L26/P-356 Predicting IVF pregnancy outcomes with deep learning-based histological analysis of the endometrium
Human Reproduction · 01.07.2026 · 41 (Supplement_1)
DOI: 10.1093/humrep/deag083.692 ↗Kurz gefasst
Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von Y.L. Lee und 10 weiteren Autor:innen.
Abstract (Originalsprache)
Study question Can deep learning models trained on endometrial histology slides predict the cumulative live birth (CLB) rate of an in vitro fertilization (IVF) cycle? Summary answer Histopathology-pretrained foundation models like UNI2-h can be repurposed for predicting CLB in IVF, offering superior generalizability and convenience compared to training models from scratch. What is known already Endometrial environment is critical for successful pregnancy, but an accurate method to predict endometrial receptivity is currently unavailable. The recent emergence of artificial intelligence (AI) technologies has garnered significant interest in using deep learning for less subjective assessment of endometrial receptivity. Although pregnancy outcome prediction models using whole slide images (WSI) of hematoxylin and eosin (H&E) stained endometrial sections from patients following hormone replacement treatment (HRT) was reported recently, the applicability of the model on natural cycle endometrial WSI is unknown. Furthermore, a significant knowledge gap remains regarding the transferability of modern foundation models to the endometrial function-oriented tasks. Study design, size, duration This is a secondary analysis of a prospective study. Infertile women (n = 183) who had completed their first IVF cycle at the Centre of Assisted Reproduction and Embryology, the University of Hong Kong - Queen Mary Hospital and the Assisted Reproduction Centre of Kwong Wah Hospital between 2017 and 2021 were included. An external cohort comprising 22 fertile patients WSIs collected at a natural cycle from the University of Hong Kong-Shenzhen Hospital was also included. Participants/materials, setting, methods H&E staining and WSIs were obtained from endometrial samples taken day 7-8 after the LH surge in a natural cycle prior to the IVF treatment. Two separate deep learning architectures, the convolutional neural network-based ResNet-18 and the histopathology-pretrained Vision Transformer foundation model (UNI2-h) were used to develop binary classification models to differentiate endometrial H&E WSIs based on their CLB outcome potentials. We also developed biologically-targeted models trained exclusively on manually annotated luminal epithelium (LE) regions. Main results and the role of chance The usability of the HRT-trained model was first tested on the 183 natural cycle WSIs. However, a relatively low area (0.58) under receiver operating characteristic curve (AUROC) indicated it failed to generate a useful prediction. New ensemble models using ResNet-18 was then trained on the 84 samples with CLB and 99 non-pregnant samples. The models exhibited robust predictive capability on the internal (balanced accuracy: 76%; AUROC: 0.84), but poor prediction in external (balanced accuracy: 49%; AUROC: 0.4) test set. We further trained a non-linear multi-layer perception (MLP) classifier on the features extracted by the frozen UNI2-h model, which yielded consistent performance for both internal (balanced accuracy: 71%; AUROC: 0.74) and external (balanced accuracy: 75%; AUROC: 0.85) test sets. We also developed biologically-targeted models trained exclusively on manually annotated LE of the endometrium. The predictive capability of UNI2-h-based LE-model was comparable to WSI models in internal (balanced accuracy: 77%; AUROC: 0.74) and external (balanced accuracy: 79%; AUROC: 0.83) test sets. This study demonstrated that histopathology-based foundation models like UNI2-h could be effectively repurposed for maternal reproductive medicine-oriented tasks. Targeted training on the LE regions highlighted the rich morphological signatures which requires specific attention in future analysis. Limitations, reasons for caution The current AI models predicting CLB in IVF were constrained by a relatively small sample size, which might have caused performance overestimation and reduced generalizability of the models to new patients. Additionally, the lack of multicentre validation limits evaluation of robustness across diverse populations, protocols, and imaging systems. Wider implications of the findings Deep learning models for analyzing endometrial histology hold great promise on pregnancy outcome predictions. By objectively quantifying subtle morphological patterns by AI models, more accurate, personalized fertility assessments could be developed to improve patient prognoses and guide targeted clinical interventions. Trial registration number No
Zitieren
APA
Lee, Y., Xu, N., Chen, A., Ruan, H., Fong, S., Yang, D., Qi, X., Liao, R., Cao, D., Yeung, W., & Ng, E. (2026). L26/P-356 Predicting IVF pregnancy outcomes with deep learning-based histological analysis of the endometrium. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.692
Vancouver
Lee Y, Xu N, Chen A, Ruan H, Fong S, Yang D, et al. L26/P-356 Predicting IVF pregnancy outcomes with deep learning-based histological analysis of the endometrium. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.692
BibTeX
@article{Lee2026,
author = {Lee, Y.L. and Xu, N. and Chen, A.C.H. and Ruan, H. and Fong, S.W. and Yang, D. and Qi, X. and Liao, R. and Cao, D. and Yeung, W.S.B. and Ng, E.H.Y.},
title = {L26/P-356 Predicting IVF pregnancy outcomes with deep learning-based histological analysis of the endometrium},
journal = {Human Reproduction},
year = {2026},
volume = {41},
number = {Supplement_1},
doi = {10.1093/humrep/deag083.692}
}
EUVIMED