Physics-Driven Deep Learning for Computational Magnetic Resonance Imaging

Abstract

Physics-driven deep learning methods have emerged as a powerful tool for computational magnetic resonance imaging (MRI) problems, pushing reconstruction performance to new limits. This article provides an overview of the recent developments in incorporating physics information into learningbased MRI reconstruction. We consider inverse problems with both linear and non-linear forward models for computational MRI, and review the classical approaches for solving these. We then focus on physics-driven deep learning approaches, covering physics-driven loss functions, plug-and-play methods, generative models, and unrolled networks. We highlight domain-specific challenges such as real-and complex-valued building blocks of neural networks, and translational applications in MRI with linear and non-linear forward models. Finally, we discuss common issues and open challenges, and draw connections to the importance of physics-driven learning when combined with other downstream tasks in the medical imaging pipeline.

AI in MRI: Computational Frameworks for a Faster, Optimized, and Automated Imaging Workflow

“…As this approach leads to image-domain artifacts, carefully designed reconstruction techniques are essential to allow for clinical-quality-preserving image recovery. Recently, DL techniques have enabled state-of-the-art results in this task, enabling high acceleration and excellent reconstruction quality [ 7 , 8 , 9 , 12 , 13 , 14 , 15 , 16 , 34 , 35 , 36 , 37 , 38 , 39 ]. Their success can be attributed to the ability to learn image priors in a data-driven manner instead of the hand-crafted manner practiced in compressed sensing and dictionary learning [ 25 , 29 , 35 ].…”

“…Their success can be attributed to the ability to learn image priors in a data-driven manner instead of the hand-crafted manner practiced in compressed sensing and dictionary learning [ 25 , 29 , 35 ]. Furthermore, physics-guided unrolled neural networks combine the benefits of DL-based artifact-removal modules with data consistency blocks, which incorporate a physics-based model of the imaging system [ 9 ]. A large body of work has demonstrated the benefits of DL for image reconstruction in 2D MRI scans [ 7 , 8 , 9 , 34 , 35 , 36 , 37 , 38 , 39 ].…”

“…Furthermore, physics-guided unrolled neural networks combine the benefits of DL-based artifact-removal modules with data consistency blocks, which incorporate a physics-based model of the imaging system [ 9 ]. A large body of work has demonstrated the benefits of DL for image reconstruction in 2D MRI scans [ 7 , 8 , 9 , 34 , 35 , 36 , 37 , 38 , 39 ]. More recently, attention has shifted to harnessing DL for accelerating higher-dimensional MRI scans, such as dynamic (temporal) MRI.…”

“…As the first breakthroughs in deep learning (DL) were largely concerned with image perception, interpretation, and analysis, it is not surprising that harnessing this technology has led to enormous progress in medical imaging. For example, DL techniques have enabled state-of-the-art results in image formation and analysis, pathology detection, and protocol planning [ 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 , 11 ].…”

Synthesizing Complex-Valued Multicoil MRI Data from Magnitude-Only Images

“…Furthermore, these publicly available datasets are typically limited in anatomy, acquisition parameters and pathology information. Recent studies have shown that such limitations can sometimes result in hallucinations of structures or artifacts during deep learning-based MRI reconstruction [ 15 , 16 ], limiting the generalization potential of these methods and their clinical use.…”