Abstract <p>Segmenting medical images plays a pivotal role in diagnosis, treatment planning, and healthcare. Recent advancements in deep learning have significantly transformed this domain. Convolutional Neural Networks (CNNs) are proficient at extracting local image features, while Vision Transformers (ViTs) excel in capturing long-range dependencies through self-attention mechanisms. However, both approaches face limitations when dealing with intricate anatomical structures, unclear lesion boundaries, and scale variations. Additionally, they often require substantial computational resources. To tackle these challenges, this study introduces CDM-UNet, designed to improve upon the limitations of existing models in feature representation and fusion. It also investigates the potential of hybrid Mamba models for advancing their application in this domain. CDM-UNet combines Mamba’s global feature modeling capabilities with lightweight attention mechanisms. The core component of CDM-UNet is Content Driven Mamba Block (CDMB), and a SCConv-Based Attention Gate (SAG) module is introduced to suppress the influence of irrelevant information. Comprehensive evaluations on the ISIC17, ISIC18, and Synapse datasets demonstrate that CDM-UNet outperforms existing mainstream models across multiple evaluation criteria, including Dice Similarity Coefficient (DSC) and mean Intersection over Union (mIoU), underscoring its excellent segmentation performance in medical image processing.</p> Graphical Abstract <p></p>