Daeun Kim¹, Jonathan Polimeni², Kawin Setsompop², and Justin Haldar^1
1Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, United States, ²Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Harvard Medical School, Massachusetts General Hospital, Charlestown, MA, United States

Noise correlations exist in multi-channel k-space data, and methods to optimally account for this correlation have been used for a long time in image-domain parallel imaging methods like SENSE.  However, methods to address noise are not widely-utilized for Fourier-domain parallel imaging methods like GRAPPA, SPIRiT, and AC-LORAKS.   In this work, we demonstrate that properly accounting for spatially-varying noise correlation can substantially reduce the noise level of coil-combined images.  Further, we demonstrate the existence of previously-unknown correlations between the real and imaginary parts of the noise in reconstructed images.  Accounting for this extra correlation can reduce the noise level even further.

Jiayang Wang¹ and Justin P. Haldar^1
1Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, United States

The g-factor is commonly used for quantifying the noise amplification associated with accelerated data acquisition and linear image reconstruction, and is frequently used to compare different k-space sampling strategies. While previous work computes g-factors in the image domain, we observe in this work that g-factors can also be used to quantify uncertainty in various transform domains (e.g., the wavelet domain and the multi-channel Fourier domain).   These transform-domain g-factor maps provide complementary information to conventional image-domain g-factor maps, and are potentially useful for k-space sampling pattern design.

Nicholas Dwork¹, Corey A. Baron², Ethan M. I. Johnson³, Daniel O'Connor⁴, Adam B. Kerr⁵, Peder E. Z. Larson¹, Jeremy Gordon¹, and John M. Pauly^6
1Radiology and Biomedical Imaging, University of California in San Francisco, San Francisco, CA, United States, ²Medical Biophysics, Western University, London, ON, Canada, ³Biomedical Engineering, Northwestern University, Evanston, IL, United States, ⁴Mathematics and Statistics, University of San Francisco, San Francisco, CA, United States, ⁵Center for Cognitive and Neurobiological Imaging, Stanford University, Stanford, CA, United States, ⁶Electrical Engineering, Stanford University, Stanford, CA, United States

In this paper, the reconstructed image is the result of a compressed sensing optimization problem that includes constraints based on fundamental physics.  The problem is solved using an alternating minimization approach: two convex optimization problems are alternately solved, one with the Fast Iterative Shrinkage Threshold Algorithm (FISTA) and the other with the Primal-Dual Hybrid Gradient method.  Results show improved detail when compared to conventional SENSE results.

Tieyuan Lu¹, Xinlin Zhang¹, Yihui Huang¹, Di Guo², and Xiaobo Qu^1
1Department of Electronic Science, Xiamen University, Xiamen, China, ²School of Computer and Information Engineering, Xiamen University of Technology, Xiamen, China

Magnetic resonance imaging has been widely applied in clinical diagnosis, however, is limited by its long data acquisition time. Although imaging can be accelerated by sparse sampling and parallel imaging, achieving promising reconstruction images with a fast reconstruction speed remains a challenge. In this work, we design the neural network structure from the perspective of sparse iterative reconstruction. The experimental results of a public knee dataset show that compared with the optimization-based method and the latest deep learning parallel imaging methods, the proposed network has less error in reconstruction and is more stable under different acceleration factors.

Edwin Versteeg¹, Tijl Van der Velden¹, Jeroen Hendrikse¹, Dennis Klomp¹, and Jeroen Siero^1,2
1Radiology, University Medical Center Utrecht, Utrecht, Netherlands, ²Spinoza Centre for Neuroimaging Amsterdam, Amsterdam, Netherlands

A silent gradient axis driven at 20 kHz can be used to decrease acoustic noise and increase patient comfort in MR-exams. However, the speed of this silent acquisition is determined by the amount of data needed for artifact free reconstructions. In this work, we present a framework for increasing the time efficiency while maintaining image quality of silent imaging with a silent gradient axis. We show that, by using a generalized conjugate gradient SENSE reconstruction, a 2-3-fold decrease in scan time is feasible both on a phantom and in-vivo.

