Predicting Embryo Viability Based on Self-Supervised Alignment of Time-Lapse Videos.

Area of Science:

• Embryology
• Computer Vision
• Machine Learning

Background:

• Automating embryo selection in fertility treatments is challenging due to limited labeled data.
• Self-supervised learning offers a way to leverage both labeled and unlabeled data for model pretraining.
• Embryo development involves complex temporal dynamics crucial for successful implantation.

Purpose of the Study:

• To apply a self-supervised video alignment method, temporal cycle-consistency (TCC), to time-lapse embryo videos.
• To extract temporal similarities between embryo videos for predicting pregnancy likelihood.
• To evaluate TCC's performance in transfer learning for semi-supervised embryo evaluation.

Main Methods:

• Applied temporal cycle-consistency (TCC), a self-supervised video alignment technique.
• Utilized 38,176 time-lapse videos of developing human embryos.
• Extracted temporal similarities from video data for predictive modeling.

Main Results:

• The TCC method achieved an AUC of 0.64 for pregnancy likelihood prediction, outperforming the time alignment measurement (TAM) at 0.56.
• Semi-supervised transfer learning using TCC achieved an AUC of 0.66, surpassing standard supervised learning (0.63) with only 16% labeled data.
• TCC-based models demonstrated competitive performance compared to existing embryo evaluation models.

Conclusions:

• Self-supervised learning with TCC effectively extracts temporal embryo features for improved pregnancy prediction.
• TCC enables robust transfer learning, significantly enhancing model performance in low-data regimes.
• This approach offers a promising avenue for advancing automated embryo selection in assisted reproductive technologies.