In this paper we present a two-stage method to segment white blood cell imags by a pixel classification model that is trained online using an extreme learning machine (ELM). During the training stage

we first locate leukocyte nucleus by mean-shift algorithm in the RGB color space. Then we dilate the leukocyte nucleus until the maximum ratio of entropy and area of the nucleus region occurs. The region including the nucleus could be regarded as positive candidate region for sampling. While the other regions excluding the positive one

are regarded as negative candidate regions. A two-class ELM could be trained with the training set via learning by sampling. Different training sets produce multiple models of ELM. In the test stage

multiple models of the ELM can be integrated to classify pixels in order to extract leukocytes. The proposed algorithm does not need to change any parameter during run-time. It is very robust to various staining and to the illumination in cell imaging. Experimental results demonstrate the effectiveness of the method.