Onur Beker^1,2, Congyu Liao^1,3, Jaejin Cho^1,3, Zijing Zhang^1,4, Kawin Setsompop^1,3,5, and Berkin Bilgic^1,3,5
1Martinos Center for Biomedical Imaging, Charlestown, MA, United States, ²Electrical and Electronics Engineering, Bogazici University, Istanbul, Turkey, ³Harvard Medical School, Boston, MA, United States, ⁴College of Optical Science and Engineering, Zhejiang University, Hangzhou, China, ⁵Harvard/MIT Health Sciences and Technology, Cambridge, MA, United States

We propose a convolutional neural network (CNN) approach that works synergistically with physics-based reconstruction methods to reduce artifacts in accelerated MRI. Given reconstructed coil k-spaces, our network predicts a k-space correction term for each coil. This is done by matching the difference between the acquired autocalibration lines and their erroneous reconstructions, and generalizing this error term over the entire k-space. Application of this approach on existing reconstruction methods show that SPARK suppresses reconstruction artifacts at high acceleration, while preserving and improving on detail in moderate acceleration rates where existing reconstruction algorithms already perform well; indicating robustness.

Aniket Pramanik¹, Hemant Kumar Aggarwal¹, and Mathews Jacob^1
1Electrical and Computer Engineering, The University of Iowa, Iowa City, IA, United States

We introduce a fast model based deep learning approach for calibrationless parallel MRI reconstruction. The proposed scheme is a non-linear generalization of structured low-rank (SLR) methods that self learn linear annihilation filters. It pre-learns non-linear annihilation relations in the Fourier domain from exemplar data which significantly reduces the computational complexity, making it three orders of magnitude faster than SLR schemes. It allows incorporation of spatial domain prior that offers improved performance over calibrated image domain MoDL approach. The calibrationless strategy minimizes potential mismatches between calibration data and the main scan, while eliminating the need for a fully sampled calibration region.

Rodrigo A. Lobos¹, Tae Hyung Kim¹, Kawin Setsompop², and Justin P. Haldar^1
1Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, United States, ²Martinos Center for Biomedical Imaging, Charlestown, MA, United States

Many autocalibrated parallel imaging reconstruction methods are based on linear-predictive/autoregressive principles, including noniterative GRAPPA-type interpolation methods, iterative SPIRiT-type annihilation methods, and structured low-rank matrix methods like PRUNO and Autocalibrated LORAKS.  In principle, all of these approaches could be adapted for simultaneous multislice (SMS) reconstruction. However, in practice, GRAPPA-type SMS methods are popular, but there has been limited exploration of more advanced annihilation-based or structured low-rank matrix SMS methods. In this work, we adapt and evaluate these advanced approaches for SMS reconstruction. Results demonstrate that these advanced approaches can offer substantial improvements over simpler GRAPPA-type methods when applied to SMS.

Abdul Basit^1,2, Omair Inam¹, Mahmood Qureshi¹, and Hammad Omer^1
1Electrical and Computer Engineering, Comsats University Islamabad, islamabad, Pakistan, ²Computer Engineering, Khwaja Fareed University of Engineering and Information Technology, Rahim Yar Khan, Pakistan

Data acquisition and reconstruction speed both are crucial for real-time MRI. However, MR image reconstruction speed is highly dependent on the processing capabilities of the hardware platforms (e.g. CPUs, GPUs). Recently, it has been observed that Field Programmable Gate Arrays (FPGA) are a potential candidate to address the computational demands of parallel MRI algorithms.  This paper presents the first design effort to implement high performance 32-bit floating-point FPGA-based coprocessor for real-time GRAPPA reconstruction. In-vivo results of 12-channel, 3.0T human-head dataset show that the proposed system speeds up the image reconstruction time up to 106x without compromising image quality.

Rida Zainab¹, Muhammad Haseeb Hassan¹, Omair Inam¹, Ibtisam Aslam^1,2, and Hammad Omer^1
1Department of Electrical and Computer Engineering, COMSATS University Islamabad, Islamabad, Pakistan, ²Department of Radiology and Medical Informatics, Hospital University of Geneva, Geneva, Switzerland

GRAPPA reconstructed images may exhibit noise modulated by the receiver coil sensitivities. Total variation (TV) regularization has been recently used to solve the image de-noising problem. However, conventional TV fails to remove staircase artifacts in the reconstructed MR images due to inhomogeneities in field strength and receiver coils. In this abstract, total generalized variation (TGV) regularization is used to de-noise the GRAPPA reconstructed images, while eliminating the limitations posed by TV. Experiments are performed on 8-channel in-vivo human-head data set. The results show that the proposed method successfully removes the noise and preserve fine details in the GRAPPA reconstructed images.

