FCMamba: frequency-aware CNN-Mamba hybrid model for skin lesion segmentation

To address the challenge of balancing global context modeling and local detail capture in skin lesion segmentation, this paper proposes Frequency-Aware CNN-Mamba (FCMamba), a novel hybrid architecture that integrates spatial-frequency collaborative modeling for high-precision medical image segmentation. Existing methods, such as CNNs, lack global dependency modeling, while Transformers and pure Mamba models suffer from high computational complexity or local detail loss. FCMamba introduces a dual-branch encoder: a CNN branch extracts multi-scale local texture features, and a Mamba branch captures global structural information through selective state space modeling. Innovatively, discrete wavelet transform (DWT) decomposes features from both branches into high- and low-frequency components. Attention mechanisms selectively fuse high-frequency edge details and low-frequency semantic information, while cross-modal skip connections optimize feature interaction. Additionally, the Mixer Mamba Conv Block (MMCB) alternates convolutional and state space operations to enhance local-global fusion.Experiments on the ISIC2017, ISIC2018, and PH2 datasets demonstrate FCMamba's superiority, achieving mIoU scores of 81.07%, 81.73%, and 86.45%, and DSC scores of 89.54%, 89.94%, and 92.73%, respectively, outperforming CNN-, Transformer-, and Mamba-based baselines. Ablation studies validate the effectiveness of frequency-domain fusion and dual-branch design, highlighting FCMamba's robustness in handling complex boundaries and small lesions.