Deep compressed sensing MRI via a gradient-enhanced fusion model

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Compressed Sensing MRI (CS-MRI) accelerates imaging but faces challenges with reconstruction speed and generalization.
• Conventional CS-MRI methods are time-consuming and struggle with multicontrast datasets.
• Existing deep learning approaches often overlook crucial prior information in MR images.

Purpose of the Study:

• To develop an iterative fusion model for Magnetic Resonance Imaging (MRI) reconstruction.
• To integrate image and gradient-based priors using convolutional neural network models.
• To enhance reconstruction quality and preserve detailed information in accelerated MRI.

Main Methods:

• Proposed a Gradient-Enhanced Fusion Network (GFN) utilizing dense blocks and dilated convolutions for efficient feature extraction.
• Incorporated decomposed gradient maps to enhance structural information and image details.
• Employed l2-norm to fuse image and gradient priors, promoting gradient sparsity for improved edge reconstruction.

Main Results:

• GFN outperformed existing CS-MRI methods in Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).
• Demonstrated superior generalization ability through cross-center training and testing experiments.
• Showcased significant improvements when applied to other deep learning methods, enhancing their reconstruction results.

Conclusions:

• Gradient-based priors reconstructed by GFNs effectively enhance edges and details in under-sampled MRI data.
• The proposed fusion model improves generalization on multicontrast datasets by integrating image and gradient priors.
• The method offers a promising approach for fast, accurate, and robust MRI reconstruction.