WIENER-BASED DECONVOLUTION METHODS FOR IMPROVING THE ACCURACY OF SPOT SEGMENTATION IN MICROARRAY IMAGES A. Daskalakis * , C. Argyropoulos * , D. Glotsos * , S. Kostopoulos * , E. Athanasiadis * , D. Cavouras ** and G. Nikoforidis * * Medical Image Processing and Analysis Group, Laboratory of Medical Physics, University of Patras, 26500 Patras, Greece ** Medical Image and Signal Processing Laboratory, Department of Medical Instruments Technology, Technological Educational Institute of Athens. daskalakis@med.upatras.gr Purpose : Microarray experiments are important tools for high throughput gene quantification. Nevertheless, such experiments are confounded by a number of technical factors, which operate at the fabrication, target labelling, and hybridization stages, and result in spatially inhomogeneous noise. Unless these sources of error are addressed, they will propagate throughout the stages of the analysis, leading to inaccurate biological inferences. The aim of this study was to investigate whether image restoration techniques may improve the accuracy of subsequent microarray image analysis steps (i.e. segmentation and gene quantification). Materials and Methods : A public dataset of seven microarrays obtained from the MicroArray Genome Imaging & Clustering Tool (MAGIC) database were used. Each image contained 6400 spots investigating the diauxic shift of Saccharomyces cerevisiae. Restoration was based on the Wiener deconvolution. Subsequently, restored images were processed with the MAGIC tool for semi-automatic griding and segmentation. The influence of the restoration process on the accuracy of spot segmentation was quantitatively assessed by the information theoretic metric of the Kullback-Liebler divergence. Results : Pre-processing based on Wiener deconvolution increased the range of divergence (0.04 – 3.01 bits) and consequently improved the accuracy of subsequent spot segmentation. Conclusion : Information theoretic metrics confirmed the importance of image restoration as a pre- processing step that significantly improved the accuracy of subsequent segmentation, thus leading to more accurate gene quantification. Introduction Complementary DNA (cDNA) microarray imaging is considered as an important tool for large-scale gene sequence and gene expression analysis [1, 2]. Molecular biologists and bioinformaticians are using microarray technology not only for identifying a gene in a biological sequence but also for predicting the function of the identified gene within a larger system, such as the human organism [3]. The basic microarray experimental procedure involves hybridization of complementary nucleic acid molecules, one of which (target) has been immobilized in a solid substrate (e.g. glass) using a robotically controlled device (arrayer). The two main techniques for printing targets are metallic pin and inkjet based systems, which lead to the formation of circular spots of known diameter and cDNA target. Those spots are located at the vertices of a rectangular lattice on the solid substrate surface. Each one of them serves as a highly specific and sensitive detector of the corresponding gene [4]. In order to create a genome expression profile of a biological system with microarrays, the messenger RNA from a particular sample is isolated, labelled and hybridized on the microarray. Although labelling and detection of hybridized probes is performed using various protocols i.e. P 32 , chromogenic systems (i.e. digoxigenin, antidigoxigenin etc [5], fluorescent dyes (e.g. Cy3, Cy5) can be characterized as the most popular. After labelling and hybridization the microarrays are “read”, using methods contingent upon the nature of the labelling reaction i.e. PhosphoImager plates, confocal laser scanners, and Charged Couple Devices [6]. The data output of the microarray experiment are two 16-bit tagged image files, one for each fluorescent dye (Cy3, Cy5). By isolating the spots for each channel via image segmentation, and by analyzing the pixel intensities of each segmented spot, it is possible to accurately quantify gene expression. These three crucial steps, experiment, image processing and gene quantification characterize the microarray analysis pipeline. Gene quantification is, nevertheless, confounded by a number of technical factors, which operate at the fabrication, target labelling, and hybridization stages, and result, in the microarray output images, not only as spatially inhomogeneous noise but also as irregularities of spot shape, size, and position.[7, 8]. Additive degradation caused by the confocal laser scanner, used as “reading” method, is furthermore complicating gene quantification. Unless these sources of error and