L26/P-362 Artificial intelligence analysis of transvaginal endometrial ultrasound images predicts clinical pregnancy in frozen embryo transfer cycles: a first-in-field exploratory retrospective study

Le résumé fourni par la source

Abstract Study question Can an artificial intelligence (AI)-derived score, from transvaginal endometrial ultrasound images and clinical data, predict clinical pregnancy (CP) in women undergoing frozen embryo transfer (FET)? Summary answer Endometrium-AI showed modest predictive ability for CP yet identified a higher probability subgroup, suggesting AI-based analysis may support non-invasive endometrial assessment pending larger validation. What is known already Assessment of endometrial receptivity remains a major challenge in assisted reproduction. Current clinical evaluation relies largely on endometrial thickness and morphology, which have limited predictive value for implantation and pregnancy. More advanced tools, including molecular and transcriptomic assays, have been developed but are invasive, costly, and not widely accessible. Ultrasound is routinely used in IVF cycles, yet interpretation remains subjective and operator dependent. AI has shown promise in reproductive medicine, particularly in embryo and oocyte assessment, by extracting subtle imaging features invisible to the human eye. However, AI-based approaches to assess endometrial receptivity from ultrasound images remain insufficiently explored. Study design, size, duration This retrospective study included 47 FETs from 46 women in a private fertility clinic (2025-2026). Transvaginal, sagittal, endometrium ultrasound images (captured on the day of exogenous progesterone administration or LH surge) plus clinical data (age and endometrial thickness) were analyzed using a novel Endometrium-AI tool. Most transfers involved good-quality (n = 44), euploid (n = 43) embryos. The primary outcome was CP, defined as the presence of a gestational sac and/or fetal heartbeat at 7 weeks’ gestation on ultrasound. Participants/materials, setting, methods The Endometrium-AI tool generates a 0-10 score, with higher scores indicating a greater predicted probability of CP. Scores were grouped into four categories (0-2.5, 2.6-5, 5.1-7.5, 7.6-10) to describe CP rates across increasing score intervals. Model performance was assessed using receiver operating characteristic (ROC) analysis and quantified by the area under the curve (AUC), sensitivity, and specificity. Continuous variables were compared using the Wilcoxon rank-sum test, and CP rates were using the chi-square test. Main results and the role of chance Mean patient age was 40±5.5 years (range 30–57). FETs included medicated (n = 36) and modified natural (n = 11) cycle protocols, with an endometrial thickness mean of 9.69±2.59 mm assessed at progesterone initiation or LH surge, accordingly. CP rate was 43% (n = 20/47). The Endometrium-AI model predicted CP with an AUC of 0.62, indicating modest discriminatory ability. CP rates increased progressively across Endometrium-AI score intervals: 35% (0–2.5), 28% (2.6–5), 54% (5.1–7.5), and 62% (7.6–10). ROC analysis identified a clinically relevant cut-off value of 5.75, yielding 50% sensitivity and 81% specificity. Women with Endometrium-AI scores ≥5.75 (n = 15) had significantly higher CP rates compared with those below this threshold (66% vs 31%, p = 0.002). Higher Endometrium-AI scores were significantly associated with younger maternal age at transfer and greater endometrial thickness. Women above the 5.75 score threshold were younger (37.33±3.03 vs 41.3±5.95; p = 0.004) and had thicker endometrium (10.9±1.90; vs 9.07±2.67; p < 0.01). No associations were observed with BMI (≥5.75: 24.67±2.27 vs < 5.75: 25.20±4.65; p > 0.05). Additionally, Endometrium-AI scores and CP rates were comparable in patients with versus without prior pregnancy (4.06 ± 2.58 vs 4.46 ± 2.70; 37.50% vs 47.82%) and prior live birth (4.37 ± 2.57 vs 4.42 ± 2.84; 44% vs 30%) (all, p > 0.05). Limitations, reasons for caution The small sample size available limits statistical power; additional data is being collected. The single-center study included predominantly good-quality, euploid embryos, which may limit generalizability. The Endometrium-AI model is still undergoing refinement and validation; further multicenter, prospective studies are needed to confirm its predictive value across diverse clinical settings. Wider implications of the findings These initial findings suggest AI-based analysis of transvaginal ultrasound images may capture unique endometrial features associated with CP potential. Although modest predictive performance, the model identified a subgroup with higher pregnancy likelihood. With further validation, Endometrium-AI could support more objective, standardized, and non-invasive endometrial assessment, complementing clinical decision-making for embryo-transfer. Trial registration number No

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.