Ramin Jafari¹, Richard K G Do², Yousef Mazaheri Tehrani^1,2, Ty Cashen³, Sagar Mandava³, Maggie Fung³, Ersin Bayram³, and Ricardo Otazo^1,2
1Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, United States, ²Radiology, Memorial Sloan Kettering Cancer Center, New York, NY, United States, ³GE Healthcare, Waukesha, WI, United States

To use deep learning to reconstruct motion-resolved dynamic images from multicoil undersampled radial data without image quality degradation and 800-fold reduction in reconstruction time compared to the iterative XD-GRASP algorithm.

Lu Wang¹, Zhen Xing², Jian Wu¹, Qinqin Yang¹, Congbo Cai¹, Shuhui Cai ¹, Zhong Chen¹, and Dairong Cao^2
1Department of Electronic Science, Xiamen University, Xiamen, Fujian, China, ²Department of Radiology, First Affiliated Hospital of Fujian Medical University, Fuzhou, Fujian, China

Intravoxel incoherent motion (IVIM) imaging is a non-invasive MR perfusion imaging that could prevent patients from the harm of exogenous reagent. Previous studies proved that the least square and Bayesian approaches are so far the best algorithms in IVIM fitting. However, they still suffer from time-consuming and high noise level. We proposed a deep neural network-based reconstruction method with synthetic training data for IVIM imaging and extended it to hybrid IVIM-DKI (diffusion kurtosis imaging) model fitting. Experimental results show that our method owns prominent performance in both image quality and accuracy of fitting results with a remarkably short reconstruction time.

Dipika Sikka^1,2, Noah Igra^3,4, Sabrina Gjerswold-Sellec¹, Cynthia Gao⁵, Ed Wu⁶, and Jia Guo^7
1Department of Biomedical Engineering, Columbia University, New York, NY, United States, ²VantAI, New York, NY, United States, ³Department of Applied Mathematics, Columbia University, New York, NY, United States, ⁴Sackler School of Medicine, Tel Aviv University, Tel Aviv, Israel, ⁵Department of Computer Science, Columbia University, New York, NY, United States, ⁶Department of Electrical and Electronic Engineering, The University of Hong Kong, Pok Fu Lam, Hong Kong, China, ⁷Department of Psychiatry, Mortimer B. Zuckerman Mind Brain Behavior Institute, Columbia University, New York, NY, United States

While the application of deep learning in MR image analysis has gained significant popularity, using raw MR k-space data as part of deep learning analysis is an underexplored area. Here we develop a completely complex U-Net deep learning architecture, CU-Net, where we apply deep learning components and operations in the complex space. CU-Net leverages k-space MR signals while training a U-Net with Attention and Residual components, as opposed to using processed spatial (real) data, typically seen with MRI deep learning applications. As part of a proof-of-concept study, the complex networks demonstrated their utility and potential superiority over their spatial counterparts.

Sonoko Oshima¹, Yasutaka Fushimi¹, Satoshi Nakajima¹, Akihiko Sakata¹, Takuya Hinoda¹, Sayo Otani¹, Krishna Pandu Wicaksono¹, Hiroshi Tagawa¹, Yang Wang¹, Yuichiro Sano², Rimika Imai², Masahito Nambu², Koji Fujimoto³, Hitomi Numamoto⁴, Kanae Kawai Miyake⁴, Tsuneo Saga⁴, and Yuji Nakamoto^1
1Department of Diagnostic Radiology and Nuclear Medicine, Graduate School of Medicine, Kyoto University, Kyoto, Japan, ²Canon Medical Systems Corporation, Otawara, Japan, ³Department of Real World Data Research and Development, Graduate School of Medicine, Kyoto University, Kyoto, Japan, ⁴Department of Advanced Medical Imaging Research, Graduate School of Medicine, Kyoto University, Kyoto, Japan

We applied four patterns of deep learning reconstruction (DLR) denoising methods to 1 NEX neuromelanin-sensitive MR images. DLR with denoising intensity coefficient of 1.0 and edge enhancement off provided the best image quality among the four types of DLR, and it was significantly better than or as good as 5 NEX images. ROC analyses using images with all DLR patterns showed good AUCs for diagnosis of Parkinson’s disease. This DLR denoising method can improve image quality of neuromelanin-sensitive MRI with good diagnostic ability to differentiate patients with Parkinson’s disease from healthy controls.

