Deep learning-based early screening for porcine embryos with different developmental potential

The developmental potential of in vitro embryos is a critical determinant of porcine artificial reproduction technology (ART) efficiency. However, their high cytoplasmic lipid content severely compromises optical clarity, making reliable prediction via microscopic observation challenging. Here, through the establishment of a novel embryonic droplet culture system, we developed and benchmarked multiple deep learning-based models to identify early-stage embryos with a high probability of developing into blastocysts. Using this system, we collected 10,041 bright-field images of individual porcine parthenogenetically activated (PA) embryos spanning from the 1-cell to the blastocyst stage, with manually curated developmental outcome annotations. By systematic benchmarking, we identified MaxviT_T as the most efficient model on our dataset with peaked prediction performance at the 4-cell stage. This robustness was further confirmed by the consistent prediction efficiency across different experimental batches with variable blastocyst formation rates. In summary, the application of the MaxViT_T predictive model can predict the potential of early-stage embryonic development, providing a novel approach to high-quality early-stage embryos screening.