BACKGROUND: The application of videofixation technologies in embryology is developing significantly. These technologies permit the objective analysis of the process of early embryogenesis of each cultured embryo without the necessity of removing the culture cup from the incubator. Timelapse technologies in routine practice allow for the guaranteed detection of embryo developmental pathologies that are inaccessible to traditional developmental monitoring methods [1, 2]. Nevertheless, the annotation and manual evaluation of all frames captured during the cultivation process can be a time-consuming process. Furthermore, video fixation itself does not eliminate the issue of objectivizing the quality of interpretation of the obtained images [3]. Intelligent technologies, in particular, solutions developed with the use of machine learning, are successfully employed in the resolution of such problems.
AIM: The aim of this study is to develop a system for the automated analysis of the morphokinetic state of the human embryo with the aim of assessing its capacity for implantation.
MATERIALS AND METHODS: The data were collected at the Family Medical Center (Ufa, Russia) and the Clinical Hospital IDK of the Mother and Child Group of Companies (Samara, Russia). Digital images of the period of preimplantation development of human embryos up to the blastocyst stage (days 0–6 from insemination) were obtained using an incubator for in vitro fertilization laboratories, the EmbryoVisor, with a timelapse (hyperlapse) video fixation system. Embryos were cultured individually in special micro-well WOW dishes (Vitrolife, Sweden). The data set was labelled using Label Studio Community Edition software. A recurrent convolutional neural network was selected to analyse the data and trained using multiple images.
RESULTS: The development of the automatic analysis system is based on the classification of the morphokinetic state of the embryo according to the stages of embryogenesis: fertilization, fragmentation, morula formation, and blastocyst formation. Segmentation of multiple objects, such as pronuclei and polar bodies at the fertilization stage or blastomeres at the fragmentation stage, will be performed depending on a certain stage of development. We plan to build a binary classification of the presence of additional features (multinucleation, heterogeneity of the endoplasmic network), classification/regression of additional features (so, fragmentation can be estimated as discrete ranges or absolute values). The result is a system for labeling the morphodynamic profile of an embryo using deep learning. This method automates and accelerates the analysis process, which previously required significant time and human resources.
CONCLUSIONS: It is anticipated that the developed system of automatic analysis of morphokinetic state of embryos will simplify the process of evaluating the quality of human embryos in in vitro fertilization laboratories, reducing the time and resources spent on this process. Furthermore, it will enhance the accuracy and reliability of assessing the implantation ability of embryos and could potentially serve as the foundation for the development of a support system for medical decision-making in embryology.
