Artificial intelligence for optimal in vitro fertilization morphokinetics

To create an artificial intelligence model able to determine the morphokinetic phases of an embryo by utilizing time-lapse imaging (TLI) videos.

The dataset utilized was the first publicly available in vitro fertilization dataset for morphokinetic parameter prediction [1]. To create the model, convolutional neural network and EfficientNetB4 (state-of-the-art deep learning model) were employed to demonstrate performance in image classification tasks. After training, the performance of the model was evaluated on a separate hold-out test set using several performance metrics, including accuracy, specificity, Matthews correlation coefficient, and Multiclass Receiver Operating Characteristic area under the curve.

The dataset was collected from 716 infertile couples who underwent intracytoplasmic sperm injection cycles in a university-based in vitro fertilization center and consented to use of monitoring those embryos with time-lapse imaging. This dataset was composed of 704 videos, recorded at 7 focal planes and annotated with 16 cellular events (ranging from polar body appearance through embryo hatching) for a total of 2.4 million images.

The exposure for this decision analysis was the utilization of convolutional neural network and EfficientNetB4 to a publicly available dataset.

To determine effectiveness of a convolutional neural network model for time-lapse imaging classification of human embryonic phases.

The performance metrics from our model include: an overall accuracy of 0.71, sensitivity of 0.59, and specificity of 0.98. Our deep learning model has an area under the curve score of 0.96, demonstrating our model's ability to differentiate between the embryonic phases and perform well in classifying them.

These results suggest that employing a convolutional neural network model is a highly effective approach for time-lapse imaging classification of human embryonic phases. The utilization of artificial intelligence ultimately offers the opportunity to select for the ideal embryo developmental stage, rather than relying on the age of the embryo. With expanding volumes of time-lapse imaging embryonic data, annotations required will outpace the ability of embryologists to review and annotate. An effective artificial intelligence model can optimize the processing of vast amounts of embryo imaging and ultimately help select an embryo at the ideal stage of development.