Human-in-the-Loop Semi-Supervised Uterine Cervix Ultrasound Image Segmentation

Medical image segmentation is today crucial for disease diagnosis and treatment planning, yet obtaining high-quality annotated data remains a significant challenge. In this paper, we propose a semi-supervised learning framework leveraging U-Net as the baseline model. The dataset consists of both labeled and unlabeled images, where the labeled subset is used for initial supervised training. To improve generalization, cross-validation is employed, and model ensembling is utilized to enhance segmentation performance. Our approach incorporates a human-in-the-loop scheme to refine the dataset and improve model performance iteratively. Predicted masks for unlabeled data are manually corrected to remove prominent errors, and these refined masks are progressively integrated into subsequent training iterations, leading to improved segmentation accuracy. Evaluated on the public FUGC 2025 leaderboard, our method achieves a Dice similarity of 90.08% and a Hausdorff distance of 40.39, demon-strating the effectiveness of iterative pseudo-label refinement in ultrasound image segmentation.