Highly undersampled magnetic resonance imaging reconstruction using autoencoding priors.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning shows promise for accelerating Magnetic Resonance Imaging (MRI).
• Learning explicit priors and integrating them into observation constraints remains a challenge for fast MRI reconstruction.
• Denoising autoencoder (DAE) networks offer a potential solution for image restoration tasks.

Purpose of the Study:

• To leverage a denoising autoencoder (DAE) as an explicit prior for highly undersampled MRI reconstruction.
• To enhance the DAE prior by training in a multichannel scenario and applying it to single-channel reconstruction.
• To improve reconstruction accuracy using a 2-sigma rule for artificial noise generation in DAE.

Main Methods:

• Utilized a denoising autoencoder (DAE) network as an explicit prior for MRI reconstruction.
• Trained the DAE in a multichannel scenario and applied it to single-channel images via variable augmentation.
• Implemented a 2-sigma rule for noise generation to enhance DAE performance.
• Employed proximal gradient descent for the reconstruction algorithm.

Main Results:

• The enhanced autoencoding priors demonstrated superior performance in highly undersampled MRI reconstruction.
• Consistent improvements were observed across varying sampling trajectories and acceleration factors.
• Key metrics like peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and high-frequency error norm (HFEN) were enhanced.

Conclusions:

• A simple and effective method for incorporating DAE priors into undersampled MRI reconstruction was developed.
• The proposed method achieves superior performance compared to state-of-the-art techniques.
• The DAE prior is versatile and applicable to diverse sampling patterns and acceleration factors.