Explainable AI consensus for grading human blastocysts and blastoids

Area of Science:

• Reproductive Biology
• Artificial Intelligence in Medicine
• Developmental Biology

Background:

• Clinical blastocyst evaluation in assisted reproductive technology (ART) is subjective and suffers from inter-observer variability.
• Stem cell-derived blastoids serve as research models for early human development but require standardized assessment.
• Existing methods lack objectivity and interpretability in morphological grading.

Purpose of the Study:

• To develop a standardized, objective, and explainable AI framework for grading human blastocysts and blastoids.
• To establish functional equivalence between clinical blastocysts and blastoids using AI-driven morphological analysis.
• To improve consistency in embryo selection for ART and validate blastoids as research models.

Main Methods:

• A multi-model consensus framework integrating four Convolutional Neural Networks (CNNs) was developed.
• A large dataset of 14,846 bright-field images of blastocysts and blastoids was used, with Gardner grading as ground truth.
• Explainability was achieved using Grad-CAM for visual attention and a large language model (LLM) for cognitive reasoning and consensus.

Main Results:

• The LLM meta-model achieved a weighted F1-score of 0.99, comparable to the statistical upper bound.
• Grad-CAM analysis confirmed AI decisions were biologically grounded, focusing on the inner cell mass (ICM).
• t-SNE visualization demonstrated morphological similarity between clinical blastocysts and H9-derived blastoids.

Conclusions:

• The AI framework provides an objective and interpretable benchmark for blastocyst and blastoid morphological assessment.
• The study validates stem cell-derived blastoids as a robust model system mirroring key features of clinical blastocysts.
• This technology has the potential to standardize embryo grading in fertility treatments and advance developmental biology research.