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

Efficient selection of in vitro-fertilized embryos is crucial for assisted reproduction in sheep, and automated microscopic analysis can enable objective evaluation. Instance segmentation is a key prerequisite for extracting downstream morphological and morphokinetic indicators. However, ovine embryo microscopic images typically exhibit abundant cytoplasmic lipid droplets and tightly packed blastomeres, causing conventional segmentation methods to fail when cell boundaries are faint. Here, we developed a boundary-enhanced deep learning network (BE-DFNet) to address high-density cell adhesion. By introducing a boundary stream as a structural prior to focus on cell-cell contact interfaces, the model was validated on 847 2D brightfield images covering the early cleavage stages (from 1-cell to 16-cell). We specifically focused on these stages because they provide critical morphokinetic indicators (e.g., t2 to t8) before significant blastomere overlapping in the compacted morula stage physically limits the reliability of 2D instance segmentation. BE-DFNet achieved a panoptic quality of 0.5088, outperforming a representative state-of-the-art method, Cellpose-SAM (0.4323). Notably, BE-DFNet reduced endpoint error, which reflects topological consistency, from 1.6265 to 0.4632, indicating improved separation of adhered blastomeres. Collectively, BE-DFNet effectively mitigates cytoplasmic texture interference and ensures the high-precision topological reconstruction of blastomere boundaries, providing a reliable foundation for fully automated morphokinetic grading.