ReconFormer: Accelerated MRI Reconstruction Using Recurrent Transformer.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Accelerated magnetic resonance imaging (MRI) reconstruction is crucial for reducing scan times but presents an ill-posed inverse problem due to k-space under-sampling.
• Existing methods struggle to balance reconstruction fidelity with acceleration factors.

Purpose of the Study:

• To introduce ReconFormer, a novel recurrent Transformer model for high-fidelity MRI reconstruction from highly under-sampled k-space data.
• To demonstrate the model's effectiveness and parameter efficiency compared to state-of-the-art methods.

Main Methods:

• Proposing a recurrent Transformer architecture, ReconFormer, built upon Recurrent Pyramid Transformer Layers (RPTLs).
• Implementing Recurrent Scale-wise Attention (RSA) to leverage multi-scale information and recurrent state dependencies.
• Validating the model on diverse MRI datasets and sequences with up to 8x acceleration.

Main Results:

• ReconFormer achieves significant improvements in MRI reconstruction quality over existing methods.
• The model demonstrates superior parameter efficiency, with only 1.1 million trainable parameters.
• High-fidelity image reconstruction is achieved even with substantial k-space under-sampling.

Conclusions:

• ReconFormer offers a powerful and efficient solution for accelerated MRI reconstruction.
• The proposed Recurrent Scale-wise Attention mechanism is key to its performance.
• This approach holds promise for faster and more accessible MRI diagnostics.