Original ContributionAn algorithm for sparse MRI reconstruction by Schatten p-norm minimization
Introduction
Magnetic resonance imaging (MRI) is a comparatively slow imaging modality. Speeding up the data acquisition time has always been of interest to the MRI research community. Until recently, most of the effort in decreasing the MR scan (data acquisition) time had been dedicated to improving the hardware of the scanner. Once the full K-space data are acquired, reconstructing the image is almost trivial — applying 2D inverse Fourier transform. In recent years, signal processing researchers have shown that it is possible to reduce the scan time by acquiring less K-space data followed by a nonlinear reconstruction technique based on the recent findings of compressed sensing [1], [2], [3], [4], [5], [6], [7]. Compressed sensing (CS)-based methods reconstruct the image by framing a nonlinear optimization problem that exploits the sparsity of the MR image in a transform domain such as wavelet, contourlet or total variation.
Our work has the same interest in mind — reconstructing the MR image from subsampled K-space data. But instead of applying CS-type methods, we will show that similar reconstruction accuracy can be achieved by exploiting the rank deficiency of the image. The way we formulate the reconstruction problem requires solving a nonlinear optimization problem that minimizes the Schatten p-norm (0<p≤1) [8] of the image. This work does not compete against CS-based MR reconstruction techniques; rather, it proposes an alternate approach. Our reconstruction accuracy is similar to standard CS-based methods, but takes considerably less time.
Rank deficiency can also be exploited to denoise MR images from legacy scanners. Current MR scanners sample the full K-space and reconstruct the image by applying an inverse Fourier transform. Since the K-space data is itself corrupted by noise, the reconstructed image is noisy as well. Such images require an additional denoising step. The MR image is assumed to be rank deficient (has a small number of high singular values); however, the noise is not. The noise is full rank but has very small singular values. Thus it is possible to denoise the image by properly thresholding the singular values of the noisy image. However, instead of applying arbitrary thresholding to the singular values noisy image, we propose to denoise it through Schatten p-norm minimization.
The rest of the article is organized into several sections. The following section discusses the CS-based MR image reconstruction. Section 3 formulates the reconstruction problem as one of Schatten p-norm minimization. In Section 4, the formulation behind the denoising problem is provided. The algorithmic development for solving the minimization problem is described in Section 5. The experimental evaluation is performed in Section 6. Finally, in Section 7, the conclusions of the work are discussed.
Section snippets
CS-based reconstruction from subsampled K-space measurements
MR images are spatially redundant, i.e., a pixel value at a particular location is highly dependent on the values of neighboring locations except at discontinuities along the edges. Since the images are spatially redundant, they are compressible, i.e., one does not require storing all the pixels, rather one can only store a few wavelet/DCT/singular value coefficients from which the image can be easily generated.
The spatial redundancy of MR images is well captured by different types of
MR image reconstruction as a Schatten p-norm minimization problem
From the discussion in Section 2, we understand that when the number of K-space samples is considerably more than the number of degrees of freedom of the solution, it is feasible to solve the reconstruction problem. Transform domain representation is not the only way the information content of the image is compactly captured. In many cases, singular value decomposition (SVD) can also efficiently represent the information content of the MR image.
Owing to the spatial redundancy, MR images do not
Denoising via Schatten p-norm minimization
An MR image is in most cases corrupted by white Gaussian noise. When the MR images are reconstructed by CS-based methods or by our proposed method, the image is denoised implicitly during reconstruction. However, in legacy MR scanners, the reconstruction algorithm does not incorporate any denoising step. They used to scan the full K-space data and the reconstruction step consisted only of applying a 2D inverse Fourier transform. Since the K-space data is itself corrupted by noise, applying the
Solving the Schatten p-norm minimization problem
There are quite a few efficient algorithms to solve the nuclear norm (Schatten 1-norm) minimization problem [34], [35], [36]. But there is no algorithm to solve our proposed nonconvex Schatten p-norm minimization problem. In this section, we derive a first-order algorithm to solve our proposed optimization problem.
The problem is to solve (14). However, it is difficult to solve the constrained problem directly. Therefore we propose to solve the following unconstrained Lagrangian version of (14)
Experimental results
We carried out the experiments on three slices: two are from the BrainWeb and one from the National Institute of Health database (see Fig. 1). All the images are of size 256 by 256 pixels. Radial sampling was employed to collect the K-space data. The reason to simulate radial sampling is that it is by far the fastest K-space sampling method [40]. However, our work is generalized and can be employed with any other sampling scheme. Nonuniform FFT (NUFFT) [41] is used as the mapping (F) between
Conclusion
Reducing the data acquisition time for MR scans has always been a challenge. In this article, we address the problem from a signal processing perspective. A new approach to reconstruct MR images from subsampled K-space measurements has been presented here. It exploits the rank deficiency of the MR images to frame the reconstruction problem.
In recent years, CS-based MR image reconstruction techniques have been successful in reconstructing image from subsampled K-space measurements. Our proposed
References (45)
- et al.
Reconstruction of dynamic contrast enhanced magnetic resonance imaging of the breast with temporal constraints
Magn Reson Med
(2010) - et al.
MR image reconstruction of sparsely sampled 3D k-space data by projection-onto-convex sets
Magn Reson Med
(2006) Wavelet packet compression of medical images
Digital Signal Processing
(2002)- et al.
Compression and reconstruction of MRI images using 2D DCT
Magn Reson Med
(1992) Medical image compression using DCT-based subband decomposition and modified SPIHT data organization
Int J Med Inform
(2007)- et al.
On the approximability of minimizing nonzero variables or unsatisfied relations in linear systems,
Theoretical Computer Science
(1998) - et al.
SVD lossy adaptive encoding of 3D digital images for ROI progressive transmission
Image Vis Comput
(2010) - et al.
MRI scan time reduction through non-uniform sampling and SVD-based estimation
Signal Process
(1996) - et al.
Sparse recovery by non-convex optimization — instance optimality
Appl Comput Harmon Anal
(2010) - et al.
Complex denoising of MR data via wavelet analysis: application for functional MRI
Magn Reson Med
(2000)