High-Quality Synthetic Annotated Tissue Data Using Conditional Generative Adversarial Networks

We present a deep-learning based pipeline for generating high-quality synthetic instance-segmentation datasets of tissues undergoing early gastrulation in chick embryo. We create point-clouds using the Lennard-Jones potential and learn an image-to-image translation from these point-clouds to create synthetic tissue images and auxiliary flow tensors using a Generative Adversarial Network, from which the segmentation masks are derived. We evaluate the downstream utility of our synthetic data by training Cellpose and Stardist models from scratch under data replacement and data augmentation scenarios, show that the synthetic datasets effectively capture the statistical properties of the real dataset, and show that our synthetic data improves segmentation performance on a held-out test set. This approach substantially reduces expert annotation time, as late-stage gastrulation data are challenging to acquire and manually label, while similar synthetic examples can be flexibly generated from easily obtainable and annotated early-stage data. Code: https://gitlab.com/siddharthsrivastava/synthetic-tissue-data