Accurate medical image segmentation faces challenges from high expert annotation costs, limiting clinical adoption of fully supervised deep learning. Existing semi-supervised methods encounter persistent challenges in addressing boundary ambiguity issues that stem from gradual pseudo-label noise accumulation in single-dimensional frame-works. This paper presents a novel Cross-dimensional Co-Training Strategy with Soft-Hard Pseudo-label (CSHP), which addresses these challenges through dynamic bidirectional optimization between 2D and 3D network architectures. Our methodology employs a 2D framework to generate multiplanar slice predictions, subsequently integrated through a three-view consensus verification mechanism to produce hybrid pseudo-labels. For voxels demonstrating prediction consistency, hard labels rein-force core region learning, while probabilistic sharpening generates adaptive soft labels for boundary-disputed areas. Comprehensive validation on the LiTS benchmark demonstrates CSHP’s superior performance compared to state-of-the-art methods in segmentation accuracy, particularly in boundary definition, confirming its effectiveness for semi-supervised medical image analysis.