Fig. 3: Shape representations in synthetic and real cellular data characterise population variation for exploratory analysis.
a Synthetic dataset and its embedding space: (Top panel) Samples from our synthetic cellular shape dataset with varying eccentricity (columns) and contour randomness (rows), for 9 classes. (Second panel) Distance matrix between robust means of class centroids in embedding space; classes with different eccentricity are further than classes with different contour randomness. (Third panel) UMAP of embedding space coloured by eccentricity and (fourth panel) randomness; these show that eccentricity classes are more separated than contour randomness classes. b For real hiPSCs cell shapes without class labels23, we use model quality evaluation tests. (Left panel) Reconstructions: the original image next to its reconstruction, which should recover the important image features. (Middle panel) Orientation tests: an image in many orientations (left column) should reconstruct images in a canonical orientation (right column). (Right panel) k-nearest neighbours: the first column is sampled images; adjacent images are the `most similar' according to the model’s shape space. c Still using hiPSCs, we learn a shape space for cells (left) and nucleus (right) segmentation masks, and do GMM clustering with 8 clusters. (Left) `prototypes', reconstructions of the cluster centroid; (middle) samples; (right) cluster frequencies. This communicates shape variation in the data (as well as dimensionality reduction in Supplementary Fig. 12). Source data for (a) and (c) are provided as a Source Data file.
