Skin lesion segmentation accuracy is often compromised by inadequate multi-scale feature representation and the loss of fine-grained details. Our solution, the Multi-Scale Dynamic Fusion Network (MSDF-Net), specifically targets these limitations through novel architectural innovations. The network is built upon two core modules: (1) the Multi-Scale Feature Fusion Module (MFFM), an extension of the Squeeze-and-Excitation network (SENet) that extracts multi-scale features through multi-branch convolution and sub-channel Sigmoid weighting to enhance feature representation. In the excitation stage, the second fully connected layer of SENet is replaced with a 1 \(\times \) 1 convolution, preserving parameter efficiency while enabling spatially aware feature recalibration. (2) the Dynamic Pathway Attention Module (DPAM), an extension of Convolutional Block Attention Module (CBAM), which adaptively balances the contribution weights of channel-first and space-first attention branches via a learnable gating mechanism instead of selecting a single path. This architecture enhances the discriminability of lesion regions while effectively suppressing background noise. The effectiveness of MSDF-Net has been validated on the ISIC2017 and ISIC2018 datasets.