DHPCL: Diffusion-Driven Hard-Soft Prototype Contrastive Learning for Semi-Supervised Medical Image Segmentation.

Semi-supervised learning (SSL) has emerged as a powerful paradigm for medical image segmentation, effectively mitigating the reliance on large-scale pixel-level annotations. However, existing SSL approaches frequently suffer from a vicious cycle of insufficient pseudo-label quality and degraded feature discriminability, where noisy predictions induce poor feature representations and vice versa. To address this problem, we propose a novel framework, Diffusion-Driven Hard-Soft Prototype Contrastive Learning (DHPCL) for semi-supervised medical image segmentation. First, we design a Discriminator-Guided Diffusion Rectification (DGDR) module. Unlike existing methods that identify noise via static thresholds, DGDR leverages pixel-wise confidence maps from a discriminator to guide a latent diffusion model, adaptively performing stronger reconstruction on low-confidence regions while preserving reliable structures through a Local Detail Enhancement (LDE) mechanism. Second, to address feature degradation, we propose a Hard-Soft Harmonized Prototype Contrast (HSH-PC) module. This component constructs dynamic class prototypes from diffusion-refined labels and enforces feature alignment using a dual-constraint strategy: imposing hard constraints on high-confidence pixels and soft distribution-based guidance for ambiguous regions. This harmonized approach significantly improves intra-class compactness and inter-class separability. Extensive experiments on the ACDC, MS-CMRSeg, and LA datasets demonstrate that the proposed framework achieves superior segmentation accuracy and boundary precision compared to state-of-the-art semi-supervised methods.