Nicholas Dwork¹, Daniel O'Connor², Corey A. Baron³, Ethan M. I. Johnson⁴, John M. Pauly⁵, and Peder E.Z. Larson^6
1Radiology and Biomedical Imaging, UCSF, San Francisco, CA, United States, ²Mathematics and Statistics, University of San Francisco, San Francisco, CA, United States, ³Robarts Research, Western University, London, ON, Canada, ⁴Biomedical Engineering, Northwestern University, Evanston, IL, United States, ⁵Electrical Engineering, Stanford University, Stanford, CA, United States, ⁶Radiology and Biomedical Imaging, University of California in San Francisco, San Francisco, CA, United States

In this work, we present a modification of the standard implementation of compressed sensing that takes advantage of the structure of the Daubechies wavelet transform.  By doing so, we show that we retain additional detail in the reconstructed images when few data samples are used.

Milica Medved¹, Marco Vicari², and Gregory S Karczmar^1
1Department of Radiology, The University of Chicago, Chicago, IL, United States, ²Fraunhofer MEVIS, Bremen, Germany

Compressed sensing (CS) was evaluated as an acceleration technique for high spectral and spatial resolution (HiSS) MRI, at acceleration factors up to R=10.  Effective spatial resolution was maintained in the readout direction, and decreased with R in the phase encoding direction, although acceleration factors of up to R = 4 are realistic. Noise amplification was not observed.  CS could improve diagnostic utility of HiSS MRI in breast by allowing longer echo trains and thus heavier T2* weighting in a fewer number of k-space lines.  CS could also facilitate use of HiSS MRI in geometrically constrained applications, such as prostate MRI.

Pak Lun Kevin Ding¹, Riti Paul¹, Baoxin Li¹, Ameet C. Patel², and Yuxiang Zhou^2
1CIDSE, Arizona State University, Tempe, AZ, United States, ²RADIOLOGY, Mayo Clinic College of Medicine, Tempe, AZ, United States

Conventional Magnetic Resonance Imaging (MRI) is a prolonged procedure. Therefore, it’s beneficial to reduce scan time as it improves patient experience and reduces scanning cost. While many approaches have been proposed for obtaining high quality reconstruction images using under-sampled k-space data, deep learning has started to show promising results when compared with conventional methods. In this paper, we propose a Variational Feedback Network (VFN) for accelerated MRI reconstruction. Specifically, we extend the previously proposed variational network with recurrent neural network (RNN). Quantitative and qualitative evaluations demonstrate that our proposed model performs superiorly against other compared methods on MRI reconstruction.

Falk Christian Mayer^1,2, Peter Bachert^1,2, Mark E Ladd^1,2,3, and Benjamin Knowles^1
1Medical Physics in Radiology, German Cancer Research Center (DKFZ), Heidelberg, Germany, ²Department of Physics and Astronomy, Heidelberg University, Heidelberg, Germany, ³Faculty of Medicine, Heidelberg University, Heidelberg, Germany

A GPU-based Fast non-Uniform Fast Fourier Transform (NUFFT) was implemented, which uses optimized libraries and algorithms such as stencils and managed memory to perform highly efficient transformations from the image to the k-space domain and vice-versa. In testing, the transform execution time was measured to be approximately 1ms for 32 channel data and a 256x256 grid size. A conjugate gradient based solution to the inverse NUFFT was also implemented, in which a solution was given within approximately 10ms. The proposed implementation has valuable applications for non-Cartesian real-time imaging.

