A Shepard wide residual network for fast MR image reconstruction on Undersampled k-space data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Undersampled k-space data in Magnetic Resonance Imaging (MRI) accelerates scans but often leads to artifacts and long reconstruction times.
• Existing MRI reconstruction techniques struggle with image artifacts, overfitting, and limited generalization, posing challenges for clinical application.

Purpose of the Study:

• To introduce a novel technique, the Shepard Wide Residual Network (ShWideResNet), for efficient and high-quality MRI image reconstruction from undersampled k-space data.
• To address limitations of current methods, including image artifacts, reconstruction speed, and generalization capabilities.

Main Methods:

• A hybrid approach combining sparse-based reconstruction using a Low-Dimensional Manifold Model (LDMM) and deep learning-based reconstruction with ShWideResNet.
• ShWideResNet integrates Scaling Wide Residual Network (SWideRNet) and Shepard Convolutional Neural Network (ShCNN) for enhanced reconstruction.
• Pre-processing involves generating an aliased image, applying a Low-Pass Filter (LPF) for anti-aliasing, and downsampling before reconstruction.

Main Results:

• The ShWideResNet achieved a Root Mean Square Error (RMSE) of 0.102.
• A Peak Signal-to-Noise Ratio (PSNR) of 33.600 dB and a Structural Similarity Index Measure (SSIM) of 0.949 indicate high image quality.
• The reconstruction time was significantly reduced to 20.288 seconds, demonstrating computational efficiency.

Conclusions:

• The proposed ShWideResNet offers a promising solution for fast and accurate MRI reconstruction.
• This technique effectively reduces artifacts and reconstruction time, potentially improving clinical workflow and patient experience.
• The fusion of sparse-based and ShWideResNet-based reconstruction outputs yields superior results compared to individual methods.