Deep learning-based embryo assessment of static images can reduce the time to live birth in in vitro fertilization

The low success rate in in vitro fertilization (IVF) may be related to our inability to select embryos with good implantation potential by traditional morphology grading and remains a great challenge to clinical practice. Multiple deep learning-based methods have been introduced to improve embryo selection. However, existing methods only achieve limited prediction power and generally ignore the repeated embryo transfers from one stimulated IVF cycle. To improve the deep learning-based models, we introduce Embryo2live, which assesses the multifaceted qualities of embryos from static images taken under standard inverted microscope, primarily in vision transformer frameworks to integrate global features. We first demonstrated its superior performance in predicting Gardner’s blastocyst grades with up to 9% improvement from the best existing method. We further validated its high capability of supporting transfer learning using the large clinical dataset of the Centre. Remarkably, when applying Embryo2live to the clinical dataset for embryo prioritization, we found it improved the live birth rates of the Top 1 embryo in patients with multiple embryos available for transfer from 23.0% with conventional morphology grading to 71.3% using Embryo2live, reducing the average number of embryo transfers from 2.1 to 1.4 to attain a live birth.