Foundation models: the next level of AI in ART

Abstract

Artificial intelligence (AI) in ART has traditionally employed narrow, task-specific models for procedures such as embryo selection and sperm analysis. Although effective, these systems depend on extensive manual annotation and address isolated tasks rather than integrating the diverse data generated in clinical practice. Recently, foundation models, pre-trained on vast, heterogeneous datasets via self-supervised learning, have emerged as promising tools for robust multimodal analysis and decision support. This Directions discusses the technical underpinnings of foundation models, explores their potential applications in ART, and integrates recent innovations that demonstrate how AI-driven methods can improve embryo selection, enable sperm epigenetics diagnostics, and personalize treatment protocols. Key challenges, including data quality, computational infrastructure, and regulatory issues, are also addressed.