Frequency Learning via Multi-Scale Fourier Transformer for MRI Reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance Imaging (MRI) acquisition is time-consuming.
• Existing acceleration methods often neglect crucial frequency and non-local information, leading to poor image structure.
• There is a need for advanced reconstruction techniques that preserve image quality during accelerated MRI.

Purpose of the Study:

• To propose a novel deep learning framework, FMTNet, for accelerated MRI reconstruction.
• To focus on repairing both low-frequency and high-frequency information for enhanced image clarity.
• To develop an efficient Transformer module capable of learning global and multi-scale information.

Main Methods:

• Frequency Learning via Multi-scale Fourier Transformer for MRI Reconstruction (FMTNet) framework.
• Dual-branch architecture: High-Frequency Learning Branch (HFLB) and Low-Frequency Learning Branch (LFLB).
• Multi-scale Fourier Transformer (MFT) module utilizing Fourier convolution for efficient global information learning and cross-scale fusion.

Main Results:

• FMTNet demonstrates superior performance compared to state-of-the-art methods in MRI reconstruction.
• Experiments conducted under various acceleration rates and sampling patterns validate the method's effectiveness.
• The proposed MFT module efficiently learns non-local information with reduced computational resources.

Conclusions:

• FMTNet successfully reconstructs MRI images with clear structures by effectively repairing frequency and non-local information.
• The Multi-scale Fourier Transformer (MFT) offers an efficient alternative to standard self-attention for learning global image features.
• The proposed method represents a significant advancement in accelerated MRI reconstruction, improving both speed and image quality.