Abstract: Medical image segmentation based on semi-supervised learning has attracted extensive attention because of the high cost and scarcity of annotated images. Effectively leveraging unlabeled data remains a challenging task. This paper proposes a two-stage segmentation model based on a multi-scale feature fusion network to make use of unlabeled data, and address the empirical distribution mismatch between labeled and unlabeled data. The model uses labeled data to train a teacher model in the first stage and both labeled and unlabeled data are used to co-train a student model in the second stage. To improve the robustness of the teacher model, a copy-paste strategy is employed to increase data diversity. To alleviate the misguidance problem caused by the pseudo-labels generated in the second stage, confidence learning based on an assumption of classified noised process is introduced, thereby reducing the potential bias caused by pseudo-labels. Extensive experiments and ablation studies on two publicly available organ datasets demonstrate that the proposed model achieves high-precision segmentation.

Key words: semi-supervised learning, organ segmentation, multi-scale feature fusion