Convolutional Neural Networks (CNNs) excel at extracting local features, but due to their restricted receptive fields, they often struggle to capture large-scale global context. Transformers leverage self-attention mechanisms to facilitate global interactions, yet the computational cost of standard self-attention scales quadratically with image resolution. To overcome these limitations, we propose MFDA-UNet, which adopts a hybrid architecture of convolution and linear attention for synergistic feature processing. To fully leverage their respective strengths, we design the Mamba-inspired Frequency-Decoupled Attention (MFDA) block. Through frequency decoupling, this block utilizes convolutions to process high-frequency local information, while employing linear attention to model the long-range dependencies of low-frequency global information. To enhance the feature representation capability of linear attention, we construct the Mamba-Enhanced Linear Attention (MELA) block. Inspired by MILA, this block injects Positional Encoding to substitute the forget gate functionality of Mamba and integrates the Mamba block structure into the linear attention mechanism. This design effectively strengthens representational power, accomplishing long-range dependency modeling with highly efficient linear complexity. Furthermore, we introduce the Gated Cross-Scale Attention (GCSA) module to optimize traditional skip connections. It aggregates features via cross-scale linear attention and incorporates Mamba’s high-performance gating mechanism for adaptive feature filtering, achieving precise feature fusion and selection. We conducted extensive experiments on four multi-modal benchmarks: ISIC 2017, ISIC 2018, Synapse, and ACDC. MFDA-UNet achieved improvements in the DSC by 0.44%, 0.15%, 0.53%, and 0.84% across the respective datasets compared to the second-best models. By capturing local and global multi-scale semantics with relatively low computational overhead, MFDA-UNet provides an efficient and robust solution for medical image segmentation.
Deng et al. (Mon,) studied this question.