Feature Fusion for Multi-Coil Compressed MR Image Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Magnetic resonance imaging (MRI) is a crucial non-invasive, radiation-free diagnostic tool.
• A major limitation of MRI is its long data acquisition time, hindering widespread clinical use.
• Deep learning (DL) methods show promise for accelerating MR image reconstruction but often overlook multi-coil correlations.

Purpose of the Study:

• To develop a novel deep learning method for efficient and high-quality undersampled MR image reconstruction.
• To address the underexploitation of inter-coil correlations in multi-coil MRI data.
• To improve the speed and image quality of MR image reconstruction.

Main Methods:

• Proposed the Multi-coil Feature Fusion Variation Network (MFFVN) for MR image reconstruction.
• Implemented an encoder for direct multi-coil feature extraction and a feature fusion operation.
• Utilized coil reshaping to enable a 2D network to process multi-coil data without significant parameter increase, preserving inter-coil information.

Main Results:

• MFFVN demonstrated improved average PSNR (0.2622 dB) and SSIM (0.0021 dB) compared to the Variation Network (VN).
• The method effectively leverages and combines multi-coil information through integrated feature extraction and fusion.
• MFFVN outperformed state-of-the-art methods on the fastMRI multi-coil brain dataset with a fourfold acceleration factor.

Conclusions:

• The MFFVN method enhances MR image reconstruction quality by effectively utilizing multi-coil information.
• The proposed network achieves superior performance without substantial computational overhead.
• MFFVN represents a significant advancement in accelerating MR image acquisition while maintaining high image fidelity.