HFIST-Net: High-throughput fast iterative shrinkage thresholding network for accelerating MR image reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Compressed sensing (CS) accelerates magnetic resonance image (MRI) reconstruction from undersampled k-space data.
• Deeply unfolded networks (DUNs) offer faster reconstruction and improved image quality compared to traditional CS-MRI methods.

Purpose of the Study:

• To propose a novel deep learning network, High-Throughput Fast Iterative Shrinkage Thresholding Network (HFIST-Net), for accelerated MRI reconstruction.
• To combine model-based CS techniques with data-driven deep learning for efficient MR image reconstruction.

Main Methods:

• The conventional Fast Iterative Shrinkage Thresholding Algorithm (FISTA) is unfolded into a deep network (HFIST-Net).
• A multi-channel fusion mechanism is introduced to enhance information transmission between network stages.
• A Gaussian context transformer (GCT) block is proposed to improve the characterization capabilities of deep Convolutional Neural Networks (CNNs).

Main Results:

• HFIST-Net was validated on T1 and T2 brain MR images from the FastMRI dataset.
• The proposed method demonstrated superior qualitative and quantitative performance compared to state-of-the-art unfolded deep learning networks.
• HFIST-Net achieved more accurate reconstruction of MR image details from highly undersampled k-space data.

Conclusions:

• HFIST-Net effectively reconstructs detailed MR images from highly undersampled k-space data.
• The network maintains fast computational speed, making it suitable for accelerated MRI applications.
• This method represents a significant advancement in combining CS and deep learning for MRI reconstruction.