MoUNets: Static-Weighted Ensemble of UNet-Based Experts for Levator Hiatus Segmentation in Pelvic Floor Ultrasound.
Authors
Affiliations (5)
Affiliations (5)
- Department of Ultrasound in Medicine, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, China.
- Department of Artificial Intelligence, Cainiao Network, Alibaba Group, Hangzhou, China.
- Department of Ultrasound in Medicine, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, China; Pelvic Medical Center, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, China.
- Department of Obstetrics and Gynecology, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, China.
- Department of Ultrasound in Medicine, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, China; Pelvic Medical Center, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, China. Electronic address: [email protected].
Abstract
This study aimed to investigate whether a static weighted ensemble of multiple UNet-based expert models (MoUNets) can achieve more accurate segmentation of the levator hiatus in pelvic floor ultrasound images than any single expert model. This retrospective single-center study included women diagnosed with pelvic organ prolapse who underwent transperineal ultrasound examinations between July 2020 and November 2023. The cohort was randomly divided into the training (n = 106), validation (n = 37) and testing (n = 43) cohorts. Model performance was evaluated on the test set using the Dice similarity coefficient (DSC) and Jaccard index (JI) and compared across all competing architectures. The study included 186 women (median age, 33.5 y [IQR, 30.0-48.8 y]). Comparative experiments on our in-house levator hiatus dataset confirmed that MoUNets is effective. Among the six conventional UNet models, the standard UNet achieved the best segmentation consistency. MoUNets, which combines different expert models of similar performance, showed higher numerical values than individual expert models. The best ensemble, MoUNets<sub>UNet&ACCUNet</sub>, yielded a DSC of 0.9653 and a JI of 0.9333. However, paired t-tests revealed no statistically significant difference between the ensemble and the UNet (p = 0.888). In contrast, statistically significant differences were observed between the best ensemble model and each of the other five individual models (all p < 0.01). The ensemble model leverages complementary advantages among different experts to significantly improve the segmentation performance of weaker models, particularly in difficult cases, demonstrating clear clinical value.