<p>Current methods in medical image segmentation have achieved promising performance in automatic lesion delineation. However, substantial uncertainty and ambiguity often persist in the fine-grained boundary regions of lesions. To address this issue, we propose a novel prior-posterior framework based on diffusion models, which utilizes multi-expert annotations fused into 0–1 probabilistic segmentation masks to guide model training. Specifically, The Beta distribution is integrated into the prior and posterior networks of the diffusion model to explicitly capture boundary uncertainty present in both expert annotations and model predictions. This approach allows the model to represent diverse plausible boundary configurations more accurately, thereby mitigating the effects of annotation ambiguity and inter-observer variability. Experimental results demonstrate that the proposed method significantly outperforms state-of-the-art probabilistic segmentation networks in terms of both segmentation accuracy and uncertainty modeling at lesion boundaries, highlighting its potential for clinical applications.</p>