LFA-Mamba: Local Feature-Augmented Medical Image Segmentation Framework with Mamba

Abstract:
Medical image segmentation, as a pivotal task in medical image analysis, plays an important role in improving clinical prognosis through precise delineation of lesion regions. While current approaches based on Convolutional Neural Networks (CNN) and Vision Transformers (ViT) have demonstrated remarkable progress, they exhibit inherent limitations: CNN excel at local feature modeling but lack global perceptual capabilities, whereas ViT capture global context at the expense of high computational complexity. To resolve this dichotomy, researchers introduces the Mamba architecture based on Spatial State Space Models (SSMs), which achieves an effective balance between global modeling and computational efficiency. Building upon the Mamba, we developed LFA-Mamba, a U-shaped segmentation network. To enhance local feature extraction, a novel Multi-channel Attention (MCA) Decoder Block was designed, whose core Largescale Channel Attention (LCA) module strengthens feature representation through reinforced cross-channel interactions. Experiments conducted on three datasets - ISIC2018 (melanoma), KvasirSEG (polyps), and a private Early Gastric Cancer (EGC) dataset - demonstrate that our model significantly outperforms existing state-of-the-art models.