Arghya Pal¹ and Yogesh Rathi^1
1Department of Psychiatry, Harvard Medical School, Boston, MA, United States

Can a random weighted deep network structure encode informative cues to solve the MR reconstruction problem from highly under-sampled k-space measurements? Trained networks update the weights at training time, while untrained networks optimize the weights at inference time. In contrast, our proposed methodology selects an optimal subnetwork from a randomly weighted dense network to perform MR reconstruction without updating the weights - neither at training time nor at inference time. The methodology does not require ground truth data and shows excellent performance across domains in T1-weighted (head, knee) images from highly under-sampled multi-coil k-space measurements.

Bo Zhou^1,2, Jo Schlemper², Seyed Sadegh Mohseni Salehi ², Neel Dey^2,3, Kevin Sheth¹, Chi Liu¹, James Duncan¹, and Michal Sofka^2
1Yale University, New Haven, CT, United States, ²Hyperfine Research, Guilford, CT, United States, ³New York University, New York, NY, United States

We present a self-supervised approach for accelerated non-uniform MRI reconstruction, which leverages self-supervision in k-space and image domains. We evaluated the performance on both simulation and real data, where fully sampled data is unavailable. The experimental results on a non-uniform MRI dataset demonstrate that the proposed method can generate reconstruction that approaches the accuracy of the fully supervised reconstruction. Furthermore, we show that the approach can be applied to highly challenging real-world clinical MRI reconstruction acquired on a low-field (64 mT) MRI scanner with no data available for supervised training while demonstrating improved perceptual quality as compared to traditional reconstruction.

Muwei Li¹, Yurui Gao¹, Adam W Anderson¹, Zhaohua Ding¹, and John C Gore^1
1Vanderbilt University Institute of Imaging Science, Nashville, TN, United States

We investigated voxel-wise spectrograms derived from the time-varying spectral patterns of BOLD signals in white matter. Finite-window spectra are non-stationary but may be categorized into five distinct modes that recur over time. Close scrutiny of the signal profiles reveals distinct spatial distributions of the occurrences and durations of these modes, and such distributions are highly consistent across individuals. In addition, two communities of white matter voxels may be identified according to a hierarchical arrangement of transitions among modes. Our findings reveal the non-stationary nature of BOLD spectral patterns, and provide a novel spatial-temporal-frequency characterization of resting-state signals in white matter.

Kanghyun Ryu¹, Zhitao Li¹, Christopher M. Sandino², and Shreyas S. Vasanawala^1
1Radiology, Stanford University, Stanford, CA, United States, ²Electrical Engineering, Stanford University, Stanford, CA, United States

Deep Learning based reconstruction methods have been vastly explored for accelerating Cartesian-based cardiac cine imaging via using unrolled neural networks. However, for non-Cartesian trajectories such as radial, these networks require substantial modifications (i.e., NUFFT-based data consistency) and requires collecting separate radial-based dataset, which may not be common in the clinics. 

Here, we investigate a method to transfer the radial k-space data to the Cartesian domain using GROG-based interpolation. We show that DL-ESPIRiT trained with Cartesian cine dataset (with pseudo radial-like under sampling pattern) can be generalizable to reconstruct actual accelerated radial cine acquired on a scanner.

Ruiying Liu¹, Jee Hun Kim², Chaoyi Zhang¹, Hongyu Li¹, Peizhou Huang¹, Xiaojuan Li², and Leslie Ying^1
1Department of Biomedical Engineering, Department of Electrical Engineering, University at Buffalo, Buffalo, NY, United States, ²Program of Advanced Musculoskeletal Imaging (PAMI), Cleveland Clinic, Cleveland, OH, United States

Most parallel imaging methods require calibration data for reconstruction. Low-rank-based methods allow calibration-free reconstruction from randomly undersampled, multi-channel data. This abstract presents a novel reconstruction method to combine multichannel blind deconvolution (MALBEC), a calibration-free method, and GRAPPA, a calibration-based method for highly accelerated imaging. The method sequentially performs MALBEC and GRAPPA with specially designed sampling masks such that the benefits of low-rank structure and the availability of calibration data can be utilized jointly. Our results demonstrate that the proposed method can achieve an acceleration factor that is the product of the factors achieved by MALBEC and GRAPPA alone. 

