L26/P-823 Automated AI-assisted preantral ovarian follicle staging in a mouse model of chemotherapy-induced primary ovarian insufficiency
Human Reproduction · 01.07.2026 · 41 (Supplement_1)
DOI: 10.1093/humrep/deag083.1150 ↗Kurz gefasst
Fachartikel in der Zeitschrift Human Reproduction, online erschienen am 08.07.2026. Verfasst von S.E. Pors und 6 weiteren Autor:innen.
Themen: Onkologische Pflege
Abstract (Originalsprache)
Study question Can an automated deep learning-based pipeline (GHOST) accurately and objectively stage preantral follicles in a mouse model of chemotherapy-induced primary ovarian insufficiency? Summary answer GHOST showed strong agreement with expert histopathologists and enables accurate, unbiased, and automated follicle staging in mouse ovaries. What is known already Manual histopathological scoring of ovarian follicles is widely used in preclinical reproductive research but is labor-intensive and susceptible to inter- and intra-observer variability. This variability can reduce the reproducibility and robustness of study outcomes, especially in models assessing ovarian toxicity or reproductive disorders. Automated image analysis and deep learning methods have shown potential to improve objectivity and throughput, but validated tools for detailed preantral follicle staging remain limited. Reliable, automated systems could greatly enhance the consistency and scalability of follicle classification in research relying on large histology datasets. Study design, size, duration Female C57BL/6JRj mice received cyclophosphamide on days 1 and 8 to induce primary ovarian insufficiency. Ovaries were collected on days 11, 18, and 30, fixed, sectioned, and stained for p63. The study developed and validated an automated pipeline combining U-Net–based oocyte segmentation and an InceptionV3 classifier for follicle staging. Automated classifications were compared with expert histopathologist scoring to assess agreement. Participants/materials, setting, methods Digitally scanned ovarian sections were used to annotate p63-positive oocytes with bounding boxes supported by SAM, generating training data for a U-Net segmentation model. Follicle-centered tiles were then extracted and manually assigned to five categories: primordial, transitional, primary, secondary, or false positive. These tiles were used to train an InceptionV3 follicle-staging classifier. Model performance was evaluated on a held-out test set against expert manual scoring. Main results and the role of chance A total of 500 annotated tiles were used to train the oocyte segmentation model, and 1000 tiles were used for follicle staging. The U-Net successfully identified p63-positive oocytes across the dataset, enabling reliable extraction of follicle-centered tiles. The InceptionV3 classifier assigned follicles to the four biological stages—primordial, transitional, primary, and secondary—while also identifying false-positive detections. In the held-out test set, automated follicle staging demonstrated good overall agreement with expert histopathologist scoring, achieving a Cohen’s kappa of 0.61. Agreement varied across categories, with the highest concordance observed for primary and secondary follicles. These stages benefited from clearer morphological features, facilitating stronger classifier performance. Primordial and transitional follicles showed moderate agreement, reflecting their subtler histological distinctions. Overall, the pipeline provided rapid, consistent follicle staging with reduced susceptibility to observer-dependent scoring variability. The approach demonstrated strong potential for scalable, objective analysis in studies requiring high-throughput evaluation of ovarian histology. Limitations, reasons for caution The study relied on limited training data and a single staining marker, which may restrict generalizability. Performance varied across follicle stages, particularly for early-stage follicles with subtle morphology. Additional datasets across strains, stains, and conditions are needed for broader validation. Wider implications of the findings GHOST offers a robust, unbiased tool for automated follicle staging, improving reproducibility in ovarian research and toxicology studies. Its validated performance supports broader adoption of AI-driven histopathology to enhance throughput and reduce subjective variability in preclinical reproductive models. Trial registration number No
Zitieren
APA
Pors, S., Riis, M., Ostenfeldt, D., Thomsen, A., Skytte, J., Butt, A., & Zachariassen, L. (2026). L26/P-823 Automated AI-assisted preantral ovarian follicle staging in a mouse model of chemotherapy-induced primary ovarian insufficiency. Human Reproduction, 41(Supplement_1). https://doi.org/10.1093/humrep/deag083.1150
Vancouver
Pors S, Riis M, Ostenfeldt D, Thomsen A, Skytte J, Butt A, et al. L26/P-823 Automated AI-assisted preantral ovarian follicle staging in a mouse model of chemotherapy-induced primary ovarian insufficiency. Human Reproduction. 2026;41(Supplement_1). doi:10.1093/humrep/deag083.1150
BibTeX
@article{Pors2026,
author = {Pors, S.E. and Riis, M.L. and Ostenfeldt, D.A. and Thomsen, A.H.R. and Skytte, J.L. and Butt, A. and Zachariassen, L.F.},
title = {L26/P-823 Automated AI-assisted preantral ovarian follicle staging in a mouse model of chemotherapy-induced primary ovarian insufficiency},
journal = {Human Reproduction},
year = {2026},
volume = {41},
number = {Supplement_1},
doi = {10.1093/humrep/deag083.1150}
}
EUVIMED