Nikolai J Mickevicius¹ and Carri K Glide-Hurst^1
1Department of Human Oncology, University of Wisconsin-Madison, Madison, WI, United States

To reduce scan time, methods to accelerate hybrid phase encoded/non-Cartesian MR fingerprinting (MRF) acquisitions for variable density spiral acquisitions have recently been developed. These methods are not applicable to MRF acquisitions wherein a single k-space spoke is acquired per frame. Therefore, we describe, implement, and test a low-rank inversion method on a 0.35T MR-guided radiation therapy system to resolve MR fingerprinting contrast dynamics from through-plane accelerated Cartesian/radial measurements.

David Parra¹, Phillip Martin², Maria Altbach^3,4, and Ali Bilgin^2,3,4,5
1Computer Science, University of Arizona, Tucson, AZ, United States, ²Electrical and Computer Engineering, University of Arizona, Tucson, AZ, United States, ³Biomedical Engineering, University of Arizona, Tucson, AZ, United States, ⁴Medical Imaging, University of Arizona, Tucson, AZ, United States, ⁵Applied Mathematics, University of Arizona, Tucson, AZ, United States

The aim of this work is to investigate the used of Transformer architectures in radial image reconstruction. While most deep-learning image reconstruction methods are based on convolutional neural networks (CNNs), recent advances in computer vision suggest that Transformer architecture may provide a favorable alternative in many vision tasks. In this work, we demonstrate that Transformer architectures can be used for sinogram interpolation and yield results comparable to CNNs.

Zhengnan Huang^1,2, Jonghyun Bae^1,2, Eddy Solomon³, Linda Moy^1,2, Sungheon Gene Kim³, Patricia M. Johnson¹, and Florian Knoll^1,4
1Center for Biomedical Imaging, Radiology, New York University School of Medicine, New York, NY, United States, ²Vilcek Institute of Graduate Biomedical Sciences, New York University School of Medicine, New York, NY, United States, ³Department of Radiology, Weill Cornell Medical College, New York, NY, United States, ⁴Friedrich-Alexander-University of Erlangen-Nürnberg, Erlangen, Germany

We proposed to use a variational network (VN) reconstruction algorithm with a 1-dimensional convolutional neural network (CNN) as a temporal regularizer for DCE-MRI reconstruction in this study. We used our newly developed breast perfusion simulation pipeline, to generate simulate data and train the reconstruction model. The machine learning (ML) reconstruction shows non-inferior structural similarity and improved visual image quality when compared with the iGRASP reconstruction. The ML reconstruction also takes much less time than the iGRASP reconstruction.

Zihan Wang^1,2, Jayse M Weaver^2,3, Daiki Tamada, PhD^2,3, Diego Hernando, PhD^2,3, and Scott B Reeder, MD, PhD^1,2,3,4,5
1Biomedical Engineering, University of Wisconsin-Madison, Madison, WI, United States, ²Radiology, University of Wisconsin-Madison, Madison, WI, United States, ³Medical Physics, University of Wisconsin-Madison, Madison, WI, United States, ⁴Medicine, University of Wisconsin-Madison, Madison, WI, United States, ⁵Emergency Medicine, University of Wisconsin-Madison, Madison, WI, United States

In this work, three methods for measuring signal-to-noise ratio (SNR) performance of deep learning (DL) denoising image reconstruction were evaluated. Images reconstructed using a vendor prototype DL reconstruction algorithm were compared to conventional Fourier reconstruction. Phantom experiments were performed using a single-channel head coil and a 32-channel head coil to assess the effects of parallel imaging acceleration in combination with DL reconstruction on SNR performance. We found excellent agreement between the three SNR measurement methods for both Fourier and DL reconstruction, and found that DL reconstructed images have similar g-factor performance patterns as Fourier reconstructed images.

