Mamba-SaCIF Net: a lightweight hybrid model for cardiac MRI segmentation using spatial and channel fusion

In this study, we propose a hybrid architecture, Mamba SaCIF Net, for improving the segmentation of cardiac structures using MRI data. The architecture leverages a U-shaped encoder-decoder framework that integrates Mamba-based modules, Global Attention Module (GAM) and Channel Attention Modules (CAM), for enhanced channel and spatial information fusion. The input image is partitioned into non-overlapping patches and projected into a high-dimensional feature space using a linear embedding layer. The encoder employs hierarchical stages with CAM and GAM to capture both local and global dependencies, while the decoder utilizes skip connections and bilinear upsampling for precise reconstruction. The proposed Spatial and Channel Information Fusion (SaCIF) mechanism merges outputs from CAM and GAM using bidirectional aggregation and shared weights, ensuring computational efficiency and robust feature extraction. Experimental evaluations on benchmark cardiac MRI datasets demonstrate that our proposed model consistently outperforms state-of-the-art methods in segmentation accuracy, parameter efficiency, and computational complexity, making it a promising solution for clinical applications.