Fig. 5: Modelling greyscale images enables joint representations of shape and texture.
a Synthetic dataset and its embedding space: (Top) Samples from our synthetic cellular shape and texture dataset with varying eccentricity (columns) and Perlin texture (rows), for 9 classes (no scale bar because data is synthetic). (Second panel) Distance matrix between robust means of class centroids in embedding space; classes with different eccentricity are more separated than classes with different texture. (Third panel) UMAP of embedding space coloured by eccentricity and (fourth panel) texture; these show that eccentricity classes are more separated than texture classes. b Linear probing scores measure representation quality by simulating a classification task on the embedding space (top scores in bold). For three experiments modelling texture or texture and shape jointly, O2-VAE has better texture representation scores (see Methods). c Sample data of real nuclei with mitosis phase classes after min-max scaling25. Scale bar in bottom-right image, 10 μm. d UMAP of representations have good separation of classes, with some mixing of interphase and prophase cells. e Representation quality (linear probing) scores for all mitosis classes and the three challenging classes (interphase, prophase, prometaphase)(top scores in bold). O2-VAE representations perform better than VAE baselines. O2-VAE representations are better when modelling texture and shape jointly (using greyscale images) compared with shape only (using segmentation images). Source data for (a) are provided as a Source Data file.
