Emerging Trends in Fast MRI Using Deep-Learning Reconstruction on Undersampled k-Space Data: A Systematic Review.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Science

Background:

• Magnetic Resonance Imaging (MRI) offers superior soft-tissue contrast but suffers from long acquisition times.
• Parallel MRI and Compressive Sensing (CS) reduce scan duration via k-space undersampling.
• Deep Learning (DL) integration with undersampled MRI has redefined fast MRI reconstruction.

Purpose of the Study:

• To systematically review and analyze deep learning models for fast MRI reconstruction.
• To highlight key elements, strengths, and weaknesses of recent deep MRI reconstruction methods.
• To provide insights for future research directions in accelerated MRI.

Main Methods:

• Systematic Literature Review (SLR) following PRISMA guidelines.
• Searched Web of Science and Scopus databases for relevant studies.
• Extracted and analyzed data on various deep learning techniques for MRI reconstruction.

Main Results:

• Analysis of techniques including residual learning, encoder-decoder architectures, data-consistency, unrolled networks, learned activations, attention, plug-and-play priors, diffusion models, and Bayesian methods.
• Discussion on loss functions and adversarial training for enhanced reconstruction.
• Exploration of applications like non-Cartesian reconstruction, super-resolution, dynamic MRI, and quantitative mapping.

Conclusions:

• Deep learning significantly advances fast MRI reconstruction, improving image quality and acquisition speed.
• This review serves as a resource for researchers and developers in accelerated MRI.
• Future directions emphasize robust generalization and artifact handling in deep MRI reconstruction.