Wall-adherent cells segmentation based on SVM

Abstract

Anti-virus experiment in vitro is a common way to screen and identify antiviral drugs. Most of the cell lines in anti-virus experiments are wall-adherent. The segmentation, recognition and counting this walladherent cells in micrograph successfully can cut down the time and cost of the experiment. Because of the wall-adherent cells characteristics, to segment this kind of image is much difficult. In this paper some research works have been done to segment the walladherent cells image by support vector machine (SVM). The number of training samples (NOTS), kernel function (KF) and features are three key facts of SVM and many experiments have been done to discuss the affection to relative error (RE) from them. Analyzing and comparing the experiments data, the SVM is determined finally with NOTS=7200, perceptron kernel function ( β = - 1 ), features of group 2. After training the SVM by samples, another two images are inputted into the SVM and the segmentation results are outputted. In the segmentation results images, the cells' edges are well connective, the noise is not much and can be removed easily. Therefore, the segmentation performance of SVM is good in this problem of wall-adherent cells image segmentation. © 2009 IEEE.