We demonstrate an end-to-end parallel imaging and compressed sensing reconstruction that does not rely on parameter tuning. We combine noise pre-whitening, auto-tuned coil sensitivity estimation, and a noise-constrained compressed sensing reconstruction to eliminate the need to select parameters such as soft threshold regularization. The method is validated across a large corpus of phantom and in vivo data at different levels of SNR and with different types of coils in 2D and in 3D. An end-to-end reconstruction is shown for 2D variable density single-shot fast spin-echo with reconstruction times of less than one minute.
Introduction
Parallel-imaging (PI) [1] and compressed-sensing (CS) [2], herein called PICS, have shown great promise at accelerating MRI. Many techniques build off of PICS to exploit redundant structure, and the success of CS has begun to penetrate the clinic [3]. Nonetheless, broader clinical use of CS is hindered by the need to carefully tune regularization parameters. Although sensitivity to these parameters may be mild in some cases, in others it can lead to noise amplification or over-smoothing due to differences in SNR, acquisition, slice, or anatomy. Coil sensitivity estimation may pose an additional challenge, as it may also depend on algorithm parameters that are sensitive to differences across scans.
In this work we implement an end-to-end PICS reconstruction that does not rely on parameter tuning. This is accomplished by (1) pre-whitening the acquired data using a noise pre-scan, (2) performing an auto-tuned and auto-calibrated coil sensitivity estimation [4,5], and (3) using a noise-constrained CS formulation [6]. The method is validated on both phantom and in vivo data at different SNRs and with different types of coils in 2D and in 3D. An end-to-end reconstruction is shown for 2D variable density single-shot fast spin-echo (VD-SSFSE) with reconstruction times under one minute.
Theory
When the noise level is known, PICS can be formulated as basis pursuit denoising [6]:$$\begin{aligned}&\underset{\mathbf{x}}{\text{minimize}}&&||\mathbf{Wx}||_1\\&\text{subject to}&&||\mathbf{y}-\mathbf{Ax}||_2\le\sigma\sqrt{n},\end{aligned}$$where $$$\mathbf x$$$ is the unknown image, $$$\mathbf W$$$ is a sparsity transform (wavelets for this work), $$$\mathbf A$$$ is the forward model incorporating estimated coil sensitivities, $$$\mathbf y\in\mathbb{R}^n$$$ are the k-space measurements, and $$$\sigma$$$ is the noise standard deviation. The l1-norm of the wavelet coefficients is minimized subject to a data inconsistency at most equal to the expected value of the noise in the acquisition. This problem can be efficiently solved with gradient-based methods, e.g. ADMM [7] and first-order primal-dual algorithm [8]. Since the noise level is known, the coil sensitivities can be reliably estimated using an auto-tuned [5] variant of ESPIRiT [6], an auto-calibration PI method.
Methods
Three experiments were performed. In the first, twenty fully sampled 3D fast spin-echo (FSE) knee datasets acquired at 3T were downloaded from http://www.mridata.org [9]. For each, nine central slices were averaged to serve as noise-free ground truth. After, noise was added to each averaged slice and the data were retrospectively under-sampled using a Poisson-disc sampling pattern (R=4). Each case was reconstructed with parameter-free PICS using the first-order primal-dual algorithm, as well as unconstrained PICS using FISTA [10] at different levels of regularization. The PSNR was evaluated and compared across the 20 cases.
In the second experiment, a fully sampled slice of the NIST/ISMRM System Phantom was acquired at 3T with a 32 channel body coil using 2D FSE, and the acquisition was repeated 16 times. A noise pre-scan was used to pre-whiten the data to unit variance. Different SNR levels were achieved by averaging an appropriate number of acquisitions, followed by retrospective Poisson-disc under-sampling (R=4) and coil compression to 16 channels [11]. Each case was reconstructed with parameter-free PICS and unconstrained PICS with different regularization parameters. The PSNR was evaluated and compared across different instances of averaged acquisitions.In the third experiment, an adult patient's abdomen was scanned with 2D prospectively under-sampled VD-SSFSE [12] at 3T with IRB approval and informed consent. The data were pre-whitened using a noise pre-scan, coil sensitivities were estimated with auto-tuned ESPIRiT, and the data were reconstructed using parameter-free and unconstrained PICS. As a proof-of-principle, an end-to-end reconstruction based on the acquisition was developed and integrated into the Stanford Hospital. The reconstructions were implemented and merged into the Berkeley Advanced Reconstruction Toolbox [13,14].