CS-MRI Reconstruction Using an Improved GAN with Dilated Residual Networks and Channel Attention Mechanism

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Compressed sensing (CS) MRI accelerates image acquisition but requires sophisticated reconstruction techniques.
• Deep learning, particularly generative adversarial networks (GANs), shows promise for CS-MRI reconstruction.
• Increasing model complexity in deep learning can lead to longer reconstruction times and convergence issues.

Purpose of the Study:

• To develop a novel, efficient GAN-based model for compressed sensing MRI reconstruction.
• To enhance reconstruction speed and quality without escalating model complexity.
• To improve the stability and performance of CS-MRI reconstruction.

Main Methods:

• A U-net based generator architecture incorporating dilated residual (DR) networks to expand the receptive field without increasing computational load.
• Integration of a channel attention mechanism (CAM) to refine channel information and reduce noise.
• Experimental validation using public domain human brain MRI datasets and ablation studies.

Main Results:

• The proposed DR-CAM-GAN model demonstrated superior performance in CS-MRI reconstruction.
• Incorporating DR networks and CAM improved peak signal-to-noise ratios (PSNR) by approximately 1.2 dB and 0.8 dB, respectively, at 10x acceleration.
• The model achieved significant gains in structural similarity index measure (SSIM) (14%) and PSNR (15%) compared to U-net, with mean squared error (MSE) reduced by a factor of two to seven.

Conclusions:

• The novel DR-CAM-GAN model offers an effective approach for enhancing the efficiency and quality of compressed sensing MRI reconstruction.
• The integration of DR networks and CAM contributes to improved image quality and reconstruction stability.
• This model presents a promising advancement for clinical applications requiring faster and more accurate MRI scans.