Yanchen Guo¹, Shichun Chen¹, Li Zhao², Manuel Taso³, David C. Alsop³, and Weiying Dai^1
1Computer Science, State University of New York at Binghamton, Binghamton, NY, United States, ²Zhejiang University, Hangzhou, China, ³Beth Israel Deaconess Medical Center and Harvard Medical School, Boston, MA, United States

Compressed sensing allowed speeding up MRI image acquisition by several folds but with increased reconstruction time. Deep-learning was introduced for fast reconstruction with comparable or improved reconstruction quality. However, its applications on non-cartesian grids are scarce, mostly based on simulated resampling from original Cartesian grids. Here, we evaluated the performance of two deep-learning networks for reconstructing images acquired with k-space spiral trajectories in real MRI scanning. We demonstrated that deep learning networks can successfully reconstruct high-resolution images from under-sampled spiral trajectories with half of data in k-space and their performance is robust to spiral trajectories different from training.

Charles John Marchini¹ and Brad Sutton ^1
1Bioengineering, University of Illinois Urbana-Champaign, Urbana, IL, United States

ASL methods result in low SNR. SVD denoising can be used to increase the temporal SNR of ASL data by throwing out the low energy singular values. The SVD can be done after a standard reconstruction, or as we show here during the reconstruction if partial separability is used. The incorporation of the SVD into the reconstruction algorithm allows for a higher tSNR within the gray matter. It also paves the path for future work in which subsampling the data results in higher temporal and spatial resolution ASL.

Shouchang Guo¹, Jeffrey A. Fessler¹, and Douglas C. Noll^2
1Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI, United States, ²Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States

Inspired by two open questions of dynamic MRI reconstruction, we propose a novel voxel-wise attention network for temporal modeling for the undersampled reconstruction. The voxel-wise design of the network enables voxel-wise training, and we further propose a two-stage training scheme that pretrains the network with voxel-wise simulated data when dynamics are easy to obtain with physical models. With a factor of 12 undersampling, our proposed model outperforms other reconstructions with higher PSNR and better fMRI performance.

Molin Zhang¹, Junshen Xu¹, Yamin Arefeen¹, and Elfar Adalsteinsson^1,2,3
1EECS, MIT, Cambridge, MA, United States, ²Harvard-MIT Health Sciences and Technology, MIT, Cambridge, MA, United States, ³Institute for Medical Engineering and Science, MIT, Cambridge, MA, United States

Resolving a time series of T2-weighted images from a fast-spin-echo (FSE) sequence with traditional techniques requires long acquisitions, but T2-shuffling enables clinically feasible scan times by combining subspace models, which reduce degrees-of-freedom, with random spatial and temporal undersampling. Supervised machine learning achieves impressive reconstruction, but lack of labeled training data preclude its use in reconstructing signal dynamics. Recent zero-shot-self-supervised-learning (ZSSS) techniques enable high quality structural MRI reconstruction without training data. In this work, we combine ZSSS with the subspace model to further accelerate 2D T2-shuffling acquisitions. Our ZSSS-subspace models show significant reconstruction improvement in comparison to standard T2-shuffling in simulation. 

Jiayang Wang¹ and Justin P. Haldar^1
1Signal and Image Processing Institute, Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, United States

This work investigates the potential value of combining non-uniform k-space averaging with advanced nonlinear image denoising-reconstruction methods in the context of low-SNR MRI.  A new data-driven strategy for optimizing the k-space averaging pattern is proposed, and is then applied to total variation and U-net reconstruction methods.  It is observed that non-uniform k-space averaging (with substantially more averaging at the center of k-space) is preferred for both reconstruction approaches, although the distribution of averages varies substantially depending on the noise level and the reconstruction method.  We expect that these results will be informative for a wide range of low-SNR MRI applications.

