Dual-Mamba: A Hybrid CNN-Mamba Architecture for Tumor Segmentation from 3D Medical Images

Accurate tumor segmentation in 3D medical images is essential for early cancer diagnosis and treatment. Despite the significant potential of deep learning-based segmentation methods, the morphological diversity and low image contrast of tumor regions limit their performance. Recent studies have sought to improve model representation capability by combining Convolutional Neural Networks (CNNs) with Mamba structures. However, existing Mamba-based methods suffer from over-discretization of features, which weakens the modeling of local dependencies and overlooks cross-channel feature relationships, thus constraining segmentation accuracy. To address this issue, we propose a novel CNN-Mamba hybrid architecture for 3D tumor segmentation, named Dual-Mamba. Specifically, Dual-Mamba introduces a Multi-scale Mamba (MSM) module to model multi-scale local-global context dependencies and a Spatial-Channel Mamba (SCM) module to enhance cross-channel feature interactions. We validate the effectiveness of our method on multiple datasets, and the experimental results demonstrate that Dual-Mamba outperforms existing methods, achieving state-of-the-art performance and exhibiting strong generalization capabilities. The source code is available at: https://github.com/wangbingxun02/Dual-Mamba.