Blood Cell Image Segmentation Based On Learning Strategy

The microscopic examination of blood cells is one of the basicmeans of diagnosis for leukemia and other diseases. By classifying, countingand comparing nucleated cells which come from bone marrow and peripheralblood, one can get the exact pathological diagnostic info rmation. While highprofessional experience may be required to the inspectors. Computer-aidedmorphological analysis of white blood cells by digital image processing andpattern recognition can greatly improve the efficiency and objectivity of thediagnosis. It is a hot topic with great application value. The imagesegmentation is the most important part in automatic image analysis system.Because the result of the segmentation determines the following partsperformances of the system. However, the differe nces of the color stain,smear preparation and imaging devices may cause large changes to cellimages especially in color. Traditional image segmentation algorithms aredifficult to deal with these complex scenes, thus they are limited in theapplication.In this paper, image segmentation could be regar ded as a two-classproblem. A two-class classifier is used to classify pixels of an image intoobject and background. The classification model of pixels is trained onlineusing supervised machine learning according to the changes of scene, thenwhich is used to classify pixels to extract leukocytes. The contributions of theresearch are as following:(1) Effective learning strategies are proposed in our method. Accordingto the prior knowledge of blood cell i mages, we firstly locate a few regions ofinteresting by mean-shift algorithm, and then sample pixels to group trainingsets. The influences of the parameters in proposed method are investigated,such as the window width of mean-shift, the gradient of pixels and theparameters of support vector machine (SVM) and so on. The high-gradientpixels are sampled as training data which can sharply reduce the number oftraining samples and improve the performance of the algorithm.(2) Machine learning methods with good generalization performance and learning ability using small training set are chosen to our method. Firstly,SVM is used to construct a segmentation model. The training of SVM needstune some parameters carefully, which is not an easy job. So a novelsingle-hidden layer feed forward networks(SLFNs)-extreme learning machine(ELM) which need not tune parameters is used to instead of SVM to product aclassification model. In order to improve the stability of the ELMs, theensemble ELMs is presented.(3) The fixational eye movement may reflect a mechanism how the brainmakes our environment visible. The localized modles are presented bysimulating eye movements in our research in order to extract single whiteblood cell from clustered cells. Experimental resu lts demonstrate thelearning-based method could reliably segment multiple-color objects fromcomplex scenes.