A new self-partition clustering method for robust identification of subsets with heterogeneous size and density and its clinical application to leukocyte differential counting

Abstract

BackgroundIdentification of clusters in 2-dimensional scatterplots generated by hematology analyzer is a classical challenge. Conventional clustering algorithms fail to process cases with complicated mixtures of overlapping clusters and noise. MethodA new method was developed that features an image processing algorithm for rational identification of initial clusters and a self-partition clustering (SPC) algorithm with iterative truncation-correction (ITC) method to handle overlapping and noise. All clusters are assumed to follow bivariate Gaussian distributions with specified means, SDs, and correlation coefficient. While, each data point is assumed to belong to all clusters but with different proportions according to the likelihood of belonging to each cluster (computed by the Mahalanobis distance) and the data size of the cluster. Bivariate cluster statistics are computed in consideration of a weight factor determined cluster by cluster by each data point. In the computation, the ITC method minimizes the effect of overlapping and data. ResultsPerformance of SPC/ITC method was evaluated by its application to differential leukocyte counting. It showed comparable performance with manual counting and much better performance than the commonly used expectation maximum algorithm. ConclusionThe SPC/ITC method showed superior performance in situations with overlapping and low-density clusters such as leukopenia or leukocytosis.