A novel deep learning approach to identify embryo morphokinetics in multiple time lapse systems.

Area of Science:

• Embryology
• Artificial Intelligence
• Computer Vision

Background:

• Time-lapse systems (TLS) in In Vitro Fertilization (IVF) provide continuous embryo monitoring.
• Deep-learning algorithms are increasingly used to analyze IVF data, but focus less on identifying specific developmental events.
• Accurate identification of key morphokinetic events is crucial for embryo assessment.

Purpose of the Study:

• To develop and validate a novel deep-learning architecture for the automatic detection of 11 key kinetic events in human embryos.
• To create a model applicable across different TLS platforms used in IVF laboratories.
• To address the gap in AI research concerning the identification of specific embryo morphokinetic events.

Main Methods:

• A dataset of 1909 embryos from multiple clinics using different TLS (EMBRYOSCOPE, GERI, MIRI) was utilized.
• A Transformer-based video backbone was trained using a custom metric to learn the ordinal structure of developmental events.
• Embeddings from the backbone were processed by a Gated Recurrent Unit (GRU) sequence model to capture kinetic dependencies.

Main Results:

• The model achieved a weighted average of 66.0% precision, 67.6% recall, and 66.3% F1-score on a test set of 278 embryos.
• The developed deep-learning architecture demonstrated applicability across multiple TLS platforms.
• The model successfully identified 11 key kinetic events from the 1-cell stage to blastocyst.

Conclusions:

• The proposed deep-learning model effectively automates the detection of crucial embryo morphokinetic events.
• This AI tool has the potential to assist embryologists in IVF labs by providing objective and consistent event identification.
• The model's cross-platform compatibility enhances its utility in diverse IVF settings.