Attention-Mamba: A Mamba-Enhanced Multi-Scale Parallel Inference Network for Medical Image Segmentation
Researchers have introduced Attention-Mamba, a novel deep learning architecture designed to enhance medical image segmentation. This model addresses the limitations of traditional U-shaped networks and Transformers by integrating Mamba, a state space model that captures long-range dependencies with linear computational complexity. The network features a dual-path architecture with lateral connections to aggregate semantic and spatial details, alongside a Recursive Alignment Module (RAM) for restoring spatial details in low-resolution features. Parallel Mamba branches establish hierarchical global representations, while a Mamba-based attention mechanism facilitates adaptive multi-scale prediction fusion. Experimental results across MRI, CT, and dermoscopy modalities demonstrate superior generalization and segmentation performance on standard datasets like Synapse, ACDC, ISIC-2018, and PH2. Notably, the model achieves state-of-the-art accuracy while maintaining high efficiency, possessing the second-smallest parameter count (14.05 million) and moderate computational complexity (8.94 GFLOPs) compared to existing CNN, Transformer, and Mamba-based counterparts.
Wire timeline
Attention-Mamba: A Mamba-Enhanced Multi-Scale Parallel Inference Network for Medical Image Segmentation
Researchers have introduced Attention-Mamba, a novel deep learning architecture designed to enhance medical image segmentation. This model addresses the limitations of traditional U-shaped networks and Transformers by integrating Mamba, a state space model that captures long-range dependencies with linear computational complexity. The network features a dual-path architecture with lateral connections to aggregate semantic and spatial details, alongside a Recursive Alignment Module (RAM) for restoring spatial details in low-resolution features. Parallel Mamba branches establish hierarchical global representations, while a Mamba-based attention mechanism facilitates adaptive multi-scale prediction fusion. Experimental results across MRI, CT, and dermoscopy modalities demonstrate superior generalization and segmentation performance on standard datasets like Synapse, ACDC, ISIC-2018, and PH2. Notably, the model achieves state-of-the-art accuracy while maintaining high efficiency, possessing the second-smallest parameter count (14.05 million) and moderate computational complexity (8.94 GFLOPs) compared to existing CNN, Transformer, and Mamba-based counterparts.
cs.AI updates on arXiv.org