Bidirectional cross-view learning with dynamic region selection for semi-supervised medical image segmentation

Semi-supervised medical image segmentation aims to alleviate the heavy reliance on dense annotations while preserving high segmentation accuracy, yet it remains challenging due to unreliable pseudo-labels and insufficient utilization of unlabeled data. In this work, we propose Bidirectional cross-view learning with dynamic region selection, a unified semi-supervised framework instantiated as cross-view dynamic network (CVDyn-Net), to effectively exploit complementary information from labeled and unlabeled samples. CVDyn-Net introduces a foreground–background cross-view modeling strategy, where foreground semantics and background context are jointly learned and mutually constrained through bidirectional consistency. This design enables more accurate boundary delineation and robust representation learning, especially under sparse supervision. Furthermore, we propose a dynamic region mixing pseudo-labeling strategy, which progressively selects high-confidence regions and performs bidirectional copy-paste operations between labeled and unlabeled data. By dynamically adjusting the region selection process according to prediction reliability, the proposed method expands the perturbation space while suppressing noisy supervision during early training stages. Extensive experiments on the ACDC and PROMISE12 benchmarks under multiple annotation ratios demonstrate that CVDyn-Net consistently outperforms state-of-the-art semi-supervised segmentation methods. With only 20% labeled data, CVDyn-Net achieves a DSC of 90.54% on ACDC and 83.74% on PROMISE12, closely approaching fully supervised performance. Notably, on PROMISE12, the proposed method attains a DSC of 84.51% with 30% labeled data, surpassing the corresponding fully supervised baseline. These results validate the effectiveness of bidirectional cross-view learning and dynamic region mixing for annotation-efficient medical image segmentation. The code will be public at: https://github.com/CVDyn-Net .