A texture approach to leukocyte recognition

Millions of white blood cells are manually classified in laboratories using microscopes, a painstaking and subjective task. A trained medical technician takes about 15 min to evaluate and count 100 cells for each blood slide, a time consuming and susceptible to error procedure. Leukocyte shape is usually insufficient to differentiate even among normal types since it varies widely. The current paper addresses the pattern recognition problem of blood image analysis and how textural information can improve differentiation among leukocytes. Cooccurrence probabilities can be used as a measure of gray scale image texture, a statistical method for characterizing the spatial organization of the gray-tones. We calculate five textural attributes based on gray level cooccurrence matrices (GLCM) as energy, entropy, inertia and local homogeneity, testing these features in leukocyte recognition. Several parameters must be estimated for obtaining GLCM, therefore we implement datamining algorithms for estimating suitable scales. Feature selection methods are also applied to define the most discriminative attributes for describing the cellular patterns. Experimental results show that texture parameters are essential to differentiate among the five types of normal leukocytes and chronic lymphocytic leukemia, evidencing the importance of biological aspects regarded by hematologists as nuclear chromatin and cytoplasmical granularity.