Xu Jiashu, Lan Yihua, Zhang Yingqi, Zhang Chi, Stirenko Sergii, Li Huizhong
School of Artificial Intelligence and Software Engineering, Nanyang Normal University, Nanyang, 473001, China.
Collaborative Innovation Center of Intelligent Explosion-proof Equipment, Nanyang, 473001, Henan Province, China.
Sci Rep. 2025 Jul 1;15(1):21357. doi: 10.1038/s41598-025-06462-3.
Recent advances in state space models (SSMs) have demonstrated remarkable efficiency in modeling long-range dependencies, yet their application to 3D medical image segmentation remains underexplored. This paper introduces CDA-Mamba (Cross-Directional Attention Mamba), a novel hybrid architecture that combines the efficiency of SSMs with the strengths of convolutional and attention mechanisms to address the unique challenges of 3D medical image segmentation. CDA-Mamba features three key innovations: a Multi-Frequency Gated Convolution (MFGC) module to enhance spatial and frequency-domain feature integration, a Tri-Directional Mamba module to capture volumetric dependencies across orthogonal dimensions, and Selective Self-Attention integration in high-semantic layers to balance computational efficiency with global context modeling. Comprehensive experiments on the BraTS2023 brain tumor segmentation dataset highlight the competitive performance of CDA-Mamba, which achieves an average Dice score of 91.44. Moreover, evaluations on the AIIB2023 airway segmentation dataset further validate its effectiveness, with CDA-Mamba attaining the highest IoU of 88.72 and a DLR of 71.01. These results underscore its ability to balance accuracy and efficiency in 3D medical image segmentation.
状态空间模型（SSMs）的最新进展在对长程依赖关系进行建模方面展现出了卓越的效率，但其在3D医学图像分割中的应用仍未得到充分探索。本文介绍了CDA-Mamba（交叉方向注意力曼巴），这是一种新颖的混合架构，它将状态空间模型的效率与卷积和注意力机制的优势相结合，以应对3D医学图像分割的独特挑战。CDA-Mamba具有三项关键创新：一个多频门控卷积（MFGC）模块，用于增强空间和频域特征整合；一个三向曼巴模块，用于捕捉正交维度上的体素依赖关系；以及在高语义层中的选择性自注意力整合，以在计算效率和全局上下文建模之间取得平衡。在BraTS2023脑肿瘤分割数据集上进行的综合实验突出了CDA-Mamba的竞争性能，其平均Dice分数达到了91.44。此外，在AIIB2023气道分割数据集上的评估进一步验证了其有效性，CDA-Mamba的最高交并比（IoU）为88.72，Dice损失率（DLR）为71.01。这些结果强调了它在3D医学图像分割中平衡准确性和效率的能力。
