Physics-Driven Deep Learning for Computational Magnetic Resonance Imaging: Combining physics and machine learning for...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Physics

Background:

• Computational magnetic resonance imaging (MRI) traditionally relies on iterative reconstruction algorithms.
• Deep learning (DL) has shown promise in accelerating and improving MRI reconstruction.
• Integrating physical principles into DL offers a powerful approach for complex inverse problems in MRI.

Purpose of the Study:

• To provide a comprehensive overview of physics-driven deep learning methods for MRI reconstruction.
• To discuss classical and modern approaches to solving inverse problems in computational MRI.
• To highlight domain-specific challenges and translational applications of these techniques.

Main Methods:

• Review of classical iterative methods for MRI reconstruction.
• Focus on physics-driven deep learning strategies: physics-guided loss functions, plug-and-play networks, generative models, and unrolled networks.
• Examination of real- and complex-valued neural network components for MRI data.

Main Results:

• Physics-driven DL methods achieve state-of-the-art performance in MRI reconstruction.
• These methods effectively handle both linear and non-linear forward models in MRI.
• Successful translational applications demonstrate the clinical potential of these advanced techniques.

Conclusions:

• Physics-driven deep learning represents a significant advancement in computational MRI.
• Addressing domain-specific challenges is crucial for further development and clinical translation.
• Combining physics-based learning with downstream medical imaging tasks offers promising future research avenues.