De-Aliasing and Accelerated Sparse Magnetic Resonance Image Reconstruction Using Fully Dense CNN with Attention Gates

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Accelerated magnetic resonance imaging (MRI) using sparse data introduces aliasing artifacts.
• Conventional reconstruction methods struggle to accurately recover image content from undersampled MRI data.
• Artifacts obscure crucial details, limiting the diagnostic value of fast MRI scans.

Purpose of the Study:

• To propose an advanced convolutional neural network (CNN) for artifact reduction in accelerated MRI.
• To enhance MRI reconstruction by improving feature learning and network generalization.
• To introduce a novel undersampling pattern for more efficient k-space data acquisition.

Main Methods:

• Developed a fully dense attention CNN (FDA-CNN) by integrating dense connectivity and an attention mechanism into the Unet architecture.
• Implemented attention gates in decoder layers to focus on relevant image features and reduce irrelevant activations.
• Designed a novel undersampling pattern in the phase direction, acquiring both low and high frequencies randomly and non-randomly from k-space.

Main Results:

• FDA-CNN demonstrated superior performance compared to five deep learning and two compressed sensing MRI reconstruction techniques.
• The proposed method reconstructed smoother and brighter MRI images.
• FDA-CNN achieved a 2 dB improvement in mean PSNR, 0.35 in SSIM, and 0.37 in VIFP compared to Unet at an acceleration factor of 5.

Conclusions:

• The FDA-CNN effectively removes aliasing artifacts in accelerated MRI, significantly improving image quality.
• The combination of dense connectivity and attention mechanisms enhances feature learning and network generalization.
• The novel undersampling pattern and FDA-CNN offer a promising approach for faster and higher-quality MRI acquisition.