Summary
For investigations into fate specification and cell rearrangements in time-lapse images of preimplantation embryos, automated 3D instance segmentation of nuclei and subsequent nuclear tracking are invaluable. Often, the images’ low signal-to-noise ratio and high voxel anisotropy and the nuclei’s dense packing and variable shapes limit the performance of many segmentation methods, while subsequent tracking of nuclear instances is complicated by low frame rates and sample movements. Supervised machine learning approaches can radically improve segmentation accuracy and enable easier tracking, but they often require large amounts of difficult-to-obtain annotated 3D data. Here we report a novel mouse line expressing near-infrared nuclear reporter H2B-miRFP720. H2B-miRFP720 is the longest wavelength nuclear reporter in mice and can be imaged simultaneously with other reporters with minimal overlap. We then generate a dataset, which we call BlastoSPIM, of 3D microscopy images of H2B-miRFP720-expressing embryos with ground truth for nuclear instance segmentation. Using BlastoSPIM, we benchmark the performance of seven convolutional neural networks and identify Stardist-3D as the most accurate instance segmentation method across preimplantation development. We then construct a complete pipeline for nuclear instance segmentation with our BlastoSPIM-trained Stardist-3D models and lineage tracking from the 8-cell stage to the end of preimplantation development (> 100 nuclei). Finally, we demonstrate BlastoSPIM’s usefulness as pre-train data for related problems, both for a different imaging modality and for different model systems. BlastoSPIM, its corresponding Stardist-3D models, and documentation of the full associated analysis pipeline are available at: blastospim.flatironinstitute.org.
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
↵* lbrown{at}flatironinstitute.org, eposfai{at}princeton.edu
This revision includes: (1) benchmarking of two additional methods on our ground-truth dataset (see Fig 2); (2) details the performance of models trained for nuclear segmentation in early embryos and late blastocysts (see Fig 3); (3) explains how the nuclear segmentations are tracked across time to generate lineage trees (see Fig 4); (4) shows an example of complete lineage trees from the 8-cell stage to the approximately 100-cell stage (see Fig 5); (5) includes additional tests of generalizability of our trained models (see Fig 6)
