Prior research has demonstrated that morphological characteristics and cell division timings can indicate the quality and developmental potential of an embryo. However, conventional evaluation techniques miss minute quantitative changes that can be identified by sophisticated computer vision analysis. Time-lapse data and embryo classifications from 2170 embryos grown for at least 110 h at a single IVF centre were examined in this retrospective cohort study. An extra 326 embryos were set aside for blind testing so that the model’s performance could be compared to that of an embryologist. The metrics were extracted by training an Attention U-Net model to segment the embryos in time-lapse images, enabling a robust analysis of the morphological patterns. The final machine learning mode was trained using extracted data from embryo surface metrics time series to perform a binary classification task, in order to discriminate between embryos that reached blastocyst stage and those who did not. The machine learning model was able to predict if an embryo will reach the blastocyst stage with an AUC of 0.85 95% CI 0.88–0.93 in a blind test. The model was able to outperform the embryologist’s prediction in accuracy by 12% (0.791 vs 0.696) and in F1 Score by 8.5% (0.811 vs 0.742).
Tsakalis et al. (Wed,) studied this question.