Both CNN models outperform the feature-based classifier in terms of tolerant class-wise recall. (A) Some morphological classes can be difficult to distinguish, so that a misclassification can be considered tolerable. Strict classification evaluation accepts only precise agreement of ground truth and network prediction, as shown in the red diagonal entries of the matrix. Mix-ups that are considered tolerable are colored blue. (B) Tolerance improvement for key classes. Error bars indicate standard deviation across 5 cross-validation folds. For segmented neutrophils and lymphocytes, performance of the feature-based classifier is slightly higher than of the neural networks. In all other classes, both CNNs consistently outperform the feature-based classifier of Krappe et al.¹³ This might be due to the distinctive signal of the nuclear shape of segmented neutrophils and lymphocytes in feature space. Additionally, ResNeXt outperforms the sequential network in several key classes, reflecting the greater complexity of the network used.