Xiaoxia Zhang¹, Xiaopeng Zong¹, Yong Chen¹, Zhenghan Fang¹, and Pew-Thian Yap^1
1Department of Radiology, Department of Radiology and Biomedical Research Imaging Center (BRIC), University of North Carolina at Chapel Hill, Chapel Hill, NC, United States

We propose a strategy based on residual U-Net to reconstruct MR images from undersampled multichannel data by considering both k-space and image space. Our method first imputes the missing data points in k-space by utilizing the intrinsic relationships among channels. Then, the image reconstructed from the imputed k-space data is fed to another network for spatial detail refinement. Our method does not necessarily require auto-calibration signal (ACS) and is hence less susceptible to motion-induced inconsistency between the ACS and the undersampled data. Comprehensive evaluation indicates that our method yields images with superior perceptual details. 

Husnain Javid Bhatti^1,2, Fariha Aamir¹, Ibtisam Aslam^1,3, Khan Afsar^1,4, and Hammad Omer^1
1Department of Electrical and Computer Engineering, COMSATS University Islamabad (CUI), Islamabad, Pakistan, ²School of Electrical Engineering and Computers Science, National University of Science and Technology (NUST), Islamabad, Pakistan, ³Department of Radiology and Medical Informatics, Hospital University of Geneva, GENEVA, Switzerland, ⁴Department of Electronic science and Technology, Xiamen University, Xiamen, China

To reduce MRI scan time, under-sampled non-Cartesian trajectories are used which lead to artifacts. This work proposes a new method ‘GROG with calibration-less pMRI for CS based p-thresholding’ to reconstruct MR images from the under-sampled radial k-space data. The proposed method is validated on the phantom and 1.5T human head data and provides significant improvement both visually and in terms of quantifying parameters (AP, RMSE & PSNR) e.g. 77% and 86% improvement in AP, 5% and 32% improvement in RMSE, 7% and 11% improvement in PSNR at AF=4 for the phantom data than POCS and pseudo Cartesian GRAPPA, respectively.

Yuanyuan Hu¹, Zhuonan He², Jinjie Zhou², Minghui Zhang², Qiegen Liu², and Dong Liang^1
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518055, China, Shenzhen, China, ²the Department of Electronic Information Engineering, Nanchang University, Nanchang 330031, China, Nanchang, China

  Ill-posed inverse problems embodied in parallel imaging remain an active research topic in several decades, with new approaches constantly emerging. Built on the observation that both dictionary learning and conventional sparse coding extract high-frequency component to model, we derived a novel strategy named HDAEP to explore the prior on high-frequency domain on the basis of denoising autoencoding. After the prior is learned from the trained network, the iteratively Gauss-Newton method is employed to jointly estimating the images and coil sensitivities. Experimental results show that the proposed method can achieve superior performances on parallel MRI reconstruction compared to state-of-the-arts. 

Eun Ji Lim¹, Guobin Li², Chaohong Wang², Zhaopeng Li², Shasha Yang², and Jaeseok Park^1
1Sungkyunkwan University, Suwon, Republic of Korea, ²United Imaging Healthcare, Shanghai, China

SMS methods typically consist of the following two steps: kernel calibration and reconstruction. However, discrepancies between calibration and imaging occur due to either different image contrasts or subject motions, resulting in residual aliasing artifacts. To tackle these problems, in this work we propose a robust SMS technique exploiting Hankel subspace learning with self-calibration and self-referencing magnitude prior. An SMS filter is designed to strictly control pass-bands and stop-bands to reduce the dependence of image contrast on reconstruction. Both external and internal calibrating signals are included in the calibration step, while a self-referencing magnitude prior is imposed in the reconstruction step. 

Stephen Cauley^1,2, Bryan Clifford³, Steffen Bollmann³, Thorsten Feiweier⁴, Berkin Bilgic^1,2, Kawin Setsompop^1,2,5, and Lawrence L. Wald^1,2,5
1Department of Radiology, A. A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA, United States, ²Harvard Medical School, Boston, MA, United States, ³Siemens Medical Solutions USA, Boston, MA, United States, ⁴Siemens Healthcare GmbH, Erlangen, Germany, ⁵Harvard-MIT Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, United States

We employ a compact phase modeling strategy for accurate multi-shot echo-planar imaging (msEPI) reconstruction. As an alternative to approaches that perform msEPI reconstruction using strict low-rank constraints, we recast the problem as an iterative relative phase estimation problem. This allows for us to utilize existing techniques such as ESPIRiT, which are formulated for determining relative magnitude and phase differences between multi-coil receive arrays. Through an iterative search we jointly estimate an artifact-free combined image and the smooth relative phase between each msEPI shot. We demonstrate the benefits of our approach for clinical and highly-accelerated multi-shot diffusion-weighted acquisitions.