L26/P-894 How consistent is artificial intelligence-based three-dimensional follicle measurement during ovarian stimulation?
Human Reproduction · 01.07.2026 · 41 (Supplement_1)
DOI: 10.1093/humrep/deag083.1218 ↗Kurz gefasst
Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von E.E. Lara Molina und 7 weiteren Autor:innen.
Abstract (Originalsprache)
Study question What is the intra- and interobserver agreement of artificial intelligence-based relaxed follicle diameters from repeated three-dimensional ovarian ultrasound volume acquisitions during ovarian stimulation monitoring? Summary answer Artificial intelligence-based relaxed follicle diameters showed high and consistent within- and between-observer agreement across repeated three-dimensional ultrasound volume acquisitions during ovarian stimulation monitoring. What is known already Follicle size assessment is essential for ovarian stimulation monitoring, yet measurements can vary due to probe positioning, plane selection, calliper placement, and differences in acquisition and interpretation. Although three-dimensional (3D) ultrasound enables volume-based assessment and artificial intelligence (AI)-based segmentation can automatically identify follicles within ovarian volumes to derive a software-generated relaxed follicle diameter, reproducibility should be characterised under routine clinical conditions before broader implementation, including repeated acquisitions by the same observer and by different observers. Study design, size, duration This prospective multicentre reliability study was conducted in three IVF clinics. Thirty participants contributed 67 ovarian stimulation monitoring visits, during which a trained physician and a trained nurse at each centre acquired two 3D ovarian volumes per visit, thereby enabling both intra- and interobserver comparisons based on repeated volume acquisitions. Statistical analyses were performed after completion of all study visits. Participants/materials, setting, methods Participants were egg donors undergoing routine ovarian stimulation monitoring who provided written informed consent. At each visit, both trained observers (physician and nurse) acquired two 3D ovarian ultrasound volumes (P1/P2; N1/N2), which were then processed with an AI-based online segmentation system that identified follicles (≥2 mm) and computed relaxed follicle diameters. Agreement was evaluated by combining complementary approaches, including Spearman correlation, intraclass correlation coefficients (ICC; 95% CI), median absolute differences (IQR) and Bland–Altman plots. Main results and the role of chance Across 67 visits, 1,694 follicles were automatically segmented and contributed to the descriptive analyses, while agreement metrics were calculated for follicles detected in both paired acquisitions. Mean (SD) relaxed diameter across all segmented follicles (≥2 mm) was 9.06 (3.92) mm for P1, 9.07 (3.94) for P2, 9.16 (4.01) for N1 and 9.11 (3.94) for N2. Missing detections were uncommon, ranging from 6.3% to 8.3% across acquisitions (107–141 follicles). For all intra- and inter-observer comparisons, Spearman correlations were consistently high (0.95–0.96; all p < 0.001). Intra-observer agreement was robust for both observers, with ICC 0.955 (95% CI 0.951–0.960) and a median absolute difference of 0.42 mm (IQR 0.19–0.86) for P1 vs P2, and ICC 0.951 (95% CI 0.946–0.956) with 0.45 mm (IQR 0.18–1.04) for N1 vs N2. Interobserver agreement was similarly strong (P1 vs N1 ICC 0.945, 0.939–0.950; 0.52 mm, IQR 0.21–1.05; P2 vs N2 ICC 0.958, 0.954–0.962; 0.44 mm, IQR 0.18–0.94), and Bland–Altman analyses suggested no systematic observer-related bias. Limitations, reasons for caution Follicle detections were occasionally missing in paired acquisitions (≤8.3%), reducing matchable follicles, although Bland–Altman analyses suggested no systematic bias. Accuracy against an independent reference standard was not assessed, and findings may not generalise beyond trained users or clinics. Wider implications of the findings Automated 3D-based relaxed follicle diameters showed low variability and strong observer agreement, enabling reproducible follicle assessment during ovarian stimulation monitoring. This AI-based approach may support standardisation, multicentre quality assurance and more consistent clinical decisions. Implementation studies should assess effects on workflow, costs and patient outcomes. Trial registration number No
Zitieren
APA
Molina, E. L., Garrigós, V., Romero, J., Coy, A., Ballesteros, A., Royo, P., Alepuz, M., & Garcia-Velasco, J. (2026). L26/P-894 How consistent is artificial intelligence-based three-dimensional follicle measurement during ovarian stimulation?. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.1218
Vancouver
Molina EL, Garrigós V, Romero J, Coy A, Ballesteros A, Royo P, et al. L26/P-894 How consistent is artificial intelligence-based three-dimensional follicle measurement during ovarian stimulation?. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.1218
BibTeX
@article{Molina2026,
author = {Molina, E.E. Lara and Garrigós, V. and Romero, J.E. and Coy, A. and Ballesteros, A. and Royo, P. and Alepuz, M. and Garcia-Velasco, J.A.},
title = {L26/P-894 How consistent is artificial intelligence-based three-dimensional follicle measurement during ovarian stimulation?},
journal = {Human Reproduction},
year = {2026},
volume = {41},
number = {Supplement_1},
doi = {10.1093/humrep/deag083.1218}
}
EUVIMED