Semi-Supervised Medical Lesion Image Segmentation Based on a Contrast-Guided Diffusion Model

Medical lesion segmentation plays a crucial role in computer-aided diagnosis, yet acquiring fully annotated images remains a significant challenge. Semi-supervised learning has shown great potential in scenarios with limited labeled data. However, pseudo-labels, commonly used for unlabeled data, may adversely affect model performance due to their inherent inaccuracies. To address this issue, we propose a semi-supervised lesion segmentation framework based on a Contrast-Guided Diffusion Model (CGDM). To mitigate the impact of inaccurate pseudo-labels, we exploit the contrastive relationship between lesion and healthy images, restoring lesion regions to a healthy-like appearance. By directly incorporating this contrastive semantic information during training, we alleviate the model’s over-reliance on pseudo-labels and mitigate its detrimental effects on model performance. Furthermore, we introduce a Structural Similarity Contrast (SSC) loss function to balance supervised and unsupervised learning. This function constructs sample pairs for contrastive learning, maximizing the disparity between paired lesion and healthy images while minimizing the resemblance of lesion regions in unpaired lesion images. Experimental results on the BUSI, BraTS2018, and KiTS19 datasets demonstrate that CGDM achieves superior performance compared to state-of-the-art semi-supervised segmentation methods.