Nicholas Dwork^1
1Radiology and Biomedical Imaging, UCSF, San Francisco, CA, United States

We present a model based reconstruction algorithm with low-rank and joint sparse regularization terms for use with hyperpolarized MRI.  Results show improved details.

Joshua Michael McAteer¹, Olivier Mougin¹, James Harkin², Paul Glover¹, and Penny Gowland^1
1Physics, University of Nottingham, Nottingham, United Kingdom, ²Medicine, University of Nottingham, Nottingham, United Kingdom

Optimising compressed sensing sampling can yield a significant increase in measured and observed image quality over heuristic sampling methods. Using an example image, such as a previous acquired scan of the same anatomy, a bespoke sampling pattern can be designed that optimally samples the data. This increase in image quality allows for greater acceleration or better SNR with the same imaging time over standard methods such as variable density Poisson disc sampling. This has been tested on a 0.5T upright MR scanner in phantoms.

Astrid van Lier¹, Yulia Shcherbakova¹, and Cornelis van den Berg^1
1UMC Utrecht, Utrecht, Netherlands

Phase cycled bGRE images disentangled into configuration modes can be used to generate alternative signal contrasts for for example for radiotherapy delineation purposes while simultaneously geometrical errors due to B0 inhomogeneity can be quantified. Numerical simulations show the change in modulation for different configurations states. In vivo experiments on the male pelvis are used to illustrate the changed image contrast and obtained B0 map against the reference method.
 

Qing-San Xiang^1,2
1Radiology, University of British Columbia, Vancouver, BC, Canada, ²Physics and Astronomy, University of British Columbia, Vancouver, BC, Canada

Susceptibility Weighted Imaging (SWI) has many applications.  One crucial step of SWI is background phase error removal that typically involves smoothing either before or after phase unwrapping. Both approaches may have to face difficulties when large isolated phase loops (around poles or singularities) are present in the image.  In this work, background correction with phase diffusor (BACOPSOR) is introduced. It is straightforward to implement and has a desirable immunity to phase loops. Application of BACOPSOR to SWI has been demonstrated with in vivo data.

satoshi ITO¹, taro SUGAI¹, kohei TAKANO¹, and shohei OUCHI^1
1Utsunomiya University, Utsunomiya, Japan

To improve the denoising performance of a convolutional neural network (CNN), a parallelized blind image denoising (ParBID) was proposed and demonstrated. ParBID procedure is similar to SENSE technique, 1) linear combination of adjacent 2D sliced noisy images, 2) blind noise level CNN denoising, and 3) separation of linearly combined and denoised images by solving linear equation. Experimental studies showed that the PSNR and the SSIM were improved for all noise levels, from 2.5% to 7.5%. ParBID showed that the greatest PSNR improvements were obtained when three slice images were used for linear image combination.

Yuxin Yang¹, Xi Xu¹, Yuanyuan Liu¹, Yanjie Zhu^1,2, Dong Liang^1,2,3, and Hairong Zheng^1,2
1Shenzhen Institute of Advanced Technology,Chinese Academy of Sciences, Shenzhen, China, ²Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen, China, Shenzhen, China, ³Research Centre for Medical AI, Shenzhen Institutes of Advanced Technology, Chinese Academy of Science, Shenzhen, China, Shenzhen, China

An optimization aimed at shortening pulse durations was carried out for three types of adiabatic spin-lock pulses by means of Bloch simulation. The variance of a part of the trajectory of Mz with respect to a range of off-resonance values was calculated to find the optimal pulse parameters and decent T1ρ-weighted imaging and T1ρ mapping results were obtained.

