Recognition of Blood and Bone Marrow Cells using Kernel-based Image Retrieval

Summary This paper presents a novel cell classification method based on image retrieval by learning with kernel. Cell image is firstly segmented into cytoplasm and nucleus regions in order to keep more spatial information. RGB color histogram of cell and two intensity histograms corresponding to those local regions compose feature vector represents the cell image. Kernel principal component analysis (KPCA) is utilized to extract effective features from the feature vector. The weight coefficients of features are estimated automatically using relevance feedback strategy by linear support vector machine (SVM). Classification depends on the decision distance obtained by SVM and the nearest center criterion. Experimental results on the ten-class task of 400 cells from blood and bone marrow smears show a 90.5% classification accuracy of the method when combined with standardized sample preparation and image acquisition.