Learning a Variational Network for Reconstruction of Accelerated MRI Data

Abstract

Purpose: To allow fast and high-quality reconstruction of clinical accelerated multi-coil MR data by learning a variational network that combines the mathematical structure of variational models with deep learning.Theory and Methods: Generalized compressed sensing reconstruction formulated as a variational model is embedded in an unrolled gradient descent scheme. All parameters of this formulation, including the prior model defined by filter kernels and activation functions as well as the data term weights, are learned during an offline training procedure. The learned model can then be applied online to previously unseen data.Results: The variational network approach is evaluated on a clinical knee imaging protocol. The variational network reconstructions outperform standard reconstruction algorithms in terms of image quality and residual artifacts for all tested acceleration factors and sampling patterns. Conclusion:Variational network reconstructions preserve the natural appearance of MR images as well as pathologies that were not included in the training data set. Due to its high computational performance, i.e., reconstruction time of 193 ms on a single graphics card, and the omission of parameter tuning once the network is trained, this new approach to image reconstruction can easily be integrated into clinical workflow.

Active MR k-space Sampling with Reinforcement Learning

“…Since this results in image blur or aliasing artifacts, traditional techniques enhance the image reconstruction using regularized iterative optimization techniques such as compressed sensing [8]. More recently, the inception of large scale MRI reconstruction datasets, such as [28], have enabled the successful use of deep learning approaches to MRI reconstruction [25,3,17,1,32,29,9,24]. However, these methods focus on designing models that improve image reconstruction quality for a fixed acceleration factor and set of measurements.…”

“…More precisely, we specify the active MRI acquisition problem as a Partially Observable Markov Decision Process (POMDP) [19,6], and propose the use of deep reinforcement learning [10] to solve it. 3 Our approach, by formulation, optimizes the reconstruction over the whole range of acceleration factors while considering the sequential nature of the acquisition process -future scans and reconstructions are used to determine the next measurement to take. We evaluate our approach on a large scale single-coil knee dataset [28] 4 and show that it outperforms common acquisition heuristics as well as the myopic approach of [31].…”

Stochastic Deep Compressive Sensing for the Reconstruction of Diffusion Tensor Cardiac MRI

“…Alternative approaches have also been proposed, which modified the deep network architectures to embed traditional optimisation algorithms. These include gradient descent [4], alternating direction method of multipliers (ADMM) [21] or optimisation algorithms inspired by variable splitting techniques [16]. In addition, various clinical applications have been explored including knee imaging [4], brain imaging [23], and dynamic cardiac imaging [19].…”

Model-Based Free-Breathing Cardiac MRI Reconstruction Using Deep Learned & Storm Priors: MODL-STORM

“…The main difference between the proposed scheme and current unrolled CNN methods [6,7] is the reuse of the CNN weights at each iteration; in addition to reducing the trainable parameters, the weight reuse strategy yields a structure that is consistent with the model-based framework, facilitating its easy use with other regularization terms. In addition, the use of the CNN as a plug and play prior rather than a custom designed variational model [7] allows us to capitalize upon the well-established software engineering frameworks such as Tensorflow and Theano.…”