Figure 6.

UMAP embedding of extracted features. UMAP embedding of the test set using the algorithm of McInnes et al.⁴³ The flattened final convolutional layer of the ResNeXt-50 network containing 2048 features is embedded into 2 dimensions. Each point represents an individual single-cell patch and is colored according to its ground-truth annotation. All annotated classes are separated well in feature space. Cell types belonging to consecutive development steps tend to be mapped to neighboring positions, reflecting the continuous transition between the corresponding classes.