Classification of peripheral blood neutrophils using deep learning

Deep learning has been used to classify the while blood cells in peripheral blood smears. However, the classification of developing neutrophils is rarely studied. Moreover, it is still unknown whether deep learning can work well on the data coming from different sources. In this study, we therefore investigate the classification performance of deep learning for immature and mature neutrophils. In particular, we used three open-access datasets obtained from different imaging systems: CellaVision DM 96, CellaVision DM 100, and iCELL ME-150. A total of 26,050 images identified by one laboratory technologist were randomly split into training, validation, and testing datasets. A total of 10 convolutional neural networks were trained to classify six blood cell types: myeloblast, promyelocyte, myelocyte, metamyelocyte, banded neutrophil, and segmented neutrophil. The experimental results showed that compared to any single model, the average ensemble model could achieve a better classification performance and provide a testing accuracy of 90.1%. The sensitivity and specificity of the average ensemble model for the six blood cell types were above 83.5% and 96.9%, respectively. Our results suggest that deep learning is a promising tool for the classification of developing neutrophils, but further improvement is required.