Explainable AI consensus for grading human blastocysts and blastoids

The morphological evaluation of human blastocysts is a critical yet subjective step in clinical assisted reproductive technology (ART), often limited by inter-observer variability. To establish a standardized and objective grading system for both clinical blastocysts and stem cell-derived blastoids, we developed an explainable artificial intelligence-driven multi-model consensus framework. Our methodology utilized a large-scale dataset of 14,846 bright-field images (14,640 blastocysts and 206 blastoids), with the Gardner grading system (Grades 1-5) established as the ground truth by experienced embryologists. High-quality embryos were defined as those reaching expansion Grades 4 or 5 with robust inner cell mass (ICM) and trophectoderm scores. To prevent data leakage and ensure generalizability, the dataset was strictly partitioned into training and testing sets at the individual embryo/blastoid level rather than the image level. The framework integrates an ensemble of four fine-tuned convolutional neural networks (ResNet18, VGG16-BN, ResNet50, and EfficientNet-B0) for visual assessment, complemented by visual explainability via Grad-CAM and cognitive reasoning via the Gemini 2.5 Flash large language model (LLM). Quantitative evaluation revealed that the LLM meta-model achieved a weighted F1-score of 0.99, demonstrating functional equivalence to the statistical upper bound established by an XGBoost meta-model (F1-score: 1.00). Grad-CAM analysis confirmed that the models' decisions were biologically grounded, with high-quality predictions consistently correlating with high focus scores on the ICM. Furthermore, t-SNE visualization provided objective evidence of morphological similarity between clinical blastocysts and H9-derived blastoids. This hybrid framework establishes a new benchmark for objective, interpretable embryo grading and validates blastoids as a robust model system.

In fertility clinics, selecting the best embryo for in vitro fertilization relies on visual inspection by experts, which can be subjective and inconsistent. In this study, 'clinical blastocysts' refers to early human embryos generated during ART procedures and imaged at the blastocyst stage before clinical decision-making. Scientists also use 'stem cell-derived blastoids', which are blastocyst-like structures generated from human stem cells in the laboratory. They are not embryos, but research models that reproduce selected structural features of early blastocysts and can help researchers study early human development under controlled laboratory conditions. To improve the evaluation of both clinical blastocysts and stem cell-derived blastoids, we developed an artificial intelligence system that acts like a panel of experts. First, the computer visually inspects thousands of images. Then, an advanced language model reviews these visual clues to agree on a final grade and explains its reasoning in simple text. We found that this system was accurate and interpretable because it highlighted which parts of the image it used to make its decision. This technology may help standardize embryo morphology assessment in fertility treatment and support early-development research by evaluating whether stem cell-derived blastoids share selected visible features with clinical blastocysts.