Deep learning methods to forecasting human embryo development in time-lapse videos

Area of Science:

• Embryology
• Artificial Intelligence
• Reproductive Medicine

Background:

• Embryo quality assessment is critical in assisted reproductive technology (ART) for selecting viable embryos and determining optimal transfer timing.
• Current AI tools automate assessment but lack predictive capabilities for future embryo development.
• There is a need for AI systems that can forecast embryo morphology dynamics over time.

Purpose of the Study:

• To develop an AI system capable of forecasting embryo morphology dynamics.
• To assist embryologists in the early assessment and selection of embryos for transfer.
• To predict future morphological changes in embryos beyond current observational capabilities.

Main Methods:

• The AI system analyzes past embryo development (2 hours) to predict future morphological changes (up to 23 hours).
• A novel predictive model utilizing Convolutional LSTM layers enables recursive forecasting of embryo development.
• The model analyzes prior video sequence changes to predict morphology.

Main Results:

• The AI system accurately forecasted embryo development at cleavage (day 2) and blastocyst (day 4) stages.
• Valuable insights were provided on cell division processes and blastocyst formation.
• Forecasts for 'transfer' category embryos showed clearer cell membranes and less distortion compared to 'avoid' categories.

Conclusions:

• The AI system provides early insights into embryo quality, assisting in the evaluation for both transfer and avoidance.
• Embryologists can utilize the forecast to visualize and understand embryo morphological changes.
• Improving image quality could enhance the clinical relevance of this predictive AI approach.