Deep learning analysis of endometrial histology as a promising tool to predict the chance of pregnancy after frozen embryo transfers.

ABSTRACT

PURPOSE: Endometrial histology on hematoxylin and eosin (H&E)-stained preparations provides information associated with receptivity. However, traditional histological examination by Noyes' dating method is of limited value as it is prone to subjectivity and is not well correlated with fertility status or pregnancy outcome. This study aims to mitigate the weaknesses of Noyes' dating by analyzing endometrial histology through deep learning (DL) algorithm to predict the chance of pregnancy. METHODS: Endometrial biopsies were taken during the window of receptivity from healthy volunteers in natural menstrual cycles (group A) and infertile patients undergoing mock artificial cycles (group B). H&E staining was performed followed by whole slide image scanning for DL analysis. RESULTS: In a proof-of-concept trial to differentiate group A (n=24) vs. B (n=37), a DL-based binary classifier was trained, cross-validated, and achieved 100% for accuracy. Patients in group B underwent subsequent frozen-thawed embryo transfers (FETs) and were further categorized into "pregnant (n=15)" or "non-pregnant (n=18)" sub-groups based on the outcomes. In the following trial to predict pregnancy outcome in group B, the DL-based binary classifier yielded 77.8% for accuracy. Its performance was further validated by an accuracy of 75% in a "held-out" test set where patients had euploid embryo transfers. Furthermore, the DL model identified histo-characteristics including stromal edema, glandular secretion, and endometrial vascularity as important features related to pregnancy prediction. CONCLUSIONS: DL-based endometrial histology analysis demonstrated its feasibility and robustness in pregnancy prediction for patients undergoing FETs, indicating its value as a prognostic tool in fertility treatment.