CSA: A Channel-Separated Attention Module for Enhancing MRI Reconstruction.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Channel attention mechanisms enhance visual task performance, including Magnetic Resonance Imaging (MRI) reconstruction.
• Traditional channel attention methods involve dimensionality reduction and cross-channel interactions, which can negatively impact MRI reconstruction due to low correlation between adjacent channels.
• Some feature maps in MRI reconstruction contain low information, making cross-channel interactions less effective.

Purpose of the Study:

• To propose a novel Channel-Separated Attention (CSA) module specifically designed for MRI reconstruction networks.
• To address the limitations of existing channel attention mechanisms in MRI reconstruction.
• To improve reconstruction quality while reducing computational complexity and parameter count.

Main Methods:

• Developed the CSA module, which avoids channel compression for lossless information transmission.
• Utilized the Hadamard product for channel-specific weight generation, eliminating cross-channel interactions.
• Integrated the CSA module into an advanced MRI reconstruction network, replacing the original channel attention module.

Main Results:

• The CSA module achieved superior MRI reconstruction performance compared to the original channel attention module.
• CSA module demonstrated significantly fewer parameters, using only approximately 1.036% of the Squeeze-and-Excitation (SE) module's parameters.
• Comparative experiments showed competitive reconstruction outcomes with state-of-the-art channel attention modules.

Conclusions:

• The CSA module offers an optimal balance between computational complexity and reconstruction quality for MRI.
• CSA module efficiently and effectively enhances MRI reconstruction networks.
• The proposed CSA module presents a promising advancement for deep learning-based MRI reconstruction.