Network prediction analysis shows focus on relevant image regions. Original images classified correctly by the network are shown in the top row. As detailed in the main text, all cells were stained using the May-Grünwald-Giemsa/Pappenheim stain, and imaged at ×40 magnification. The middle row shows analysis using the SmoothGrad algorithm. The lighter a pixel appears, the more it contributes to the classification decision made by the network. Results of a second network analysis method, the Grad-CAM algorithm, are shown in the bottom row as a heat map overlaid on the input image. Image regions containing relevant features are colored red. Both analysis methods suggest that the network has learned to focus on the leukocyte while ignoring background structure. Note the attention of the network to features known to be relevant for particular classes, such as the cytoplasmic structure in eosinophils or the nuclear configuration in plasma cells.