Automated precise recognition of ovine embryo microscopic images using a boundary-enhanced deep learning network.

Area of Science:

• Animal Science
• Biotechnology
• Computer Vision

Background:

• Efficient selection of in vitro-fertilized sheep embryos is vital for successful assisted reproduction.
• Automated microscopic analysis offers objective evaluation but struggles with ovine embryo segmentation due to lipid droplets and cell adhesion.
• Accurate instance segmentation is essential for extracting key morphological and morphokinetic indicators.

Purpose of the Study:

• To develop an advanced deep learning model, BE-DFNet, for precise instance segmentation of ovine embryos.
• To address challenges in segmenting high-density ovine blastomeres with faint boundaries.
• To improve the reliability of automated morphokinetic grading in sheep embryo development.

Main Methods:

• Developed a boundary-enhanced deep learning network (BE-DFNet) incorporating a boundary stream to focus on cell-cell interfaces.
• Validated the model on 847 2D brightfield images of early cleavage stage sheep embryos (1-cell to 16-cell).
• Focused on early stages to leverage critical morphokinetic indicators before compaction limits 2D segmentation accuracy.

Main Results:

• BE-DFNet achieved a panoptic quality of 0.5088, surpassing the state-of-the-art Cellpose-SAM (0.4323).
• Significantly reduced endpoint error from 1.6265 to 0.4632, demonstrating improved separation of adhered blastomeres.
• Effectively mitigated cytoplasmic texture interference and ensured high-precision topological reconstruction of blastomere boundaries.

Conclusions:

• BE-DFNet provides a robust solution for accurate ovine embryo instance segmentation, even with challenging image characteristics.
• The improved segmentation accuracy lays a reliable foundation for fully automated morphokinetic grading in sheep assisted reproduction.
• This advancement has the potential to enhance efficiency and success rates in sheep breeding programs.