Kanghyun Ryu¹, Cagan Alkan², and Shreyas S. Vasanawala^1
1Radiology, Stanford University, Stanford, CA, United States, ²Electrical Engineering, Stanford, Stanford, CA, United States

In this study, we propose a novel refinement method using auto-calibrated k-space null-space kernel to reduce the k-space errors and enable reconstruction of improved high-frequency image details and textures. The refinement algorithm can be easily plugged in after DL-based reconstructions. We show that our method enables the reconstruction of sharper images with significantly improved high-frequency components measured by HFEN and GMSD while maintaining overall error in the image measured by PSNR and SSIM.

Arghya Pal¹ and Yogesh Rathi^2
1Department of Psychiatry, Harvard Medical School, Boston, MA, United States, ²Harvard Medical School, Boston, MA, United States

Deep learning methods are increasingly being used for accelerating magnetic resonance imaging (MRI) acquisition and reconstruction strategies. However, it is important to understand whether or not deep learning models have an inherent structural bias that may effectively create concerns in real-world settings. In this abstract, we show a strategy to understand and then reduce structural prior bias in deep models. The proposed approach decouples the spurious structural bias (prior) of a deep learning model by intervening in the input. Our proposed debiasing strategy is fairly robust and can work with any pre-trained deep learning MR reconstruction model.

Zachary Miller¹ and Kevin Johnson^1
1University of Wisconsin-Madison, Madison, WI, United States

The goal of this work is to reconstruct high resolution, volumetric non-cartesian acquisitions using model based deep learning. This is a challenging problem because unlike cartesian acquisitions and hybrid trajectories like stack of spirals that can be decoupled into 2D sub problems, fully non-cartesian trajectories require the complete 3D volume for data-consistency pushing GPU memory capacity to its limits even using cluster computing. Here, we investigate blockwise learning, a method that has the memory saving benefits of patch-based methods while maintaining data-consistency in k-space. We demonstrate the feasibility of this method by reconstructing high resolution pulmonary non-contrast and MRA acquisitions.

Sizhuo Liu¹, Philip Schniter¹, and Rizwan Ahmad^1
1The Ohio State University, Columbus, OH, United States

Plug-and-play (PnP) methods can reconstruct images by employing iterative algorithms that leverage the knowledge of the forward model and a sophisticated denoiser. The performance of PnP can be improved by utilizing an application-specific denoiser. However, training such denoisers may not be feasible for many MRI applications. Here, we describe a PnP-inspired method that does not require data beyond the single, incomplete set of measurements. The proposed method, called recovery with a self-calibrated denoiser (ReSiDe), trains the denoiser from the patches of the image being recovered. For validation, ReSiDe is applied to T1-weighted brain and myocardial first-pass perfusion data.

Meiyappan Solaiyappan¹, Santosh Kumar Bharti¹, Mohamad Dbouk², Wasay Nizam³, Malcolm V. Brock^3,4, Michael G. Goggins^2,4,5, and Zaver M. Bhujwalla^1,2,6
1Department of Radiology, The Johns Hopkins University School of Medicine, Baltimore, MD, United States, ²Department of Pathology, The Johns Hopkins University School of Medicine, Baltimore, MD, United States, ³Department of Surgery, The Johns Hopkins University School of Medicine, Baltimore, MD, United States, ⁴Department of Oncology, The Johns Hopkins University School of Medicine, Baltimore, MD, United States, ⁵Department of Medicine, The Johns Hopkins University School of Medicine, Baltimore, MD, United States, ⁶Radiation Oncology and Molecular Radiation Sciences, The Johns Hopkins University School of Medicine, Baltimore, MD, United States

Early detection of cancers using blood-based analytes for routine screening is a rapidly advancing area. Here, we developed a neural-network based solution to detect pancreatic ductal adenocarcinoma (PDAC) and non-small cell lung cancer (NSCLC) with high sensitivity and specificity using human plasma and serum samples to discriminate between subjects with no known pancreatic or lung disease, subjects with benign disease and subjects with PDAC or NSCLC.

