<p>Using deep learning models to accurately segment lesions is crucial for developing effective treatment plans. However, segmentation models based on CNNs and Transformers often produce overly-smoothed results due to their inherent deterministic mapping paradigm, and the application of standard diffusion models in segmentation is often hindered by information loss in the conditional guidance. To address these challenges, we introduce DiffMg-Seg, a novel conditional diffusion framework for medical image segmentation. DiffMg-Seg enhances the guidance and generation processes by first employing the Conditional Prior Feature Perceptron (CPFP) to enrich the conditioning network’s ability to perceive low-level spatial details for more robust guidance. It then leverages the Dynamic Alignment Module (DAM) to dynamically align conditional features with the evolving mask features during the denoising process, ensuring a more precise and context-aware synthesis. We conducted extensive experiments on three challenging datasets: a private Intracerebral Hemorrhage (ICH) dataset, the public ISIC16 skin lesion dataset, and the SpineSagT2W dataset. The results demonstrate that DiffMg-Seg achieved IoU values of 87.72(±4.70)%, 86.08(±7.54)% and 86.74(±7.78)% on the three datasets, respectively. It significantly outperforms state-of-the-art methods across all evaluation metrics, proving its effectiveness and robustness across different medical imaging modalities. This work presents a promising new direction for generative-based medical image segmentation, showing its significant potential to enhance diagnostic accuracy in various clinical applications.</p>