Fig. 5: Automatic detection of developmental epochs.
From: Uncovering developmental time and tempo using deep learning
a, Strategy for the calculation of similarities between a test embryo and images from previous acquisition timepoints of the same embryo. Left, for each embryo image, the cosine similarity with all the images from previous acquisition timepoints is calculated. Given the symmetry of pairwise comparison, such similarity profiles can be stored in a symmetric squared matrix (right) that describes the self-similarity of an embryo during development over time. b, Representative similarity matrix showing pairwise comparison of an image series of one zebrafish embryo with itself. High similarity between neighboring timepoints defines clusters that correspond to developmental epochs. The Twin Network detects and partitions embryo development into phases that are in line with the classical zebrafish staging atlas6. Imaging was started at 2 hpf (64-cell stage). Full analysis shown in Extended Data Fig. 7a–c. Scale bar, 500 μm.
