AI System 100% Accurate in Tracking Identity of IVF Embryos
- Massachusetts General Hospital researchers previously developed convolutional neural networks (CNNs) that can detect subtle morphologic differences among embryos and a genetic algorithm that generates a unique identifier for each embryo of a patient
- In this study, the combination of those tools was trained and validated on a cohort of 4,889 time-lapse imaging videos of embryos that had been collected using an EmbryoScope (Vitrolife, Sweden)
- The system was tested in 400 patients—each patient's cohort of embryos was grouped with seven other randomly selected embryo cohorts for testing on days 3 and day 5 of development
- On both days of development, the system was 100% accurate in matching embryos to the correct patient
- It should be possible to readily incorporate this novel application of artificial intelligence (AI) into any in vitro fertilization lab that images embryos throughout development
Although extremely rare, human error occurs in the in vitro fertilization (IVF) lab. It can result in the loss of embryos or, worse, the implantation of embryos that don't belong to the patient.
Subscribe to the latest updates from OB/GYN Advances in Motion
The current gold standard of identifying gametes is initial labeling with supervision, or "double witnessing," where two people physically verify the identity of the patients and their embryos. The additional time and personnel required are a serious limitation and the possibility of misidentification still exists. Electronic witnessing systems such as barcodes or radiofrequency identification are used in some laboratories but are not standardized or routine.
Victoria S. Jiang, MD, a clinical fellow at Massachusetts General Hospital's Fertility Center, Charles L. Bormann, PhD, director of the Vincent Embryology Laboratory, and colleagues have created the first artificial intelligence (AI)–based witnessing software for embryo identification in the IVF laboratory. They describe their system in the Journal of Assisted Reproduction and Genetics.
Developing the System
Mass General researchers previously developed two deep convolutional neural networks (CNNs) capable of detecting subtle morphologic differences among embryos (published in Helivon Open Access). CNNs, which process an image as millions of data points, are trained on a vast number of visual images and "learn" how to interpret additional images. In the IVF setting, they can detect far more features in a cell than even the most skillful human eye.
In this study, the researchers combined those CNNs with a genetic algorithm they previously developed (published in eLife) that generates a unique identification score for each patient embryo. The resulting model was trained and validated on a cohort of 4,889 time-lapse imaging videos of embryos collected using EmbryoScope.
The new witnessing software processes images of a set of embryos from each patient. It generates identifiers that can be used later to determine whether the embryos originated from the same patient.
Testing the System
The system was tested on 400 patients (2–12 embryos per patient) who had undergone fresh IVF. Each patient's cohort of embryos was grouped with seven other randomly selected cohorts.
Embryos were assessed at the:
- Cleavage stage, day 3—A identifier was generated for each embryo at 65 hours post-insemination (hpi); at 70 hpi the embryos were reassessed and were assigned an identifier that was matched to a patient identification key
- Blastocyst stage, day 5—Embryos were assessed at 105 and 110 hpi
For each patient, the researchers calculated whether embryos were identified accurately at the 70 hpi time point vs. 65 hpi, and at the 110 hpi time point vs. 105 hpi.
The CNN was 100% accurate in matching embryos to the correct patient identifier on both day 3 and day 5.
Ready for the Clinic
Because the new witnessing system was tested at standard time points for embryo morphologic assessment, IVF labs will be able to integrate it easily into any workflow that includes EmbryoScope time-lapse imaging.
view original journal article Subscription may be required
Learn more about fertility care at Mass General
Refer a patient to the Mass General Fertility Center