Combined Input Deep Learning Pipeline for Embryo Selection for In Vitro Fertilization Using Light Microscopic Images...

Area of Science:

• Reproductive Medicine
• Artificial Intelligence in Healthcare
• Embryology

Background:

• Current in vitro fertilization (IVF) embryo selection relies on subjective morphological evaluation by embryologists.
• This manual assessment can lead to variability and potential inaccuracies in embryo viability classification.

Purpose of the Study:

• To develop and evaluate a deep learning-based pipeline for objective embryo viability classification.
• To integrate microscopic images with clinical and pseudo-features for enhanced prediction accuracy.

Main Methods:

• A deep learning pipeline was created using combined inputs: microscopic embryo images and patient data (age, Istanbul grading scores).
• Convolution-based transfer learning models (EfficientNet-B0) and self-supervised learning (SimCLR) with generative adversarial networks (GANs) were employed.
• Hyperparameter optimization was performed using Optuna for model tuning.

Main Results:

• The best model, an optimized EfficientNet-B0, achieved an F1 score of 65.02%, accuracy of 69.04%, sensitivity of 56.76%, and AUC of 66.98%.
• The deep learning approach demonstrated an advantage in accuracy and comparable AUC to existing ensemble methods.

Conclusions:

• The developed deep learning pipeline offers a promising, objective alternative to traditional subjective embryo assessment in IVF.
• This AI-driven method has the potential to improve the efficiency and accuracy of embryo selection for better IVF outcomes.