Investigation for Improvement of Automatic Classification Accuracy of Leukocyte Image using Machine Learning

Abstract

Classifying leukocyte and examining their proportions is a very important for disease examination and diagnosis. This examination needs the knowledge of experts and a lot of time. Therefore, many automatic leukocyte images classification algorithms have been proposed. There is a method to classify 13 types blood cells using 1 vs 1 Support Vector Machine in one of them. In the conventional method, leukocyte images are classified with the 26-dimensional feature vectors. The feature vectors are composed of 18-dimensional feature vectors relating to the color of cell and 8-dimensional feature vectors relating to the form of cell. The classification accuracy, however, is poor with these feature vectors regarding granulocytes in this method. In this paper, we propose new feature vectors to improve the classification accuracy of promyelocytes, myelocytes, myelocytes, basophils, eosinophils with low classification accuracy among the leukocyte fractions. That is, we add two feature vectors in the proposed method.