Knowledge-driven deep learning for fast MR imaging: Undersampled MR image reconstruction from supervised to...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging (MRI)

Background:

• Deep learning (DL) is a powerful tool for accelerating MRI acquisition and reconstruction.
• MRI reconstruction presents unique challenges due to physics-based processes and data properties, differing from natural image restoration.
• Integrating domain knowledge with data-driven DL approaches is crucial for accurate MRI.

Purpose of the Study:

• To review significant challenges in knowledge-driven DL for fast MRI.
• To present notable solutions and trends in DL-based MRI reconstruction.
• To discuss MR vendor adoption, open questions, and future directions in DL for reliable MRI systems.

Main Methods:

• Review of existing literature on DL applications in MRI.
• Analysis of knowledge-driven DL approaches, including neural network architectures.
• Examination of learning paradigms: supervised, semi-supervised, and unsupervised learning in MRI reconstruction.

Main Results:

• Identified key challenges in applying DL to MRI's physics-based nature.
• Highlighted successful integration of domain knowledge with DL models.
• Documented the evolution of learning strategies from supervised to unsupervised methods.

Conclusions:

• Knowledge-driven DL offers significant potential for accelerating MRI.
• Addressing specific MRI challenges requires tailored DL solutions and learning strategies.
• Future research should focus on open questions for robust and reliable DL-based MRI systems.