Abstract—Two-Dimensional Polyacrylamide Gel Electro- phoresis (2D PAGE) is a proteomic technique that allows the analysis of large collections and complex mixtures of proteins. The 2D-PAGE gel images depict protein signals as spots of various intensities and sizes. In this paper, we present a novel approach to unsupervised protein spot detection in 2D-PAGE images based on a genetic algorithm. This approach involves three main steps: a) wavelet-based noise reduction, b) segmentation of the input images into regions around the local maxima of the image intensities, c) detection and model-based quantification of the spots within each region using a genetic algorithm. This algorithm searches within a multidimensional parameter space to determine, in parallel, the parameters of multiple diffusion models that optimally fit the characteristics of possible spots. The detection and quantification of the spots is achieved by superposition of diffusion functions modeling adjacent spots. Experiments with 16-bit 2D-PAGE images show that the proposed method is effective and results in low spurious spot detection rate. I. INTRODUCTION ROTEOMIC research deals with the systematic analysis of protein profiles expressed in a given cell, tissue or biological system at a given time. In this field, two- Dimensional Polyacrylamide Gel Electrophoresis (2D- PAGE) analysis [1], is a well-established and widely used technique for the analysis of large collections and complex mixtures of proteins. Images produced by digitization of 2D-PAGE gels contain spots, of various intensities and sizes, that correspond to proteins. Detection and quantification of the protein spots may reveal alterations in protein expression within a given biological system. However, this is not a straightforward process. It can be rather complicated due to the presence of noise, the inhomogeneneous background, and the overlap between the spots. A variety of software packages have been developed for protein spot detection [2]. Many of these packages implement image segmentation methods based on edge detection algorithms such as Laplacian filtering, in conjunction with smoothing or morphological operators [2]- Manuscript received June 30, 2006. This work was supported by the Greek General Secretariat of Research and Technology and the European Social Fund, through the PENED 2003 program (grant no, 03ED332). D. K. Iakovidis, D. Maroulis, and E. Zacharia, are with the Department of Informatics and Telecommunications, University of Athens, GR 15784 Panepistimiopolis, Ilisia, Greece (e-mail: rtsimage@di.uoa.gr). S. Kossida, is with the Division of Biotechnology, Center of Basic Research, Foundation of Biomedical Research of the Academy of Athens, Greece (e-mail: skossida@bioacademy.gr). [4]. However, if a 2D-PAGE image contains artifacts it is likely that the boundaries of the artifacts have similar characteristics with the boundaries of the actual spots, leading to spurious spot detection. Moreover, the segmentation produced by the edge detection methods is particularly dependent on the preparation of the 2D-PAGE gels. The watershed algorithm has also been a popular choice for 2D-PAGE image segmentation [4]. It usually performs better than the methods based on edge detection algorithms, however it tends to oversegment the images. To alleviate possible oversegmentation effects, post-processing techniques, such as region merging, are usually applied. State of the art approaches to spot detection by image segmentation include geometric algorithms [5], and the pixel value collection method [6]. 2D-PAGE image segmentation is usually followed by characterization and representation of the protein spots with a list of parameters over which further analysis can be carried out. Spot characterization algorithms span two categories: parametric and nonparametric. Nonparametric methods [6]-[7] involve heuristic post-processing of the segmentation boundaries for the delineation of spots, which are then represented by a set of measurements calculated over the detected spot regions. These methods do not impose any explicit constraint on the shape of the boundaries or the appearance of the spots. However, they exhibit poor performance with complex images. Parametric methods utilize model functions to parameterize protein spots. Models represent prior knowledge used to impose constraints on the analysis procedure. This in turn improves the robustness of the solution. Early approaches to modeling protein spots in 2D- PAGE images include the use of 2D-Gaussian functions [8][9]. This model provides a good representation of some spots, but has proved inadequate as a general model. More precisely, in [10] it is noted that when the local concentration of protein is high, saturation effects occur and the spot can not be accurately modeled by a Gaussian function. Instead, a simplified diffusion model is suggested as more appropriate. Optimization of a model’s parameters usually involves supervised techniques. For example, Melanie, a popular software package for 2D-PAGE analysis, uses the Polak- Ribiere variant of the conjugate gradient method to optimize the parameters of a Gaussian model [10]. Latest advances in spot modeling include the construction of spot models based A Genetic Approach to Spot Detection in Two-Dimensional Gel Electrophoresis Images Dimitris K. Iakovidis, Member, IEEE, Dimitris Maroulis, Member, IEEE, Eleni Zacharia, and Sofia Kossida P