{"paper_id":"47ec1e78-aeac-4e3f-9006-801afb1933a7","body_text":"ECR 2025 / C-28020\nAdvancing adenomyosis detection through deep learning-assisted uterus segmentation in MRI\nCongress:\nECR 2025\nPoster Number:\nC-28020\nType:\nScientific Exhibit\nKeywords:\nArtificial Intelligence, MR, Neural networks, Segmentation, Outcomes\nAuthors:\nC. Tappermann, M. S. May, L. Siegler, M. Fenske, M. B. Bauer, T. Rüttinger, S. Arndt, B. Lassen-Schmidt\nDOI:\n10.26044/ecr2025/C-28020\nPurpose\nAdenomyosis is an under-researched gynecological condition where endometrial glands and stroma infiltrate the myometrium. It affects more than 20 % of women during their reproductive years and can cause chronic pelvic pain, social restrictions, and infertility [1, 2].The RACOON FADEN project aims to advance research on the early detection of adenomyosis from MRI in a multicenter study. Thirteen German university hospitals are participating in this study. Thickening of the junctional zone, the hypointense layer between endometrium and myometrium, is widely accepted as a surrogate for...\nMethods and materials\nThe RACOON FADEN study was approved by the ethics committee of all participating university hospitals. The study protocol complies with the declaration of Helsinki. Funding was provided by the Bundesministerium für Bildung und Forschung via Netzwerk Universitätsmedizin (NUM 2.0, FKZ: 01KX2121).We included a cohort of thirteen female patients and volunteers (27 ± 5.44 years, 21.8 ± 2.96 BMI) from three German university hospitals in our first MRI datasets from the FADEN study preparation to develop the initial models. This dataset will expand throughout the project...\nResults\nThe 2D u-net achieves an average DSC of 0.71 (MM), 0.62 (JZ) and 0.66 (EM) on the test data, improving slightly to 0.78 (MM), 0.64 (JZ) and 0.66 (EM) when considering only the slices up to the cervical junction. The 3D u-net differs slightly from the performance of the 2D u-net. This model achieves an average DSC of 0.52 (MM), 0.68 (JZ), and 0.73 (EM). Again, considering the slices up to the cervical junction, the DSC increases to 0.66 (MM), 0.70 (JZ), and 0.77 (EM)....\nConclusion\nThese results serve as an initial proof-of-concept for the methods employed in supporting clinicians with this project's segmentation task. While the current approach demonstrates feasibility, there is significant potential for improvement. By acquiring more data, we will iteratively improve the model quality through a continuous training loop, allowing for better generalization and robustness across diverse cases. Furthermore, we aim to extend these methods to additional sequences and explore the possibility of combining them into a single model. In parallel, we plan to incorporate more complex...\nPersonal information and conflict of interest\nC. Tappermann:\nNothing to disclose\nM. S. May:\nNothing to disclose\nL. Siegler:\nNothing to disclose\nM. Fenske:\nNothing to disclose\nM. B. Bauer:\nNothing to disclose\nT. Rüttinger:\nNothing to disclose\nS. Arndt:\nNothing to disclose\nB. Lassen-Schmidt:\nNothing to disclose\nReferences\n1. Agostinho, L., Cruz, R., Osório, F., Alves, J., Setúbal, A., & Guerra, A. (2017). MRI for adenomyosis: a pictorial review. Insights into Imaging,8(6), 549–556. doi:10.1007/s13244-017-0576-z2. Bazot, M., Cortez, A., Darai, E., Rouger, J., Chopier, J., Antoine, J.-M., & Uzan, S. (2001). Ultrasonography compared with magnetic resonance imaging for the diagnosis of adenomyosis: correlation with histopathology. Human Reproduction,16(11), 2427–2433. doi:10.1093/humrep/16.11.24273. Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press.4. Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical...","source_license":"CC0","license_restricted":false}