DPFF-net: dual-path feature fusion network with vision Mamba for skin lesion segmentation

Abstract Accurate segmentation of skin lesions is crucial for the early screening and treatment of skin cancer. In recent years, Convolutional Neural Networks (CNNs) and Transformer-based approaches have achieved remarkable progress in image segmentation tasks. However, convolutional neural networks are limited in their ability to model long-range dependencies, which limits their effectiveness in capturing fine-grained details within medical images, while Transformers are constrained by their quadratic computational complexity. Recent studies have demonstrated that State Space Models (SSMs), particularly those represented by the Mamba architecture, can efficiently model long-range interactions with linear computational complexity. Inspired by the Mamba architecture, this paper proposes a lightweight medical image segmentation network, DPFF-Net, which integrates Mamba into a dual-path framework. The core of this network is the Dual Path Feature Fusion Module (DPFFM), which achieve efficient extraction and fusion of local and global features through a convolutional branch and an SSM branch. The Parallel Vision Mamba (PVM) layer further enhances the modeling of spatial dependencies, while the Detail Aware Convolution Block (DACB) employs various differential convolution strategies to reinforce boundary feature extraction. In addition, the Attention-Integrated Feature Aggregation Module (AIFAM) combines attention mechanisms with feature aggregation to promote deep interaction between local and global representations. The results of comparison and ablation studies on the ISIC2017 and ISIC2018 datasets indicate that DPFF-Net exhibits strong competitiveness in the skin lesion segmentation task. Compared with UltraLight VM-UNet, on the ISIC2017 dataset, DPFF-Net improves mIoU and DSC by 2.61% and 1.63% respectively. On the ISIC2018 dataset, mIoU and DSC increase by 1.62% and 1.00% respectively.