EUVIMEDEUVIMED entdecken
Crossref

L26/P-255 Generating synthetic human blastocyst images for in-vitro fertilization blastocyst grading

P. Narahari, S. Rajendran, L. Bori, J. Malmsten, Q. Zhan, Z. Rosenwaks, N. Zaninovic, I. Hajirasouliha

Human Reproduction · 01.07.2026 · 41 (Supplement_1)

DOI: 10.1093/humrep/deag083.591 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von P. Narahari und 7 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Study question Can latent diffusion models generate high-fidelity synthetic blastocyst images with granular control on Gardner-based categories that improve the performance of downstream morphological grading classification models? Summary answer When used for data augmentation, blastocyst images generated by latent diffusion models significantly improved downstream Gardner grade classification accuracy of day 5 blastocysts. What is known already The success of in vitro fertilization (IVF) at many clinics relies on the accurate morphological assessment of day 5 blastocysts, a process that is often subjective and inconsistent. While artificial intelligence can help standardize this evaluation, models require large, diverse, and balanced datasets, which are often unavailable due to data scarcity, natural class imbalance, and privacy constraints. Existing generative embryo models face several limitations in mitigating these issues, such as poor image quality, small training datasets, and non-robust evaluation. Furthermore, their synthetic outputs lack direct utility for downstream clinical tasks within the IVF workflow. Study design, size, duration Diffusion model training used 85716 human embryo images for training and 8716 images for validation from Weill Cornell Medicine. Images were labeled with grades for expansion, inner cell mass, trophectoderm, and z-axis focal depth. The framework was evaluated using Fréchet Inception Distance (FID), a memorization metric (SSCD), manually by four embryologists, and three downstream grade classification experiments using five-fold cross-validation. Participants/materials, setting, methods Images were captured using the EmbryoScope+ TLI system at Weill Cornell Medicine. Four latent diffusion models were developed: three providing granular control over z-axis focal depth and one Gardner-based score (Expansion, Inner Cell Mass, or Trophectoderm), and a fourth integrating z-axis depth with all three scores. Downstream classification evaluation involved training ResNet-50 classifiers to address class imbalance, augmentation of already large and balanced datasets, and replacing parts of real-image only datasets. Main results and the role of chance We present the Diffusion Based Imaging Model for Artificial Blastocysts (DIA) framework, a set of latent diffusion models trained to generate high-fidelity, novel day 5 blastocyst images with granular control of Gardner-based morphological categories and z-axis focal depth. DIA models achieved low FID scores (21.68-23.63), and SSCD analysis confirmed images were novel rather than memorized (threshold < 0.85). Embryologists’ accuracy in distinguishing real vs. synthetic images was near chance (58.6%-64.3%). Most importantly, we demonstrated clear clinical value. Augmenting an imbalanced dataset with synthetic images significantly improved classification accuracy (p < 0.05) and improved performance on initially underrepresented morphological classes. Also, adding synthetic images to an already large, balanced dataset yielded statistically significant performance gains, suggesting that synthetic augmentation benefits even high-resource clinical settings. Finally, synthetic images successfully replaced up to 40% of real clinical data without a statistically significant loss in accuracy, offering a robust solution for clinics facing data scarcity. These findings highlight DIA as a viable solution for low-resource clinics to overcome data limitations in developing AI blastocyst assessment tools and improving standardization in reproductive medicine. Limitations, reasons for caution The models were trained on data from a single high-resource institution using specific imaging (Embryoscope+) and grading (Gardner) standards. Results may require clinic-specific adaptation for different technologies. Wider implications of the findings DIA generates high-fidelity, novel, and controllable synthetic images to address data scarcity and class imbalance in IVF datasets. These augmented datasets can enhance the performance, fairness, and standardization of embryo assessment AI tools. Additionally, the ability to generate specific morphological combinations on demand offers a new tool for embryologist education. Trial registration number No

Zitieren

APA

Narahari, P., Rajendran, S., Bori, L., Malmsten, J., Zhan, Q., Rosenwaks, Z., Zaninovic, N., & Hajirasouliha, I. (2026). L26/P-255 Generating synthetic human blastocyst images for in-vitro fertilization blastocyst grading. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.591

Vancouver

Narahari P, Rajendran S, Bori L, Malmsten J, Zhan Q, Rosenwaks Z, et al. L26/P-255 Generating synthetic human blastocyst images for in-vitro fertilization blastocyst grading. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.591

BibTeX
@article{Narahari2026,
  author = {Narahari, P. and Rajendran, S. and Bori, L. and Malmsten, J. and Zhan, Q. and Rosenwaks, Z. and Zaninovic, N. and Hajirasouliha, I.},
  title = {L26/P-255 Generating synthetic human blastocyst images for in-vitro fertilization blastocyst grading},
  journal = {Human Reproduction},
  year = {2026},
  volume = {41},
  number = {Supplement_1},
  doi = {10.1093/humrep/deag083.591}
}
Kostenlos starten →