Victoria Liu¹, Kanghyun Ryu², Cagan Alkan², John Pauly², and Shreyas Vasanawala^2
1California Institute of Technology, Pasadena, CA, United States, ²Stanford University, Stanford, CA, United States

Model-based accelerated MRI reconstruction networks leverage large datasets to reconstruct diagnostic-quality images from undersampled k-space. To deal with inherent dataset variability, the current paradigm trains separate models for each dataset. This is a demanding process and cannot exploit information that may be shared amongst datasets. In response, we propose multi-task learning (MTL) schemes that jointly reconstruct multiple datasets. Introducing inductive biases to the network allows for positive information sharing. We test MTL architectures and weighted loss functions against single task learning (STL). Our results suggest that MTL can outperform STL across a range of dataset ratios for two knee contrasts.

W. Scott Hoge^1,2 and Jonathan R. Polimeni^2,3
1Radiology, Brigham and Women's Hospital, Boston, MA, United States, ²Dept of Radiology, Harvard Medical School, Boston, MA, United States, ³Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA, United States

The structural relationship between the linear system of equations used to calibrate inplane-GRAPPA, slice-GRAPPA, and LeakBlock for SMS is analyzed. This analysis reveals that LeakBlock is structurally identical to inplane-GRAPPA in the specific case of a (synthetic) zero-slice-gap. Signal leakage in the original slice-GRAPPA formulation is thus revealed to be due to signal cancellation in the original formulation. This relationship between inplane-GRAPPA and LeakBlock further justifies its preferred usage over the original slice-GRAPPA formulation for the separation of slice-accelerated data.

Silu Han¹, Chidi Patrick Ugonna¹, and Nan-kuei Chen^1,2
1Biomedical Engineering Department, The University of Arizona, Tucson, AZ, United States, ²Medical Imaging Department, The University of Arizona, Tucson, AZ, United States

An integrated post-processing algorithm has been developed for effectively removing various types of aliasing artifact due to consecutive echo asymmetry, in-plane k-space under-sampling, through-plane acceleration with multi-band (MB) imaging and motion-induced phase variations in EPI based fMRI data. This algorithm uses a novel two-dimensional (2D) coil-signature-based phase-cycled correction method for 2D Nyquist artifact removal without calibration scans, and is capable of simultaneously removing: (1) aliasing artifacts due to in-plane and through-plane acceleration in single-shot and multi-shot EPI; (2) motion-induced phase errors in multi-shot EPI. Experimental results illustrate the effectiveness of the developed method, and its successful application to fMRI studies.

Andres Saucedo¹ and M. Albert Thomas^1
1Radiological Sciences, University of California, Los Angeles, Los Angeles, CA, United States

Radial echo planar spectroscopic imaging (REPSI)  is applied in healthy prostate and compared to Cartesian EPSI acquisitions.  Due to the small spatial extent of the anatomy-of-interest, REPSI is well-suited for acceleration in a reduced-field-of-view context. In this proof-of-concept study, we acquired both prospectively and retrospectively undersampled REPSI and Cartesian EPSI datasets in prostate phantom and in vivo, at multiple acceleration factors, and we compared both quantitative and qualitative results. 

Jo Schlemper¹, Neel Dey², Seyed Sadegh Mohseni Salehi¹, Kevin Sheth³, W. Taylor Kimberly⁴, and Michal Sofka^1
1Hyperfine, Guilford, CT, United States, ²New York University, New York City, NY, United States, ³Yale University, New Haven, CT, United States, ⁴Massachusetts General Hospital, Boston, MA, United States

In clinical low-field MRI, prolonged data acquisition is impractical, limiting the achievable SNR during imaging. In the absence of ground truth, unsupervised denoising is desirable, but many of them underperform on correlated noise structure of reconstructed MR images. In this work, we present an effective two step training framework for removing correlated MR noise without ground truth. We demonstrate that the proposed approach outperforms the existing denoising methods when applied to the low-field (64mT) diffusion-weighted images and demonstrate that significant noise reduction is possible. 82.5% of processed images were expertly rated clearly/far better overall.