An artificial intelligence algorithm to select most viable embryos considering current process in IVF labs

Area of Science:

• Reproductive Medicine
• Artificial Intelligence in Healthcare
• Embryology

Background:

• Morphological assessment of embryos during In-Vitro Fertilization (IVF) is subjective and limits success rates.
• Current embryo selection methods rely on limited visual parameters, leading to suboptimal outcomes.
• Artificial intelligence (AI) offers an objective approach to embryo selection, potentially improving IVF success.

Purpose of the Study:

• To develop and evaluate an AI tool for predicting pregnancy outcomes using embryo images.
• To assess the utility of multiple embryo images versus a single image for pregnancy prediction.
• To compare the performance of AI-based embryo selection with that of experienced embryologists.

Main Methods:

• Collected 252 time-lapse videos of IVF embryos from 2017-2020.
• Extracted key frames at specific post-insemination hours (19, 43, 67).
• Applied Convolutional Neural Network (CNN) architectures with transfer learning, comparing single-image and multi-image AI models against embryologist assessments.

Main Results:

• AI models using three images (DeepEmbryo) demonstrated superior pregnancy prediction accuracy compared to single-image models.
• DeepEmbryo outperformed five experienced embryologists in predicting pregnancy outcomes.
• CNN architectures achieved up to 75.0% accuracy in predicting pregnancy chances via transfer learning.

Conclusions:

• Developed DeepEmbryo, an AI tool utilizing three static embryo images for objective pregnancy prediction.
• DeepEmbryo leverages readily available images from current IVF laboratory processes.
• AI-based tools show significant potential to enhance pregnancy prediction and improve future IVF outcomes.