Compressed sensing MRI via a multi-scale dilated residual convolution network.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Compressed Sensing (CS) MRI algorithms are iterative and time-consuming, with limited model capacity.
• Deep learning (DL) based CS-MRI faces challenges balancing performance and network size.
• Existing methods struggle with speed and accuracy in MRI reconstruction.

Purpose of the Study:

• To develop a novel multi-scale dilated network for high-speed and high-performance MRI reconstruction.
• To address the limitations of conventional and deep learning-based CS-MRI methods.
• To improve reconstruction accuracy and visual quality while maintaining model efficiency.

Main Methods:

• A multi-scale dilated network architecture was developed for MRI reconstruction.
• Dilated convolutions were employed to reduce parameters and expand receptive fields.
• Global and local residual learnings were integrated to capture image details and edges.
• Concatenation layers were used to fuse multi-scale features and residual learnings.

Main Results:

• The proposed method demonstrated superior reconstruction accuracy and visual improvements compared to non-deep and deep learning CS-MRI algorithms.
• The network achieved high speed and outstanding performance in MRI reconstruction.
• The model exhibited stability in noisy settings and was successfully extended to MRI super-resolution tasks.

Conclusions:

• The developed multi-scale dilated network offers a significant advancement in MRI reconstruction, balancing speed and accuracy.
• This approach overcomes key limitations of existing CS-MRI techniques.
• The model shows potential for broader applications in medical imaging, including super-resolution.