CAMS: Convolution and Attention-Free Mamba-based Cardiac Image Segmentation

Our Contributions

• To the best of our knowledge, we are the first to propose a convolution and self-attention-free Mambabased segmentation network, CAMS-Net.
• We propose a Linearly Interconnected Factorized Mamba (LIFM) block to reduce the trainable parameters of Mamba and improve its non-linearity. LIFM implements a weight-sharing strategy for different scanning directions, specifically for the two scanning direction strategies of vision Mamba [55], to reduce the computational complexity further whilst maintaining accuracy.
• We propose a Mamba Channel Aggregator (MCA) and Mamba Spatial Aggregator (MSA) and demonstrate how they can learn information along the channel and spatial dimensions of the features, respectively.
• Extensive experimental validation, including ablation studies, is conducted to showcase the efficacy of our proposed model. Our proposed CAMS-Net outperforms existing state-of-the-art segmentation models on the CMR and the Multi-Disease, Multi-View, and Multi-Center (M&Ms-2) segmentation datasets, including pure CNN, self-attention, and hybrid self-attention, as well as methods using the original Mamba-based architecture combined with CNNs.

Paper Available Here