The Best Texture Features for Leukocytes Recognition

Differential counting of white blood cells (WBCs or leukocytes) is a common task to diagnose many diseases such as leukemia, and infections. An accurate process for recognizing leukocytes is to evaluate a blood smear under a microscope by an expert. Since, this procedure is manual, time-consuming and tedious, making the procedure automatic would overcome these problems. In an automated CAD (Computer-Aided-Design) system for this purpose, a crucial module is leukocytes recognition. In this paper, we are looking for the best features in order to recognize five types of leukocytes (Monocyte, Lymphocyte, Neutrophil, Eosinophil and Basophil) from microscopic images of blood smear in an automated cell counting system.

In this work, we focus on the texture features and seven categories: GLCM features, Haralick features, Spectral texture features, Wavelet-based features, Gabor-based features, CoALBP and RICLBP are analyzed to find the best features for leukocytes detection. The best features of each category are selected using stepwise regression and finally three well-known classifiers called K-NN, LDA and NB are utilized for classification.

The proposed system is tested on a self-provided dataset composed of 200 cell images. In our experiments, to evaluate the process, the accuracy of each leukocyte type and the mean accuracy are computed. RICLBP features achieved the best mean accuracy (85.53%) for LDA classifier.

In our experiments, although the maximum mean accuracy (85.53%) went with RICLBP features, but the accuracies of all five leukocyte types weren't maximized for RICLBP features. This result directs us to design and develop a system based on multiple features and multiple classifiers to maximize the accuracies even for each individual cell type in our future work.