cell sample appearance as seen in the test images. II. PROPOSED ALGORITHM The proposed algorithm aims at identifying and classifying each cell as either a single cell or a cell cluster to obtain the population statistics that includes the cell count, cell size and cell position. The different steps used in the proposed algorithm are elaborated in the subsequent sub sections.