A densely interconnected network for deep learning accelerated MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Accelerated MRI acquisition is crucial for reducing scan times and improving patient comfort.
• Deep learning frameworks have shown promise in reconstructing MRI images from undersampled data.
• Existing cascading deep learning models can be further optimized for improved reconstruction fidelity.

Purpose of the Study:

• To enhance accelerated MRI reconstruction using a novel densely connected cascading deep learning framework.
• To investigate the impact of specific architectural modifications on MRI reconstruction quality.
• To evaluate the proposed framework's performance across different acceleration factors.

Main Methods:

• Modified a cascading deep learning framework with input-level dense connections, an improved sub-network, and long-range skip-connections.
• Trained five model configurations on the NYU fastMRI neuro dataset for four- and eightfold acceleration.
• Evaluated reconstruction quality using structural similarity index measure (SSIM), normalized mean square error (NMSE), and peak signal to noise ratio (PSNR).

Main Results:

• The proposed densely interconnected residual cascading network (DIRCN) achieved an 8% and 11% SSIM improvement for four- and eightfold acceleration, respectively.
• NMSE improved by 14% and 23%, and PSNR by 2% and 3% for four- and eightfold acceleration, respectively.
• Individual architectural modifications contributed to performance gains, improving SSIM, NMSE, and PSNR by approximately 2-4%, 4-9%, and 0.5-1%.

Conclusions:

• The proposed architectural modifications offer a straightforward method to enhance existing cascading deep learning frameworks for MRI reconstruction.
• These modifications lead to significant improvements in image reconstruction quality for accelerated MRI.
• The DIRCN framework presents a valuable advancement for efficient and high-fidelity accelerated MRI acquisition.