The University of Washington Department of Electrical Engineering

The Microscale Life Sciences Center desires an automated count of fluorescing live/dead biological cells from images obtained via microscopy. Various cell counting procedures were explored in the literature. Four methods were tried: threshold segmentation, watershed segmentation, crosscorrelation, and a threshold “real-time” cross-correlation hybrid. The threshold technique produced the best results in overall accuracy, closely followed by the hybrid and then the cross-correlation, which performed the worst. The watershed technique could not be used with the images to produce a result substantially different or better than threshold segmentation due to image intensity saturation in the sample set. INTRODUCTION The Microscale Life Sciences Center, under the direction of Dean Meldrum and Vice Provost Lidstrom, is developing a system for biologists to study single cells through the use of microscopy and sensors. One such sensor tracks oxygen consumption by cells in a 3-by-3 array of tiny microwells etched in the surface of a glass slide. The purpose is to track the metabolic rates of cells in response to a variety of stimuli. During operation, the oxygen supplies are strictly controlled and images of the cells are captured by a CCD camera via a microscope. To make an accurate estimate of per-cell O2 consumption, the micro-well consumption is divided by the number of living cells in the micro-well. The number of dead cells in each micro-well is also relevant since dead cells disaggregate, contaminating the well and influence the metabolic rates of surrounding cells. This necessitates an accurate method for counting the number of living/dead cells in each micro-well. Motivation While it is possible for a human to do the actual counting, the shear amount of images makes an automated approach favorable in not only efficiency, but also uniformity. An automated count of live/dead biological cells would therefore reduce the workload of potential investigators while yielding significant amounts of accurate data. Such an algorithm would be an invaluable asset for the Center. Problem Formation To distinguish between the living and dead cells in the images, two fluorescing dyes are added to the media with non-overlapping emission spectra. The live stain is a cell membrane permeable, ionbinding dye, Calcein AM, while the dead stain is nucleic acid binding, non-permeable SYOTX orange. Since the cell membrane disaggregates as the cell dies the amount of the dead dye that can reach the DNA becomes live/dead dependant. During the experiments two images of the micro-well array are taken, each with a passband corresponding to one of the two dye emission spectra. The two gray scale images are then combined into a single color image with the live stain and dead stain mapped to the green and red channels respectively. Figure 1 shows an example live/dead composite image. In addition, the fluorescence spectra of an oxygen sensitive polymer that rings the inside of each micro-well overlaps that of the dead cell stain causing it to appear on the red channel as well. While this addition lends itself to determining the well locations it further complicates automating a dead cell count. Figure 1: Micro-array with live (green) and dead (red) macrophage cells. The red circles are the boundaries of the wells and glow red because dyes bind to the corners. Only cells within the wells are of interest. Cells are able to move and reproduce, asexually, within the confines of the wells. A key problem to address is when to count one green spot as one cell or two. Furthermore, there is the added problem of non-cellular debris and/or noise, which is imaged as a uniform red dusting over the entire field of view. The counting of dead cells becomes considerably more difficult with these considerations. Additionally, cells may partially or completely eclipse each other. In the case of a complete overlap, nothing may be done but the algorithm should be robust enough to recognize partially overlapped live cells and/or dead ones. Counting the cells can be formulated as a classification/detection problem. In this case the three classes are live, dead, and noise. A fourth possible classification is dying. To determine the effectiveness of the automation, its results will be compared to human segmentation and classification of the same data sets. BACKGROUND AND PRIOR WORK Methods for computer vision based counting can generally be broken-down into four steps: preprocessing, segmentation, feature extraction, and classification. During pre-processing, filtering techniques are used to remove noise. During segmentation, the image is subdivided into points or regions of interest. During feature extraction, descriptive values are extracted from each point or region. Finally, during classification, each point or region is assigned a class based on its extracted features. For the most basic case consider two classes ‘cell’ and ‘background.’ A block diagram of the steps to classify between the two can be seen below in Figure 2. Since computer vision based counting is a well studied problem, a significant number of methods have been published which can be utilized in the creation of an automated cell counting system. In addition to publications on automated cell counting, related works such as automated cell segmentation and cell colony counting [3] offer a wealth of knowledge that can be drawn on. Cell segmentation, in particular, offers a powerful method for feature extraction which forms a solid foundation for classification. Segmentation Feature Extraction Classification Captured Image Contiguous regions Means Circularity Radial Variance {Cell, Noise} Labeled Dataset Figure 2: Computer vision based counting/classification can be subdivided into three blocks. First, the image is segmented with the output being ether a set of contiguous regions or a single pixel subset to be operated on. Next, each segment or pixel is transformed into a set of features based on its properties. Finally, the set of features for each pixel or segment is evaluated and a class is assigned. Cell Segmentation The cell segmentation methods in the literature can be broken up into three rough categories: thresholding, watershed, and active contour. An empirical comparison of three main cell segmentation methods was performed by Bamford [2]. The comparison showed that the active contour segmentation method ranked higher then the watershed method, which ranked higher then the thresholding method when evaluated against manual segmentation. In contrast, the thresholding method required the least computation time. Threshold Segmentation Threshold segmentation transforms an intensity image into a binary image and groups pixels together into contiguous regions [2][1][3]. This method is the gateway to a wide array of powerful binary image processing techniques such as erosion, dilation, opening, and closing. The primary advantages of threshold segmentation are its simplicity and efficiency. The methods of threshold segmentation are shown in Figure 3. Image Threshold Filtering Cluster Assignment Cluster Information {S1,..., Sn} Where Si is defined as: IDi Membersi Fi Identity of Cluster (1, 2, 3, et cetera) Pixels Belonging to Cluster Features of Cluster (area, radius , perimeter , ...) h (Threshold value )