Choosing which embryos to implant falls to a fertility clinic’s embryologists, specialists who use a handful of methods — including microscopy and time-lapse monitoring of an embryo at the stage in which it consists of just 200 to 300 cells — to determine an embryo’s viability. Determining which embryo has the best chance of developing into a healthy pregnancy is a subjective process, and even experienced embryologists disagree.
Now, with the use of AI, embryo selection is poised to become far more precise.
“We wanted to develop an objective method that can be used to standardize and optimize the selection process to increase the success rates of IVF,” says Dr. Nikica Zaninovic, director of the Embryology Lab at Weill Cornell Medicine’s Center for Reproductive Medicine.
Dr. Zaninovic is a co-senior author of the team’s study, published in the April 2019 edition of NPJ Digital Medicine, in which they used 12,000 photos of human embryos taken 110 hours after fertilization to train an AI algorithm to discriminate between poor and good embryo quality.
To arrive at the designations, each embryo was first assigned a grade by embryologists who considered various aspects of the embryo’s appearance. Then the team, co-led by Dr. Pegah Khorasvi, postdoctoral associate, and Dr. Iman Hajirasouliha, assistant professor of physiology and biophysics and a member of the Englander Institute for Precision Medicine, both of Weill Cornell Medicine, performed a statistical analysis to determine the probability of an embryo becoming a successful pregnancy. Embryos were considered good quality if the chances were greater than 58 percent and poor quality if the chances were below 35 percent. After training and validation, the algorithm, called Stork, was able to classify the quality of a new set of images with 97 percent accuracy.
This type of AI is also called “deep learning,” an approach that is roughly modeled after the neural networks of the brain, which analyze information in increasing layers of complexity. As the computer is fed new information, its ability to recognize the desired patterns, whether they are the features of a healthy embryo or the cells of a lung cancer tumor, improves automatically. The size of the training data set is critically important to the success of the algorithm, with more data leading to better outcomes.