Nuclear instance segmentation and tracking for preimplantation mouse embryos.

Area of Science:

• Developmental Biology
• Biotechnology
• Bioimaging

Background:

• Automated 3D nuclear segmentation and tracking are crucial for studying preimplantation embryo development.
• Challenges include low signal-to-noise, high density, and variable shapes, complicating accurate analysis.
• Supervised machine learning improves accuracy but requires extensive annotated 3D data.

Purpose of the Study:

• To develop a novel mouse line and dataset for improved 3D nuclear instance segmentation and lineage tracking.
• To benchmark convolutional neural networks for accuracy in this context.
• To establish a complete pipeline for nuclear tracking in early embryonic development.

Main Methods:

• Generated a new mouse line expressing near-infrared nuclear reporter H2B-miRFP720.
• Created the BlastoSPIM dataset of 3D embryo images with ground truth nuclear instances.
• Benchmarked seven convolutional neural networks, identifying Stardist-3D as the top performer.
• Developed a complete segmentation and lineage tracking pipeline using trained Stardist-3D models.

Main Results:

• Identified Stardist-3D as the most accurate instance segmentation method among seven tested.
• Successfully constructed a comprehensive pipeline for nuclear segmentation and lineage tracking from the eight-cell stage to the end of preimplantation development.
• Demonstrated the utility of the BlastoSPIM dataset for pre-training models on related imaging tasks and model systems.

Conclusions:

• The BlastoSPIM dataset and trained Stardist-3D models provide a robust solution for 3D nuclear segmentation and lineage tracking in preimplantation embryos.
• This resource facilitates high-throughput analysis of developmental processes.
• BlastoSPIM serves as valuable pre-training data for advancing bioimaging analysis.