Human blastoids—3D structures that mimic the early human blastocyst—offer a powerful, ethically viable alternative for studying embryogenesis and testing drug safety. While these models can be generated in massive quantities (up to 30,000 per plate), their evaluation has traditionally relied on laborious manual assessment by experts. This manual bottleneck is prone to human error and subjectivity, often leading to missed biological insights and limiting the scale of experimental applications.
1. A First human blastoid images Dataset
The research team established the first curated brightfield image dataset for human blastoids, comprising 17,133 images. A subset of 2,407 images was meticulously labeled by experts into five morphological categories:

• Class A: Well-formed cavity with an inner cell mass (ICM).
• Class B: Cavity present but no ICM.
• Class C: ICM present but with an irregular trophectoderm.
• Class D: Cellular debris without a cavity.
• Class W: Empty microwells.

After evaluating several architectures, the team selected ResNet-18 for its optimal balance between accuracy and speed. The trained deepBlastoid model achieves accuracy up to 87% and processes 273.6 images per second. This speed represents a high throughput compared to manual evaluation, allowing researchers to analyze entire experimental plates in minutes.
To further enhance reliability for sensitive biological applications, the researchers introduced a Confidence Rate (CR) metric. This system identifies samples where the AI is less certain and flags them for human expert review. By setting a CR threshold of 0.8, the overall classification accuracy improves to 97%, maintaining a high level of automation while ensuring expert-level precision for complex cases.
The utility of deepBlastoid was demonstrated in two real-world use cases: LPA Dosage Optimization: The model analyzed over 10,000 images to identify 0.5 uM as the minimum effective concentration for blastoid formation. It also revealed a significant increase in Class B blastoids at this dosage, a subtle morphological shift easily overlooked by the human eye. DMSO Safety Assessment: The AI confirmed that 0.1% DMSO, a common solvent, does not negatively impact cavitation efficiency, validating its use in drug screening protocols. Additionally, the platform uses the "empty ratio" (Class W) to monitor cell seeding density, providing an automated quality assurance check that enhances experimental reproducibility.
deepBlastoid provides a standardized, automated platform that significantly reduces the manual labor burden on researchers. Beyond basic research, the tool has potential applications in drug toxicity screening, teratogenic effect assessment, and potentially In Vitro Fertilization (IVF) clinical technologies by optimizing embryo quality evaluation. The model and dataset are now publicly available, enabling global researchers to train customized models tailored to their specific imaging protocols.
Author Introduction
Corresponding Authors:

• Professor Mo Li: Associate Professor at the Division of Biomedical Sciences (BioMed), KAUST. He is a core member of the KAUST Center of Excellence for Smart Health (KCSH). His research focuses on stem cell biology, organoid modeling, and regenerative medicine.
• Professor Peter Wonka: Professor of Computer Science at the Computer, Electrical and Mathematical Science and Engineering Division (CEMSE), KAUST. He is a core member of the Visual Computing Center, specializing in computer graphics and deep learning.

Academic Event Preview
One of the study's lead authors, Zejun Fan, will present this research at the upcoming International Society for Stem Cell Research (ISSCR) webinar.

• Theme: Modeling Human Development: Gene Networks, Organoids, and AI Tools
• Time: February 26, 2026, 4:00 PM EST

DOI: https://doi.org/10.1093/lifemedi/lnaf026