The network exhibits fair performance on an external data set. Of 627 single-cell images from an external data set, the network predicted classes on 380 images (61% of the data set). The confusion matrix shows fair agreement between the annotations of Choi et al²⁰ and the network predictions, which possess slightly different, but compatible, classification schemes. Good agreement is generally observed with the exception of confusions within the myeloid and erythroid lineages.