A review on deep learning MRI reconstruction without fully sampled k-space

Abstract

Background Magnetic resonance imaging (MRI) is an effective auxiliary diagnostic method in clinical medicine, but it has always suffered from the problem of long acquisition time. Compressed sensing and parallel imaging are two common techniques to accelerate MRI reconstruction. Recently, deep learning provides a new direction for MRI, while most of them require a large number of data pairs for training. However, there are many scenarios where fully sampled k-space data cannot be obtained, which will seriously hinder the application of supervised learning. Therefore, deep learning without fully sampled data is indispensable. Main text In this review, we first introduce the forward model of MRI as a classic inverse problem, and briefly discuss the connection of traditional iterative methods to deep learning. Next, we will explain how to train reconstruction network without fully sampled data from the perspective of obtaining prior information. Conclusion Although the reviewed methods are used for MRI reconstruction, they can also be extended to other areas where ground-truth is not available. Furthermore, we may anticipate that the combination of traditional methods and deep learning will produce better reconstruction results.

Image guidance for FLASH radiotherapy

“…Alternative solutions deemed more in reach for clinical deployment rely on the online capturing of the anatomical location, currently limited to fast 2D cine MRI sequences at a few Hz frame rate at modern MR‐Linac, 62 combined with dosimetric calculations based on log‐files and synthetic CT generation from the acquired MR data. This may improve with fast reconstruction methods utilizing deep learning running on fast graphical processing units (GPU) in the future 63 . However, such approaches are especially meant to overcome one of the still outstanding great challenges of modern RT, namely, to properly account for organ motion and deformation during beam delivery.…”

“…This may improve with fast reconstruction methods utilizing deep learning running on fast graphical processing units (GPU) in the future. 63 However, such approaches are especially meant to overcome one of the still outstanding great challenges of modern RT, namely, to properly account for organ motion and deformation during beam delivery. However, the achievable temporal resolution of all above mentioned technologies might still be a challenge for application in time-F I G U R E 1 An idealized illustration of a beam FLASH-radiotherapy (FLASH-RT) pulse structure, with high instantaneous dose rate per pulse ( Ḋpulse ) and short pulse lengths (Δt pulse ).…”

SPICE: Self-Supervised Learning for MRI with Automatic Coil Sensitivity Estimation

“…There is a growing interest in DL-based image reconstruction to reduce the dependence of training on high-quality ground-truth data (see recent reviews [47][48][49]). Some well-known strategies include Noise2Noise (N2N) [50], Noise2Void (N2V) [51], deep image prior (DIP) [52], Compressive Sensing using Generative Models (CSGM) [53], and equivariant imaging [54].…”

CoRRECT: A Deep Unfolding Framework for Motion-Corrected Quantitative R2* Mapping

“…In the past decade, DL has gained great popularity for solving MRI inverse problems due to its excellent performance (see reviews in [6,30,31]). A widely-used supervised DL approach is based on training an image reconstruction CNN R θ by mapping a corrupted image A † y to its clean target x, where A † is an operator that maps the measurements back to the image domain.…”