AUTOMATED CLASSIFICATION OF NORMAL AND ABNORMAL LEUKOCYTES

Abstract

The development of an automated system for counting and classifying normal and abnormal leukocytes in peripheral blood smears is described. General requirements are discussed and the results of a simulation experiment are presented. A sample of 1572 leukocytes, divided equally among 17 types, was photographed and analyzed using computerized pattern recognition techniques. Various geometrical, color and texture parameters were extracted from the cell images and an optimal set of 20 were used in several computerized classification runs. Training on one-half of the sample and classifying the other half resulted in an over-all correct classification of between 67 and 77% depending on the definition of classification error. When only normal cells are considered, correct classification is obtained for 9l.5% of the cells.