Corey Edward Cruttenden¹, Wei Zhu¹, Yi Zhang¹, Xiao-Hong Zhu¹, Rajesh Rajamani², and Wei Chen^1
1Center for Magnetic Resonance Research, University of Minnesota, Minneapolis, MN, United States, ²Mechanical Engineering, University of Minnesota, Minneapolis, MN, United States

Acquiring neuro-electrophysiology signal simultaneously with fMRI is hindered by electromagnetic field interactions that generate artifacts, including fMRI gradient induced artifacts in the neuro-electrophysiology data. This abstract presents a novel method using a separation boundary on the singular value decomposition of the first difference of artifact-contaminated data to accurately reconstruct clean neural signals. The separation boundary can be estimated from a brief baseline recording period followed by simultaneous fMRI and neuro-electrophysiological data acquisition. The method is successfully demonstrated on neural recording data acquired simultaneously with time-series echo planar imaging at 16.4T. 

Chang Sun¹, Roido Manavaki¹, Jason Tarkin², Christopher Wall², James HF Rudd², Fiona J Gilbert¹, and Martin J Graves^1
1Department of Radiology, University of Cambridge, Cambridge, United Kingdom, ²Division of Cardiovascular Medicine, University of Cambridge, Cambridge, United Kingdom

Bias correction in the thoracic region is challenging due to the low proton density in the lung. Traditional retrospective bias correction techniques, such as surface fitting method and the histogram-based method, suffer from over suppression in the lung regions or increased noise in the tissue. We propose a hybrid bias correction method that combines the advantages of the surface fitting and the histogram-based methods. The hybrid method normalized the signal intensity in lung and reduced the signal variation in tissue.

Rolf F Schulte¹, Mary A McLean², Joshua D Kaggie², Stephan Ursprung², Ramona Woitek², Ferdia A Gallagher², Esben S S Hansen³, Nikolaj Bogh³, and Christoffer Laustsen^3
1GE Healthcare, Munich, Germany, ²Department of Radiology, University of Cambridge, Cambridge, United Kingdom, ³MR Research Centre, University of Aarhus, Aarhus, Denmark

Multi-channel receive coils can improve coverage for metabolic imaging with hyperpolarised ¹³C compounds. Combining different receive channels in an SNR-optimal way is challenging due to the difficulties in determining sensitivity maps. The main aim of this work was to implement and optimise a Singular-Value-Decomposition (SVD) based sensitivity map extraction from metabolic images with a single spectral point per metabolite and to investigate its performance in SNR-limited metabolic imaging experiments.

Catarina Rua^1,2, Alberto Llera^3,4, Olivier Mougin⁵, Mauro Costagli⁶, Renat Yakupov⁷, Richard Bowtell⁵, James B Rowe^1,8, and Christopher T Rodgers^9,10
1Department of Clinical Neurosciences and University of Cambridge Centre for Parkinson-plus, University of Cambridge, Cambridge, United Kingdom, ²Wolfson Brain Imaging Centre, University of Cambridge, cambridge, United Kingdom, ³Cognitive Neuroscience, Radboud University Medical Centre, Nijemen, Netherlands, ⁴Donders Institute, Centre for Cognitive Neuroimaging, Radboud University, Nijmegen, Netherlands, ⁵Sir Peter Mansfield Imaging Centre, School of Physics and Astronomy, University of Nottingham, Nottingham, United Kingdom, ⁶IMAGO7 Foundation, Pisa, Italy, ⁷German Centre for Neurodegenerative Diseases (DZNE), Magdeburg, Germany, ⁸Medical Research Council Cognition and Brain Sciences Unit, University of Cambridge, Cambridge, United Kingdom, ⁹Wolfson Brain Imaging Centre, University of Cambridge, Cambridge, United Kingdom, ¹⁰Department of Clinical Neurosciences, University of Cambridge, Cambridge, United Kingdom

Multi-site studies are an attractive option at ultra-high field (7T) as it is possible to pool larger number of datasets of healthy and patients increasing the statistical power of neuroimaging studies. However, differences on scanner hardware and software increase variability in the measurements obtained on across imaging sites.In this study we have piloted the application of a multimodal ICA approach for denoising scanner effects across two different 7T MRI scanner platforms.