MR image reconstruction using densely connected residual convolutional networks.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Deep learning methods offer improved MRI reconstruction performance and reduced acquisition times compared to traditional analytical techniques.
• Real-time, clinical-grade MR image reconstruction is increasingly feasible due to advancements in deep learning methodologies.
• Short connections between layers are known to enhance deep network performance in image processing tasks.

Purpose of the Study:

• To introduce a novel MRI reconstruction method employing dense connections within residual blocks.
• To investigate the efficacy of these densely connected residual blocks in enhancing MR image reconstruction.
• To integrate these blocks into various deep network architectures for improved performance.

Main Methods:

• Developed a novel MRI reconstruction approach utilizing densely connected residual blocks.
• Implemented feature map concatenation within blocks to propagate information effectively.
• Augmented existing deep network models with these novel blocks in new configurations.

Main Results:

• Quantitative and qualitative evaluations demonstrated significant improvements in MRI reconstruction performance.
• The proposed method, incorporating densely connected blocks, outperformed existing techniques.
• The novel integration of dense connections proved effective for MR image reconstruction.

Conclusions:

• The introduction of densely connected residual blocks represents a significant advancement in MR image reconstruction.
• This novel approach enhances reconstruction performance, paving the way for more efficient MRI.
• The findings support the utility of dense connections for complex inverse problems in medical imaging.