Deep learning based on images of human embryos obtained from high-resolusion time-lapse cinematography for predicting good-quality embryos

Abstract

Evaluating the potential of embryos to develop into good-quality embryos is very important in human assisted reproductive technologies. For this reason, time-lapse monitoring of human embryos cultured in vitro has been widely used to select embryos in a clinical setting. However, there are significant limitations with the accuracy of predicting good-quality embryos even if morphokinetic analyses are applied. Here we examined whether deep learning based on images of human embryos obtained from high-resolution time-lapse cinematography (hR-TLC) could predictively determine good-quality embryos. Retrospective cohort study. Time-lapse images of 118 normally fertilized embryos from conventional in vitro fertilization (cIVF) and intracytoplasmic sperm injection (ICSI) obtained between April 2003 and November 2008 were used in this study. Images of the embryos in culture (37°C; pH 7.35 ± 0.02) were captured every two minutes for 40 hours (approximately 2,000 photographs in total) using an inverted microscope with differential interference contrast. For both cIVF (n = 48) and ICSI (n = 70) embryos, images captured for 30 hours starting immediately after the second polar body extrusion were selected and used as data input for deep learning. Only 64 embryos developed into good-quality embryos (Grades 1 and 2 based on modified Veeck’s criteria; cIVF: n = 19, ICSI: n = 45). The remaining embryos were of poor quality (cIVF: n = 29, ICSI: n = 25). Good- and poor-quality embryos (n = 45 from each group) were randomly selected for training, while five embryos from each group were used for validation. Deep learning was performed with the Keras neural network library. We conducted two different examinations to predict embryo quality. Firstly, good-quality embryos at the four-cell stage were predictively determined using one image obtained immediately after the first cleavage as the input value for deep learning. After 50 learning sessions, we achieved 94% correct answers for the training dataset and 70% for the validation dataset. Secondly, good-quality embryos at the four-cell stage were predictively determined using 31 images captured hourly for 30 hours. After 50 learning sessions, we achieved 92% correct answers for the training dataset and 80% for the validation dataset. This study showed that deep learning based on hR-TLC images was able to predict good-quality embryos with an accuracy of 80%. Although the prediction accuracy needs improving, this can be achieved by updating the deep learning system and increasing the number of cases. Once a system for predicting the quality of human embryos is constructed by deep learning based on hR-TLC images, it may become possible to predetermine good-quality embryos at early embryonic stages, such as before syngamy.