Although in vitro fertilization (IVF) has advanced in the past 30 years, grading of embryos at the blastocyst stage remains subjective, with results differing between embryologists. A recent study, published in Nature, looked at the efficacy of using artificial intelligence (AI) to accurately predict the quality of human blastocysts and select the best single embryo for transfer.
For the study, an AI approach was trained to recognize embryo quality using time-lapse images from 10,148 embryos—6000 from 877 good-quality embryos and 6001 from 887 poor-quality embryos. The AI was based on deep neural networks (DNNs), various node layers through which data pass in a multi-step process of pattern recognition.
Training of the DNN, called STORK, involved 50,000 steps. The technology’s performance was evaluated using a randomly selected independent test set with 964 good-quality images from 141 embryos and 966 poor-quality images from 142 embryos.
The results indicated that the algorithm was able to correctly identify good-quality and poor-quality images with 96.94% accuracy (1871 correct predication out of 1930 images). Looking at the fair-quality images, STORK classified 82% of them as good-quality and 18% as poor quality. On examination, the embryos with fair-quality images that STORK classified as poor quality were found to have a lower likelihood of live birth (50.9%) than those classified as good quality (61.4%).
To further evaluate the algorithm’s accuracy, the authors also tested its performance by using additional datasets of embryo images obtained from two other IVF centers. Although the scoring systems used by these centers were different from the system used to train the model, STORK successfully identified and registered score variations and discriminated between them.
Based on their findings, the authors believe that AI can help improve success rates for IVF by removing some of the subjectivity involved in embryo selection. STORK’s ability to correctly distinguish between good-quality and poor-quality embryos, as well as apply its algorithms to unfamiliar embryo images in a manner that agrees with, and at time supersedes, the ratings of human embryologists, the authors said, may help bring uniformity to the field and increase IVF success.