Jingyuan Lyu

Dynamic magnetic resonance imaging using compressed sensing with self-learned nonlinear dictionary (NL-D)

Compressed Sensing (CS) is a new paradigm in signal processing and reconstruction from sub-nyquist sampled data. CS has shown promising results in accelerating dynamic Magnetic Resonance Imaging (dMRI). CS based approaches hugely rely on sparsifying transforms to reconstruct the dynamic MR images from its undersampled k-space data. Recent developments in dictionary learning and nonlinear kernel based methods have shown to be capable of representing dynamic images more sparsely than conventional linear transforms. In this paper, we propose a novel method (NL-D) to represent the dMRI more sparsely using self-learned nonlinear dictionaries based on kernel methods. Based on the proposed model, a new iterative approach for image reconstruction relying on pre-image reconstruction is developed within CS framework. Simulation results have shown that the proposed method outperforms the conventional CS approaches based on linear sparsifying transforms.

Fast GRAPPA reconstruction with random projection

To address the issue of computational complexity in generalized autocalibrating partially parallel acquisition (GRAPPA) when several calibration data are used.

Recovery of parametric manifold from reduced measurements: Application to magnetic resonance parameter mapping

Magnetic resonance (MR) quantitative imaging such as parameter mapping has great potential in clinical applications. It requires acquisition of a sequence of images at different time points to extract the quantitative parameters. A practical challenge is the tradeoff between spatial resolution and acquisition speed. A number of methods have been proposed to address the challenge. In this paper, a novel manifold recovery approach is proposed to obtain the quantitative map from highly reduced measurements. The performance of the proposed method is demonstrated using simulated and real datasets on parameter mapping.

Accelerating MR parameter mapping using nonlinear manifold learning and supervised pre-imaging

In this paper, we propose a new reconstruction framework that utilizes nonlinear models to sparsely represent the MR parameter-weighted image in a high dimensional feature space. Different from the prior work with nonlinear models where the image series is reconstructed simultaneously, each image at a specific time point is assumed to lie in a low-dimensional manifold and is reconstructed individually. The low-dimensional manifold is learned from the training images generated by the parametric model. To reconstruct each image, among infinite number of solutions that satisfy the data consistent constraint, the one that is closest to the manifold is selected as the desired solution. The underlying optimization problem is solved using kernel trick and split Bregman iteration algorithm. The proposed method was evaluated on a set of in-vivo brain T2 mapping data set and shown to be superior to the conventional compressed sensing methods.

Efficient GRAPPA reconstruction using random projection

As a data-driven technique, GRAPPA has been widely used for parallel MRI reconstruction. In GRAPPA, a large amount of calibration data is desirable for accurate calibration and thus estimation. However, the computational time increases with the large number of equations to be solved, which is especially serious in 3-D reconstruction. To address this issue, a number of approaches have been developed to compress the large number of physical channels to fewer virtual channels. In this paper, we tackle the complexity problem from a different prospective. We propose to use random projections to reduce the dimension of the problem in the calibration step. Experimental results show that randomly projecting the data onto a lower-dimensional subspace yields results comparable to those of traditional GRAPPA, but is computationally significantly less expensive.

M-MRI: A manifold-based framework to highly accelerated dynamic magnetic resonance imaging

High-dimensional signals, including dynamic magnetic resonance (dMR) images, often lie on low dimensional manifold. While many current dynamic magnetic resonance imaging (dMRI) reconstruction methods rely on priors which promote low-rank and sparsity, this paper proposes a novel manifold-based framework, we term M-MRI, for dMRI reconstruction from highly undersampled k-space data. Images in dMRI are modeled as points on or close to a smooth manifold, and the underlying manifold geometry is learned through training data, called “navigator” signals. Moreover, low-dimensional embeddings which preserve the learned manifold geometry and effect concise data representations are computed. Capitalizing on the learned manifold geometry, two regularization loss functions are proposed to reconstruct dMR images from highly undersampled k-space data. The advocated framework is validated using extensive numerical tests on phantom and in-vivo data sets.

Accelerating dynamic magnetic resonance imaging by nonlinear sparse coding

Although being high-dimensional, dynamic magnetic resonance images usually lie on low-dimensional manifolds. Nonlinear models have been shown to capture well that latent low-dimensional nature of data, and can thus lead to improvements in the quality of constrained recovery algorithms. This paper advocates a novel reconstruction algorithm for dynamic magnetic resonance imaging (dMRI) based on nonlinear dictionary learned from low-spatial but high-temporal resolution images. The nonlinear dictionary is initially learned using kernel dictionary learning, and the proposed algorithm subsequently alternates between sparsity enforcement in the feature space and the data-consistency constraint in the original input space. Extensive numerical tests demonstrate that the proposed scheme is superior to popular methods that use linear dictionaries learned from the same set of training data.

Highly accelerated cardiac cine parallel MRI using low-rank matrix completion and partial separability model

This paper presents a new approach to highly accelerated dynamic parallel MRI using low rank matrix completion, partial separability (PS) model. In data acquisition, k-space data is moderately randomly undersampled at the center kspace navigator locations, but highly undersampled at the outer k-space for each temporal frame. In reconstruction, the navigator data is reconstructed from undersampled data using structured low-rank matrix completion. After all the unacquired navigator data is estimated, the partial separable model is used to obtain partial k-t data. Then the parallel imaging method is used to acquire the entire dynamic image series from highly undersampled data. The proposed method has shown to achieve high quality reconstructions with reduction factors up to 31, and temporal resolution of 29ms, when the conventional PS method fails.

Conflict-Cost Based Random Sampling Design for Parallel MRI with Low Rank Constraints

In compressed sensing MRI, it is very important to design sampling pattern for random sampling. For example, SAKE (simultaneous auto-calibrating and k-space estimation) is a parallel MRI reconstruction method using random undersampling. It formulates image reconstruction as a structured low-rank matrix completion problem. Variable density (VD) Poisson discs are typically adopted for 2D random sampling. The basic concept of Poisson disc generation is to guarantee samples are neither too close to nor too far away from each other. However, it is difficult to meet such a condition especially in the high density region. Therefore the sampling becomes inefficient. In this paper, we present an improved random sampling pattern for SAKE reconstruction. The pattern is generated based on a conflict cost with a probability model. The conflict cost measures how many dense samples already assigned are around a target location, while the probability model adopts the generalized Gaussian distribution which includes uniform and Gaussian-like distributions as special cases. Our method preferentially assigns a sample to a k-space location with the least conflict cost on the circle of the highest probability. To evaluate the effectiveness of the proposed random pattern, we compare the performance of SAKEs using both VD Poisson discs and the proposed pattern. Experimental results for brain data show that the proposed pattern yields lower normalized mean square error (NMSE) than VD Poisson discs.

Highly accelerated 3D dynamic contrast enhanced MRI from sparse spiral sampling using integrated partial separability model andJSENSE

Dynamic contrast enhanced MRI requires high spatial resolution for morphological information and high temporal resolution for contrast pharmacokinetics. The current techniques usually have to compromise the spatial information for the required temporal resolution. This paper presents a novel method that effectively integrates sparse sampling, parallel imaging, partial separable (PS) model, and sparsity constraints for highly accelerated DCE-MRI. Phased array coils were used to continuously acquire data from a stack of variable-density spiral trajectory with a golden angle. In reconstruction, the sparsity constraints, the coil sensitivities, spatial and temporal bases of the PS model are jointly estimated through alternating optimization. Experimental results from in vivo DCE liver imaging data show that the proposed method is able to achieve high spatial and temporal resolutions at the same time.