Medical image segmentation is essential for clinical diagnosis and treatment; however, existing methods still suffer from insufficient robustness in cross-modal and other complex tasks. In recent years, diffusion models have attracted attention for their stability and strong generative capabilities, and have been increasingly applied to segmentation tasks. Nevertheless, most studies are confined to single scenarios and lack systematic validation. This paper systematically evaluates several conditional diffusion-based segmentation methods and conducts comparative experiments on multiple public datasets. Experimental results show that DiffAtlas achieves the best performance on core metrics, IoU, Dice, and SEN, and, in particular, demonstrates stronger stability and generalization in few-shot and cross-domain tasks. EIDiffSeg attains performance on Dice and SEN close to that of DiffAtlas, while EnsemDiff stands out in PPV, although its overall robustness is slightly lower. These findings validate the effectiveness of diffusion models for medical image segmentation and highlight the advantages of a generative framework that jointly models images and masks in complex medical scenarios, providing valuable references for clinical applications and future research.