Effectiveness of feature groups for automated pairwise leukocyte class discrimination.

Automated blood cell differentials using statistical classification techniques have been implemented in several commercial machines. The machine-derived features used to classify leukocytes resemble the descriptors used by humans performing visual classification, e.g. size, content, shape, color, and texture. However, because of our crude modeling of vision there is no universally accepted measure of characteristics such as shape, color, or texture. One expects, therefore, that features which are powerful discriminators for humans may perform poorly when quantified by machine, while other parameters may be more precisely measured automatically and so prove more useful in cell classification. This paper reports the results of a feature evaluation study for automated discrimination between all pairs of a large set of leukocyte classes consisting of both normal and abnormal types. To provide a framework for comparing automated feature ordering with the ranking attached by medical technologists, the machine-derived features were divided into six groups: size and content, mean and mode, cytoplasm/nucleus comparison, contrast and texture, color, and nuclear shape. A sequential procedure was used to select the best five-feature set from each group and the globally best five-feature set for each pairwise classification. The results illustrate the strengths and weaknesses of machine-derived descriptors for each type of decision.