Segmentation of Blood and Bone Marrow Cell Images via Learning by Sampling

Abstract

This paper presents an automatic machine-learning method to segment blood and bone marrow cell images. Different from traditional methods, we focus on a few significant samples rather than all of them. Firstly, three mean-shift procedures are used to seek the local clustering modes corresponding to the regions of nuclei, mature erythrocytes and background respectively. And then a SVM is trained by uniform sampling from three modes in order to find more nuclei pixels. So we could dilate the nuclei regions only in high gradient pixels to get the part pixels of cytoplasm. Finally, we train a new SVM by a training set sampling from cytoplasm and three modes to extract the whole leukocytes. SVM with fixed parameters is used here to yield two classification models via learning by sampling on-line. The segmentation results of the new method are closer to the human visual perception. It can achieve higher accuracy of segmentation in complex scenes and more robust to color confusion and changes. Experiments have demonstrated the validity of the new method compared with the thresholding and the watershed algorithm.