Siddharth Srinivasan Iyer^1,2, Frank Ong³, and Kawin Setsompop^2,3
1Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ²Department of Radiology, Stanford University, Stanford, CA, United States, ³Department of Electrical Engineering, Stanford University, Stanford, CA, United States

This work aims to accelerate the convergence of proximal gradient algorithms (like FISTA) by designing a preconditioner using polynomials that targets the eigenvalue spectrum of the forward linear model to enable faster convergence. The preconditioner does not assume any explicit structure and can be applied to various imaging applications. The efficacy of the preconditioner is validated on four varied imaging applications, where it seen to achieve 2-4x faster convergence.

Nolan K Meyer^1,2, Myung-Ho In², Daehun Kang², Yunhong Shu², John Huston III², Matt A Bernstein², and Joshua D Trzasko^2
1Mayo Clinic Graduate School of Biomedical Sciences, Rochester, MN, United States, ²Department of Radiology, Mayo Clinic, Rochester, MN, United States

Echo planar imaging (EPI) is commonly used clinically for its speed, but is sensitive to non-idealities including system  field inhomogeneity and subject-specific susceptibility effects. Multi-shot techniques encoding an auxiliary point-spread-function (PSF)-encoding dimension provide robustness to off-resonance effects, with demonstrated potential for diffusion and anatomic applications, but incur scan time penalties motivating acquisition and reconstruction strategies to increase acceleration. This work proposes a comprehensive model-based iterative reconstruction framework for anatomic PSF-encoded EPI scans incorporating low-rank constraints, directly reconstructing undistorted images from undersampled data. Advantages are demonstrated using brain MRI data acquired on a compact 3T MRI system.

Ajin Joy¹, Rajakumar Nagarajan¹, Andres Saucedo¹, Zohaib Iqbal¹, Manoj K Sarma¹, Neil E Wilson¹, Ely R Felker¹, Robert E Reiter², Steven S Raman¹, and M Albert Thomas^1
1Radiological Sciences, University of California-Los Angeles, Los Angeles, CA, United States, ²Urology, University of California-Los Angeles, Los Angeles, CA, United States

Prospectively undersampled 5D echo-planar J-resolved spectroscopic imaging (EP-JRESI) data were acquired in 9 prostate cancer patients and 3 healthy controls. The 5D data  was reconstructed using Dictionary learning (DL), Total Variation (TV), Perona-Malik (PM) and a hybrid DLTV method combining DL and TV. DLTV uses the gradient sparsity of TV and the learned dictionary-based sparsity of DL to further increase the transform sparsity of the data. The DLTV method unambiguously resolved 2D J-resolved peaks including myo-inositol, citrate, creatine, spermine and choline with an improved reconstruction that facilitates higher acceleration factors, leading to significant reduction in scan time.

Nolan K Meyer^1,2, Daehun Kang², Zaki Ahmed², Myung-Ho In², Yunhong Shu², John Huston III², Matt A Bernstein², and Joshua D Trzasko^2
1Mayo Clinic Graduate School of Biomedical Sciences, Rochester, MN, United States, ²Department of Radiology, Mayo Clinic, Rochester, MN, United States

Functional magnetic resonance imaging (fMRI) has inherent limitations of fast acquisitions due to low signal-to-noise ratio (SNR) and artifacts. Multi-echo fMRI acquires images at multiple TEs, increasing robustness to off-resonance based signal loss and improving sensitivity to neural activity. However, noise is substantial given limitations of EPI, although consistent across TEs contributing to prolonged scan times required for sufficient statistical power. Reduction of acquisition time with reconstruction and processing techniques is of interest. This study extends preliminary work on locally low-rank denoising of multi-echo fMRI data to explore scan time reduction through processing with retrospective truncation.

Avrajit Ghosh¹, Shijun Liang², Anish Lahiri ³, and Saiprasad Ravishankar ^1,2
1Department of Computational Mathematrics Science and Engineering, Michigan State University, East Lansing, MI, United States, ²Department of Biomedical Engineering, Michigan State University, East Lansing, MI, United States, ³Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, United States

This work examines optimized parallel combinations of deep networks and conventional regularized reconstruction for improved quality of MR image reconstructions from undersampled k-space data. Features learned by deep networks and typical model-based iterative algorithms (e.g., sparsity-penalized reconstruction) could complement each other for effective reconstructions. We observe that combining the image features from multiple approaches in a parallel fashion with appropriate learned weights leads to more effective image representations that are not captured by either strictly supervised or (unsupervised) conventional iterative methods.