Artificial intelligence deep learning model assessment of leukocyte counts and proliferation in endometrium from women with and without polycystic ovary syndrome.

Objective: To study whether artificial intelligence (AI) technology can be used to discern quantitative differences in endometrial immune cells between cycle phases and between samples from women with polycystic ovary syndrome (PCOS) and non-PCOS controls. Only a few studies have analyzed endometrial histology using AI technology, and especially, studies of the PCOS endometrium are lacking, partly because of the technically challenging analysis and unavailability of well-phenotyped samples. Novel AI technologies can overcome this problem.

Design: Case-control study.

Setting: University hospital-based research laboratory.

Patient(s): Forty-eight women with PCOS and 43 controls. Proliferative phase samples (26 control and 23 PCOS) and luteinizing hormone (LH) surge timed LH+ 7-9 (10 control and 16 PCOS) and LH+ 10-12 (7 control and 9 PCOS) secretory endometrial samples were collected during 2014-2019.

Intervention(s): None.

Main Outcome Measure(s): Endometrial samples were stained with antibodies for CD8+ T cells, CD56+ uterine natural killer cells, CD68+ macrophages, and proliferation marker Ki67. Scanned whole slide images were analyzed with an AI deep learning model. Cycle phase differences in leukocyte counts, proliferation rate, and endometrial thickness were measured within the study populations and between the PCOS and control samples. A subanalysis of anovulatory PCOS samples (n = 11) vs. proliferative phase controls (n = 18) was also performed.

Result(s): Automated cell counting with a deep learning model performs well for the human endometrium. The leukocyte numbers and proliferation in the endometrium fluctuate with the menstrual cycle. Differences in leukocyte counts were not observed between the whole PCOS population and controls. However, anovulatory women with PCOS presented with a higher number of CD68+ cells in the epithelium (controls vs. PCOS, median [interquartile range], 0.92 [0.75-1.51] vs. 1.97 [1.12-2.68]) and fewer leukocytes in the stroma (CD8%, 3.72 [2.18-4.20] vs. 1.44 [0.77-3.03]; CD56%, 6.36 [4.43-7.43] vs. 2.07 [0.65-4.99]; CD68%, 4.57 [3.92-5.70] vs. 3.07 [1.73-4.59], respectively) compared with the controls. The endometrial thickness and proliferation rate were comparable between the PCOS and control groups in all cycle phases.

Conclusion(s): Artificial intelligence technology provides a powerful tool for endometrial research because it is objective and can efficiently analyze endometrial compartments separately. Ovulatory endometrium from women with PCOS did not differ remarkably from the controls, which may indicate that gaining ovulatory cycles normalizes the PCOS endometrium and enables normalization of leukocyte environment before implantation. Deviant endometrial leukocyte populations observed in anovulatory women with PCOS could be interrelated with the altered endometrial function observed in these women.