Quantitative blastocyst expansion features derived via deep learning from time-lapse sequences predict clinical outcomes

At the blastocyst stage, embryo selection is still performed based on morphology assessment following the Gardner and Schoolcraft grading system, despite its subjectivity and lack of quantitative criteria. Continuous culture monitoring through time-lapse incubators enables artificial intelligence (AI) systems based on deep learning to analyze time-lapse images and rank embryos according to their likelihood of leading to a pregnancy. However, the specific developmental features contributing to AI-based embryo ranking remain largely unclear. We aim to develop an AI-based tool to automatically measure the cross-sectional area (CSA) of blastocysts. Using the CSA-derived expansion features, both independently and in combination with other features, we assess the predictive value for clinical outcomes. Retrospective cohort study including 280 patients that had a fresh embryo transfer of a single blastocyst after time-lapse embryo culture. A segmentation model was developed to measure the blastocyst CSA from time-lapse images, from start of blastulation until 112-116 h post-fertilization, resulting in expansion curves. From these curves, quantitative features of blastocyst expansion were extracted. Expansion features, static image convolutional neural network derived features, and clinical parameters such as age were used in different prediction models to evaluate which factors have predictive value. The segmentation model achieved a high accuracy in CSA measurement (Jaccard index of 97.6%). Blastocysts resulting in either biochemical or clinical pregnancy with fetal heartbeat expanded statistically significantly faster than those that did not. Prediction models using clinical and expansion features for biochemical and clinical pregnancy with fetal heartbeat achieved area under the receiver operating characteristics curve (ROC-AUCs) of 0.69 and 0.67, respectively. This performance is comparable to that of the commercially available AI-tool iDAScore v2.0. Our segmentation model allows automatic determination of quantitative expansion features that are associated with pregnancy outcomes. The quantitative nature of these features may complement the existing scoring system, providing a more robust framework for embryo selection. Further studies in larger cohorts are required to confirm our findings and to explore additional quantitative features with predictive value.