<p>Medical image segmentation is a critical task in the field of computer vision, aimed at segmenting specific organs or lesions from input medical images. In current research, multiple deep learning architectures including Convolutional Neural Networks (CNNs) and Transformers are widely employed to tackle this challenge. Nevertheless, CNN-based methods remain constrained in establishing long-range dependencies, whereas Transformer architectures, despite demonstrating superior modeling capacities, are hindered by quadratic computational complexity. To overcome these dual limitations, this study introduces MSS-UNet, a novel architecture integrating State-Space Models (SSMs) that achieves linear computational scaling while maintaining robust long-range modeling capabilities, thereby effectively mitigating the respective shortcomings of CNN and Transformer approaches. We use the U-Net architecture as the overall framework and stack four MSS blocks in the encoder part to capture broad contextual information. The proposed Dual Cross-Attention (DCA) is a newly designed attention module tailored for skip-connection fusion. Unlike conventional cross-attention, DCA performs channel-wise and spatial-wise interactions in parallel, enabling more discriminative multi-scale aggregation. In the skip connection part, we introduce the DCA module to enhance the fusion of low-level and high-level features. Experimental results on the ISIC 2017 and ISIC 2018 datasets demonstrate that MSS-UNet achieves a 1.5–1.7% increase in mIoU and a 2.1–2.2% improvement in DSC over state-of-the-art Mamba-based UNet variants. These gains verify that our multi-directional selective scanning and Dual Cross-Attention effectively enhance long-range dependency modeling and multi-scale fusion. The proposed framework offers a promising and computationally efficient solution for real-world dermatological image segmentation. The codes are available at <a href="https://github.com/mmm587/MSS-UNet">https://github.com/mmm587/MSS-UNet</a></p>