Extended Data Fig. 1: Architecture of the Twin Network to analyze developmental dynamics.

(a) High-throughput imaging pipeline and ResNet101-based image segmentation to generate developmental trajectories of individual embryos. Embryos are individually tracked, as indicated by equally colored bounding boxes in the segmentation steps. (b) Model architecture of the core of the Twin Network based on ResNet50. (c) Image triplets consist of an anchor image, a positive image, and a negative image, and are passed to Twin Network for training with triplet loss. Anchor and positive images contain similar objects, while anchor and negative images show dissimilar objects. Triplet loss is used during the training to reduce the Euclidian distance between embeddings generated for the anchor image and positive image, and increase the distance between embeddings of the anchor image and negative image. Embryos for illustration also shown in Fig. 1a. Scale bar, 500 μm.