In this study, post-implantation human epiblast and amnion development are modeled using a stem cell-based embryoid system. A dataset of 3,697 fluorescent images, along with tissue, cavity, and cell masks, is generated from experimental data. A computational pipeline analyzes morphological and marker expression features, revealing key developmental processes such as tissue growth, cavity expansion, and cell differentiation. To uncover hidden developmental dynamics, a novel deep manifold learning framework is introduced. This framework uses an autoencoder to project embryoid images into a 20-dimensional latent space and models the dynamics using a mean-reverting stochastic process of mixed Gaussians. The approach accurately captures phenotypic changes observed at discrete experimental time points. Moreover, it enables the generation of artificial yet realistic embryoid images at finer temporal resolutions, providing deeper insights into the progression of early human development.

(1) Acquisition of fluorescent images 

The experimental protocol for the generation of embryoids is provided in the work: Kejie Chen, Yi Zheng, Xufeng Xue, Yue Liu, Agnes M. Resto Irizarry, Huaijing Tang, and Jianping Fu(*). Branching development of early post-implantation human embryonic-like tissues in 3D stem cell culture. Biomaterials, 275:120898, 2021.

Specifically, the isolated human embryoid stem cells (hESCs) were aggregated in the PDMS pyramidal wells and cultured for 6 hours using mTeSR1 medium (STEMCELL Technologies) supplemented with Y2763 (10 μM). hESC aggregates were then transferred into the 128-well plate for suspension culture in 5% Geltrex. A proportion of embryoids were collected from the well plates and stained with fluorescent markers including DAPI, NANOG, GATA3 and T. Four-channel fluorescent images were acquired using an Olympus IX81 fluorescent microscope at ×20 magnifaction. 

(2) Image pre-processing

First, a proportion of embryoid images was selected. Boundaries of the embryoid tissues, cavities, and single cells in these images (DAPI channel) were manually annotated using an open-source annotation software. The annotation results were automatically read using the custom Python scripts. Next, a Transformer-CNN model was developed for accurate segmentation of embryoid tissues, cavities, and single cells. Annotated images were used to train and validate the Transformer-CNN model. Then, the model was applied to all the images. The predicted masks were verified by annotators visualizing the raw images and the predicted masks image-by-image. Finally, the centers of embryoid tissues were moved to the center of the images for both raw images and masks. The longest axes of the embryoids were aligned vertically.