WISE: whole-scenario embryo identification using self-supervised learning encoder in IVF.

Area of Science:

• Reproductive biology
• Artificial intelligence in medicine
• Computer vision

Background:

• Accurate embryo identification is crucial in in vitro fertilization (IVF) for successful outcomes.
• Current methods face challenges with varying imaging conditions and data sources.
• Self-supervised learning (SSL) offers a potential solution for improving identification accuracy.

Purpose of the Study:

• To evaluate the effectiveness of a novel self-supervised learning encoder, WISE (Whole-scenario Embryo Identification using Self-supervised learning), for embryo identification in IVF.
• To assess WISE's performance across time-lapse, cross-device, and cryo-thawed embryo image scenarios.
• To compare WISE's accuracy against human embryologists.

Main Methods:

• WISE was developed using a Vision Transformer (ViT) architecture and Masked Autoencoders (MAE) for SSL.
• Three datasets were utilized: SSL pre-training, time-lapse identification, and cross-device identification.
• Embryo images underwent pre-processing including object detection, cropping, padding, and resizing before being fed into WISE.

Main Results:

• WISE achieved high accuracy in embryo identification: 99.89% for time-lapse and 83.55% for cross-device scenarios.
• Performance on a cryo-thawed subset reached 82.22% accuracy, demonstrating real-world applicability.
• SSL integration led to approximately 10% improvement in cross-device and cryo-thawed identification.
• WISE outperformed embryologists by 9.5% to 18% across the evaluated scenarios.

Conclusions:

• SSL methods significantly enhance embryo identification accuracy, particularly in challenging cross-device and cryo-thawed conditions.
• This study marks the first application of SSL in embryo identification, showcasing its potential.
• The WISE model demonstrates considerable promise for future integration into automated embryo witnessing systems.