EFFICIENT SEGMENTATION FRAMEWORK OF CELL IMAGES IN NOISE ENVIRONMENTS EunSang Bak 1 , Kayvan Najarian 2 , John P. Brockway 3 1 Electrical and Computer Engineering Department, 2 College of Information University of North Carolina at Charlotte 9201 University City Blvd, Charlotte, NC 28223, U.S.A. 3 Memory Testing Corporation P.O. Box 1621 Davidson, NC 28036, U.S.A. ABSTRACT In this paper, we propose an efficient segmentation method that exploits local information for automated cell segmenta- tion. This method introduces a new criterion function based on statistical structure of the objects in cell image. Each pixel is initially assigned to the most probable region and then the pixel assignment process is iteratively updated by a new criterion function until steady state is reached. We apply the proposed method to cervical cell images as well as the corresponding noisy images that are contaminated by Gaussian noise. The performance of the proposed method is evaluated based on the results from both normal and noisy cell images. Keywords: Cell segmentation, local information, itera- tive algorithm 1. INTRODUCTION The capabilities of the computer-aided cell segmentation and classification have been dramatically increasing for the last decade. Such automated systems allow the study of not only individual cell nuclei as local objects but also complex cell structures. Pap smear test is considered as one of the most routine cytological screenings. Screening is conducted by a trained pathologist based on a standard rule-base often referred to as “The Bethesda System” (TBS) [1]. Although many of the rules and criteria in the TBS have been clearly defined for each cell class and category, the evaluation of these rules may be performed rather differently by each pathologist. This is due to the fact that pathologists make decisions by visual inspection of a large number of cells and evaluating diagnostic features such as the ratio of the size of the nu- cleus to the size of cytoplasm or the shape of nucleus in each cell. Accurate measurement of such features from all cells is, while critical for reliable clinical determinations, a difficult and tiring process. An automated segmentation and classification system [2, 3] could be an alternative tool to avoid false clinical decisions due to fatigue or other types of human error. As mentioned above, the key step in any automated cell classification is cell segmentation as a preprocess. There are several methods for cell segmentation that empowers automated cell screening systems. Some of them [4, 5] in- volve a thresholding algorithm which determines the thresh- old value either adaptively or recursively. In other methods [6, 7], geometrical shape or configuration is utilized in some manners. In this paper, we propose an efficient segmentation scheme for cell segmentation and evaluate its performance using Pap smear samples in the presence of heavy additive noise. In the proposed method, in order to improve the segmen- tation performance, both global and local characteristics of the each object (nucleus, cytoplasm and background) of the cell image are exploited in the segmentation process. A locally adapted likelihood, called local spatial likelihood (LSL), is defined to reflect not only the global but also local characteristics. Combined with local spatial prior probabili- ties, the LSL forms a new function, called local spatial pos- terior (LSP), that is treated as a criterion function of the pro- posed segmentation method. The LSP iteratively updates a segmented image. We illustrate the performance of the local spatial posterior as a criterion function for cell segmentation in the presented experimental results. The paper is organized as follows: In Section 2, our pro- posed iterative segmentation process is described. The cri- terion function, LSP used in the iterative algorithm, is intro- duced in Section 3. The experimental results are presented in Section 4, and finally conclusions are given in Section 5. 2. ITERATIVE SEGMENTATION From the segmentation point of view, a cell image has three dominant regions that are occupied by nucleus, cytoplasm,