Dynamic changes in AI-based analysis of endometrial cellular composition: Analysis of PCOS and RIF endometrium

Abstract

Background: The human endometrium undergoes a monthly cycle of tissue growth and degeneration. During the mid-secretory phase, the endometrium establishes an optimal niche for embryo implantation by regulating cellular composition (e.g., epithelial and stromal cells) and differentiation. Impaired endometrial development observed in conditions such as polycystic ovary syndrome (PCOS) and recurrent implantation failure (RIF) contributes to infertility. Surprisingly, despite the importance of the endometrial lining properly developing prior to pregnancy, precise measures of endometrial cellular composition in these two infertility-associated conditions are entirely lacking. Additionally, current methods for measuring the epithelial and stromal area have limitations, including intra- and inter-observer variability and efficiency.

Methods: We utilized a deep-learning artificial intelligence (AI) model, created on a cloud-based platform and developed in our previous study. The AI model underwent training to segment both areas populated by epithelial and stromal endometrial cells. During the training step, a total of 28.36 mm2 areas were annotated, comprising 2.56 mm2 of epithelium and 24.87 mm2 of stroma. Two experienced pathologists validated the performance of the AI model. 73 endometrial samples from healthy control women were included in the sample set to establish cycle phase-dependent dynamics of the endometrial epithelial-to-stroma ratio from the proliferative (PE) to secretory (SE) phases. In addition, 91 samples from PCOS cases, accounting for the presence or absence of ovulation and representing all menstrual cycle phases, and 29 samples from RIF patients on day 5 after progesterone administration in the hormone replacement treatment cycle were also included and analyzed in terms of cellular composition.

Results: Our AI model exhibited reliable and reproducible performance in delineating epithelial and stromal compartments, achieving an accuracy of 92.40% and 99.23%, respectively. Moreover, the performance of the AI model was comparable to the pathologists' assessment, with F1 scores exceeding 82% for the epithelium and >96% for the stroma. Next, we compared the endometrial epithelial-to-stromal ratio during the menstrual cycle in women with PCOS and in relation to endometrial receptivity status in RIF patients. The ovulatory PCOS endometrium exhibited epithelial cell proportions similar to those of control and healthy women's samples in every cycle phase, from the PE to the late SE, correlating with progesterone levels (control SE, r2 = 0.64, FDR < 0.001; PCOS SE, r2 = 0.52, FDR < 0.001). The mid-SE endometrium showed the highest epithelial percentage compared to both the early and late SE endometrium in both healthy women and PCOS patients. Anovulatory PCOS cases showed epithelial cellular fractions comparable to those of PCOS cases in the PE (Anovulatory, 14.54%; PCOS PE, 15.56%, p = 1.00). We did not observe significant differences in the epithelial-to-stroma ratio in the hormone-induced endometrium in RIF patients with different receptivity statuses.

Conclusion: The AI model rapidly and accurately identifies endometrial histology features by calculating areas occupied by epithelial and stromal cells. The AI model demonstrates changes in epithelial cellular proportions according to the menstrual cycle phase and reveals no changes in epithelial cellular proportions based on PCOS and RIF conditions. In conclusion, the AI model can potentially improve endometrial histology assessment by accelerating the analysis of the cellular composition of the tissue and by ensuring maximal objectivity for research and clinical purposes.

Keywords: Artificial intelligence; Computational histology; Endometrium; IVF; Polycystic ovary syndrome; Recurrent implantation failure.

Graphical abstract

Fig. 1

Graphical overview of sample information and the AI analysis protocol. Twenty (20) control women and 18 PCOS women provided multiple samples in different cycle phases during different menstrual cycles. The endometrial biopsy samples were categorized according to menstrual cycle phase, ovulation, and endometrial receptivity. The AI model was trained to segment the epithelium (red) and stroma (yellow) using IHC slides. PE, proliferative phase; SE, secretory phase; P4, progesterone; HRT, hormone replacement treatment; IHC, immunohistochemistry; CNN, convolutional neural network.

Fig. 2

Examples of convolutional neural network (CNN)1 training. CNN1 training was performed within the area indicated by the solid black line. (A) Original image, (B) CNN1 training for segmenting the epithelium and stroma, and (C) the training result. The epithelium is marked red, and the stroma is marked yellow.

Fig. 3

Epithelium percentages according to the cycle phases and PCOS diagnosis. Epithelium percentages in the controls (PE, n=12; ESE, n=15; MSE, n=26; LSE, n=20) and women with PCOS (PE, n=24; ESE, n=14; MSE, n=21; LSE, n=19; Anovulatory, n=12). Each symbol represents an individual sample in the controls (blue dot) and the PCOS samples (orange triangle), and the lines represent the upper quartiles, medians, and lower quartiles. ^ap<0.05 and ^a′p<0.001 when compared to the PE samples, ^bp<0.001 when compared to the MSE samples, and ^cp<0.01 and ^c′p<0.001 when compared to the anovulatory samples by the mixed model analysis. PE, proliferative phase; ESE, early secretory phase; MSE, mid-secretory phase; LSE, late secretory phase.

Fig. 4

Epithelial proportion according to endometrial receptivity status in RIF patients. Epithelium percentages according to distinct endometrial receptivity profiles (Pre-receptive (Pre) RIF, n=9; Receptive (Re) RIF, n=10; Post-receptive (Post) RIF, n=10). Each symbol represents an individual sample in Pre (blue dot), Re (green triangle), and Post RIF (red square), and the lines represent the upper quartiles, medians, and lower quartiles. The statistical differences were calculated by the Kruskal–Wallis test. Pre, pre-receptive; Re, receptive; Post, post-receptive.