Generative artificial intelligence to produce high-fidelity blastocyst-stage embryo images

Area of Science:

• Embryology
• Artificial Intelligence
• Medical Imaging

Background:

• AI integration in IVF can improve objectivity and automate embryo selection.
• Data scarcity and privacy concerns limit AI effectiveness in IVF.
• Generative adversarial networks (GANs) offer a solution by creating synthetic data.

Purpose of the Study:

• To assess the capability of generative AI models in producing high-fidelity human blastocyst images.
• To develop robust AI models for IVF by generating substantial training datasets.
• To explore the potential of GANs in overcoming data limitations in assisted reproductive technologies.

Main Methods:

• A style-based GAN was fine-tuned using 972 blastocyst images from time-lapse microscopy videos.
• Models were configured with data augmentation and pretrained weights for optimization.
• Quantitative metrics (FID, KID) and a visual Turing test with 60 evaluators assessed image quality and fidelity.

Main Results:

• The AUG + Pretrained-R model achieved the best performance with FID of 15.2 and KID of 0.004 after 5000 iterations.
• Pretrained models demonstrated significantly lower FID and KID scores compared to baseline models.
• Expert and non-expert evaluations in the visual Turing test showed comparable specificity for synthetic images.

Conclusions:

• Generative AI can produce high-fidelity human blastocyst images, creating valuable training datasets.
• This approach addresses data scarcity and enhances patient data privacy in IVF.
• Generative models show potential to transform embryo selection and improve IVF outcomes.