Ground truth
This animation shows ground truth segmentation of nuclei in a late blastocyst embryo. The segmentation was determined by trained biologists using manual annotation.
Click the image for a detailed view.
Robust 3D Nuclear Instance Segmentation of the Early Mouse Embryo
We introduce new ground truth datasets of 3D nuclear instance segmentation in pre-implantation mouse embryos as well as models trained on those datasets. One ground truth dataset (BlastoSPIM 1.0) is primarily composed of embryos from the 8-nuclei stage to the 64-nuclei stage. The other ground truth dataset (BlastoSPIM 2.0) is primarily composed of embryos from the 64-nuclei stage to the >100-nuclei stage. BlastoSPIM 1.0 and 2.0 are two of the largest and most complete datasets of their kind with 573 and 80 annotated light-sheet images, respectively. Each image has an xy-resolution of 0.208 micron and a z-resolution of 2.0 microns. For each of the 18336 nuclei across the two datasets, the ground-truth instance is defined by drawing a contour in each z-slice in which the nucleus appears.
Using this dataset, we have trained two deep-learning models to perform accurate 3D nuclear instance segmentation. With these models, we have begun several quantitative investigations into pre-implantation development, including the dynamics of fate specification, nuclear volume oscillations, and nuclear shape deformations.
Significant progress has been made in providing a means to segment nuclei in 2D [Cellpose 2.0] but there are very few biological datasets with complete ground truth instance segmentation of nuclei in 3D. This dataset can be used to help enable transfer learning for other applications via pre-training, fine-tuning, simulating, augmenting, evaluating etc.
Automatic segmentation
This animation shows automatically derived segmentation of nuclei in a late blastocyst embryo. This segmentation was derived using machine learning and computer vision methods.
Click the image for a detailed view.
Microscopy source data
This animation shows a time-lapse sequence of max-projected 3D image volumes of nuclei in an early mouse embryo. With a light-sheet microscope, a 3D image was acquired every 15 minutes, and the total movie duration is 29 hours and 45 minutes of development.
Click the image for a detailed view.
Model performance
Number of nuclei over time for 4 embryo sequences as inferred by the early embryo model (left) and late blastocyst model (right). Dashed horizontal lines: 16, 32, 64, 100 nuclei respectively. Note that the early model's estimates of nuclear count lose accuracy at around 64 nuclei, whereas the late model's estimates are able to accurately count nuclei up to the >100 nuclei stage.
Click the image for a detailed view.
Paper
"A novel ground truth dataset enables robust 3D nuclear instance segmentation in early mouse embryos"
https://www.biorxiv.org/content/10.1101/2023.03.14.532646v1
Hayden Nunley1, Binglun Shao1,2, Prateek Grover1, Jaspreet Singh1, Bradley Joyce3, Rebecca Kim-Yip3, Abraham Kohrman3, Aaron Watters1, Zsombor Gal,3, Alison Kickuth3, Madeleine Chalifoux2,3, Stanislav Shvartsman1,2,3,4, Eszter Posfai3, Lisa M. Brown1
1 Center for Computational Biology, Flatiron Institute - Simons Foundation, New York, United States of America
2 Department of Chemical and Biological Engineering, Princeton University, Princeton, New Jersey, United States of America
3 Department of Molecular Biology, Princeton University, Princeton, New Jersey, United States of America
4 The Lewis-Sigler Institute for Integrative Genomics, Princeton University, Princeton, New Jersey, United States of America