A domain-agnostic MR reconstruction framework using a randomly weighted neural network.

Area of Science:

• Medical Imaging
• Machine Learning
• Neural Networks

Background:

• Magnetic Resonance Imaging (MRI) reconstruction often requires large datasets for training deep learning models.
• Current state-of-the-art methods necessitate extensive in-vivo data and ground truth for accurate MR image reconstruction.

Approach:

• A Weight Agnostic randomly weighted Network (WAN-MRI) is proposed for MR image reconstruction from undersampled k-space data.
• The network architecture comprises dimensionality reduction, a reshaping layer, and upsampling layers, without weight updates.
• The methodology is validated on fastMRI knee and brain datasets, demonstrating domain-agnostic capabilities.

Key Points:

• WAN-MRI achieves performance comparable to deep learning techniques while requiring minimal training data (20 samples) and no ground truth.
• The method shows significant improvements in SSIM and RMSE scores at R=4 and R=8 undersampling factors.
• Qualitative analysis reveals WAN-MRI captures clinically relevant details missed by classical methods like GRAPPA and SENSE.

Conclusions:

• WAN-MRI is a domain-agnostic algorithm capable of reconstructing MR images across different modalities and body organs.
• The method demonstrates superior generalization to out-of-distribution examples and achieves excellent SSIM, PSNR, and RMSE metrics.
• This approach eliminates the need for ground truth data and extensive training, offering a more efficient solution for MR image reconstruction.