Embryo Prediction Network: Determining Embryo Transfer Strategies with Time-Lapse Imaging for In-Vitro Fertilization Treatment

In this paper, we propose the Embryo Prediction Network, named EPNet, which aims to assist reproductive specialists in predicting blastocyst grades from day 3 embryo images, thereby optimizing embryo culture times and improving IVF success rates. EPNet comprises three subnetworks: Morula, Blastocyst, and Embryo Prediction, leveraging guided learning and a multi-loss function to address data imbalance and enhance prediction accuracy. The network was trained and evaluated on an internal testing set using 5,586 images from 1,862 embryos across days 3, 4, and 5, and further externally validated using 596 images from 298 embryos across days 3 and 5. Its performance was compared with six representative machine learning-based models and four modern deep learning-based models. EPNet achieved accuracies of 82.82% on the internal testing set and 82.52% on the external validation set, outperforming the competing models. In a pilot clinical comparison involving 18 cases per group, the EPNet-assisted strategy yielded improvements exceeding 10 percentage points in overall pregnancy rate, implantation rate, and ongoing pregnancy rate, compared with the traditional decision-making strategy.