Endo-MedSAM: a promptable vision foundation model adaptation for uterus segmentation on pelvic MRI in endometriosis

Abstract

Background Endometriosis is a common gynecologic condition in which pelvic MRI plays an important role in diagnosis and preoperative assessment. AI-enabled automated uterus segmentation on pelvic MRI could support endometriosis care by enabling standardized volumetric measurements and quantitative imaging analyses. However, developing robust AI models for this task is challenging because endometriosis often distorts pelvic anatomy through adhesions, uterine displacement, and coexisting fibroids or adenomyosis. Although promptable medical vision foundation models, such as MedSAM-2, provide a promising framework for interactive segmentation, their zero-shot performance on endometriosis MRI remains limited. Objective To develop Endo-MedSAM, a uterus-focused adaptation of MedSAM-2 for pelvic MRI in endometriosis, and to systematically evaluate its segmentation performance across institutions and prompting strategies. Methods We used a pelvic MRI dataset comprising 74 subjects and 3,449 T2-weighted slices from two institutions (D1, multicenter; D2, single-center). Endo-MedSAM was initialized with MedSAM-2 weights and fine-tuned by training the prompt encoder and mask decoder while unfreezing the final layers of the image encoder. Slice-wise predictions were then reconstructed into 3D volumes for volumetric evaluation. Performance was assessed using slice-level and 3D Dice scores and the 95th percentile Hausdorff distance (HD95). Results Endo-MedSAM was evaluated in three configurations: training on the single-center D2 cohort and testing on the multicenter D1 cohort, training on D1 and testing on D2, and using a mixed D1 + D2 split with 75% of patients for training and 25% for testing. Across these settings, Endo-MedSAM achieved mean 3D Dice scores of 0.81-0.88 with bounding-box prompts and 0.68-0.76 with 1-click and 2-click point prompts. Compared with zero-shot MedSAM-2, this represented absolute 3D Dice improvements of approximately 0.27-0.34 for bounding-box prompting. Endo-MedSAM also showed markedly better performance across all prompting modes, with the largest gains observed for bounding-box prompting.