Dual-domain faster Fourier convolution based network for MR image reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning accelerates Magnetic Resonance Imaging (MRI) by reconstructing images from undersampled k-space data.
• Existing methods face limitations in receptive field size, data consistency rigidity, and refinement structures, hindering performance.

Purpose of the Study:

• To develop an advanced deep learning framework for enhanced MRI reconstruction quality and speed.
• To address limitations of current dual-domain reconstruction networks.

Main Methods:

• Introduced Faster Inverse Fourier Convolution (FasterIFC) to expand receptive fields in k-space domain networks.
• Developed a novel softer Data Consistency (softerDC) layer for flexible data consistency strategies.
• Proposed the Dual-Domain Faster Fourier Convolution Based Network (D2F2) with a parallel structure leveraging FasterIFC and softerDC.

Main Results:

• D2F2 demonstrated superior performance on the NYU fastMRI dataset across multiple acceleration factors.
• The framework achieved significant improvements in both quantitative and qualitative MRI reconstruction evaluations.
• FasterIFC effectively utilized long-range information in k-space data.

Conclusions:

• The proposed D2F2 framework significantly enhances MRI reconstruction quality and efficiency.
• The FasterIFC operator and softerDC layer are key innovations for overcoming limitations in deep learning-based MRI.
• This approach represents a substantial advancement in fast MRI techniques.