Automatic classification of red cells in peripheral blood smears

Abstract

Introduction: The process of erythrocyte classification in peripheral blood smear is normally done manually from microscopic observation. This implies not only a considerable investment of time and resources but also brings potential problems of subjectivity and difficulty in the reproducibility of diagnosis. Objective: To develop an application that allows the automatic classification of red blood cells in peripheral blood smears, as a diagnostic aid tool. Methodology: Image processing techniques were used in order to segment erythrocytes in the microscopic photographs and to measure characteristics as area, perimeter, solidity, circularity, eccentricity, texture and box-counting dimension. An artificial neural network was used to classify the red blood cells in the images in seven classes, including normal and six pathological changes, according to their characteristics. The network was trained according to the classification of 262 erythrocytes by an expert hematologist. The developments were made in matlab ®, a powerful scientific computing platform. Results: The chosen network reaches 97.3% correct in the validation data. Mistakes in the network correspond to cells with various pathological classifications features, which make them difficult to classify even for an expert. Conclusions: The developed application classifies quickly and accurately the different types of red blood cells in a microscopic sample of peripheral blood smear, so it could be useful as a diagnostic support tool.