Title:White blood cell classification

Abstract:This paper proposes a novel automatic classification framework for the recognition of five types of white blood cells. Segmenting complete white blood cells from blood smears images and extracting advantageous features from them remain challenging tasks in the classification of white blood cells. Therefore, we present an adaptive threshold segmentation method to deal with blood smears images with non-uniform color and uneven illumination, which is designed based on color space information and threshold segmentation. Subsequently, after successfully separating the white blood cell from the blood smear image, a large number of nonlinear features including geometrical, color and texture features are extracted. Nevertheless, redundant features can affect the classification speed and efficiency, and in view of that, a feature selection algorithm based on classification and regression trees (CART) is designed. Through in-depth analysis of the nonlinear relationship between features, the irrelevant and redundant features are successfully removed from the initial nonlinear features. Afterwards, the selected prominent features are fed into particle swarm optimization support vector machine (PSO-SVM) classifier to recognize the types of the white blood cells. Finally, to evaluate the performance of the proposed white blood cell classification methodology, we build a white blood cell data set containing 500 blood smear images for experiments. By comparing with the ground truth obtained manually, the proposed segmentation method achieves an average of 95.98% and 97.57% dice similarity for segmented nucleus and cell regions respectively. Furthermore, the proposed methodology achieves 99.76% classification accuracy, which well demonstrates its effectiveness.