Sidharth Kumar¹ and Jonathan I Tamir^1,2,3
1Electrical and Computer Engineering, University of Texas at Austin, Austin, TX, United States, ²Oden Institute for Computational Engineering and Sciences, University of Texas at Austin, Austin, TX, United States, ³Department of Diagnostic Medicine, University of Texas at Austin, Austin, TX, United States

Synthetic MRI has emerged as a tool for retrospectively generating contrast weightings from tissue parameter maps, but the generated contrasts can show mismatch due to unmodeled effects. We looked into the feasibility of refining arbitrary synthetic MRI contrasts using conditional GANs. To achieve this objective we trained a GAN on different experimentally obtained inversion recovery contrast images. As a proof of principle, the RefineNet is able to correct the contrast while losing out on finer structural details due to limited training data.

Tuneesh K Ranota¹, Fumin Guo², Tingting Wu³, Matthew S Fox⁴, and Alexei Ouriadov^3
1School of Biomedical Engineering, The University of Western Ontario, London, ON, Canada, ²Sunnybrook Research Institute, University of Toronto, Toronto, ON, Canada, ³Physics and Astronomy, The University of Western Ontario, London, ON, Canada, ⁴Physics and Astronomy, Lawson Health Research Institute, London, ON, Canada

Hyperpolarized ¹²⁹Xe MRI provides a way to investigate and assess pulmonary diseases.  To improve the performance of ¹²⁹Xe MRI in lung imaging, a method for acquiring two VDP calculations using high-resolution 3D static ventilation imaging from FGRE combined with a key-hole method was proposed and demonstrated in participants with ventilation abnormalities. SNR was calculated as well as semi-automated VDP and fully automated VDP, which showed a strong positive linear correlation with a zero intercept and close to unity slop. This study confirms the feasibility of using isotropic-voxel ¹²⁹Xe MRI acquired in a single breath-hold to study ventilation heterogeneity.

Neha Koonjoo^1,2, Bo Zhu^1,2, Danyal Bhutto^1,2,3, Suresh E Joel⁴, and Matthew S Rosen^1,2,5
1Department of Radiology, A.A Martinos Center for Biomedical Imaging / MGH, Charlestown, MA, United States, ²Harvard Medical School, Boston, MA, United States, ³Department of Biomedical Engineering, Boston University, Boston, MA, United States, ⁴GE Healthcare, Bangalore, India, ⁵Department of Physics, Harvard University, Cambridge, MA, United States

AUTOMAP has proved itself to be robust to noise especially in the low SNR regimes, however due to the fully connected architecture of the input layer, its applicability to large matrix size datasets has been limited. Here we propose a patched-based trained network that enables the reconstruction of larger datasets. Low SNR single-channel volume coil brain images were acquired at 1.5T with different pulse sequences and reconstructed with the trained model. Results obtained show significant denoising potential. An increase in SNR of 1.5-fold as well as an increase in SSIM was also observed.

Chang Gao^1,2, Shu-Fu Shih^1,3, Vahid Ghodrati^1,2, Paul Finn^1,2, Peng Hu^1,2, and Xiaodong Zhong^4
1Department of Radiological Sciences, University of California Los Angeles, Los Angeles, CA, United States, ²Department of Physics and Biology in Medicine, University of California Los Angeles, Los Angeles, CA, United States, ³Department of Bioengineering, University of California Los Angeles, Los Angeles, CA, United States, ⁴MR R&D Collaborations, Siemens Medical Solutions USA, Inc., Los Angeles, CA, United States

Deep learning-based undersampled MRI reconstruction generally requires the k-space data consistency term to constrain the output. However, this requires gridding onto a Cartesian basis for radial data, which slows down the training process profoundly and may even make it impractical. To avoid the repeated gridding process in training, we developed a transformer network to directly predict unacquired radial k-space spokes. The developed network was evaluated in vivo, accurately predicted the unacquired k-space spokes and generated better image intensity and less streaking artifacts compared to the undersampled images.

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

The goal of this work is to reconstruct highly undersampled frames from dynamic Non-Cartesian acquisitions using deep learning. In this setting, supervised data is difficult to obtain due to organ motion necessitating self-supervision.  The challenge is that acquisitions are often so data-starved that self-supervised reconstructions that use only spatial correlations fail to recover fine details. Here, we leverage correlations across time frames, and show that even when data is misaligned, it is possible to reconstruct highly accelerated frames using self-supervised  methods. We demonstrate the feasibility of this technique by reconstructing end-inspiratory phase images from respiratory binned Pulmonary UTE acquisitions.

Phillip Martin¹, Maria Altbach^2,3, and Ali Bilgin^1,2,3,4
1Electrical and Computer Engineering, University of Arizona, Tucson, AZ, United States, ²Biomedical Engineering, University of Arizona, Tucson, AZ, United States, ³Medical Imaging, University of Arizona, Tucson, AZ, United States, ⁴Applied Mathematics, University of Arizona, Tucson, AZ, United States

In this work we propose a novel algorithm of denoising accelerated diffusion weighted MRI (dMRI) acquisitions using deep learning and self-supervision. This method effectively enables the prediction of diffusion-weighted images (DWIs), without the need for large amounts of training data with high directional encodings. We demonstrate that accurate diffusion tensor metrics can be obtained with as few as 6 DWIs using only a few training datasets with high directional encodings.

Peizhou Huang¹, Chaoyi Zhang², Hongyu Li², Ruiying Liu², Xiaoliang Zhang¹, Xiaojuan Li³, Dong Liang⁴, and Leslie Ying^1,2
1Biomedical Engineering, State University of New York at Buffalo, Buffalo, NY, United States, ²Electrical Engineering, State University of New York at Buffalo, Buffalo, NY, United States, ³Program of Advanced Musculoskeletal Imaging (PAMI), Cleveland Clinic, Cleveland, OH, United States, ⁴Paul C. Lauterbur Research Center for Biomedical Imaging, SIAT, CAS, Shenzhen, China

In this abstract, we propose a novel reconstruction method, named DURED-Net, that enables interpretable unsupervised learning for MR image reconstruction by combining an unsupervised denoising network and a plug-and-play method. Specifically, we incorporate the physics model into the denoising network using Regularization by Denoising (RED), and unroll the underlying optimization into a deep neural network. In addition, a sampling pattern was specially designed to facilitate unsupervised learning. Experiment results based on the knee fastMRI dataset exhibit marked improvements over the existing unsupervised reconstruction methods.

Chang Gao^1,2, Vahid Ghodrati^1,2, Shu-Fu Shih^1,3, Holden H. Wu^1,2,3, Yongkai Liu^1,2, Marcel Dominik Nickel⁴, Thomas Vahle⁴, Brian Dale⁵, Victor Sai¹, Ely Felker¹, Qi Miao¹, Xiaodong Zhong⁶, and Peng Hu^1,2
1Department of Radiological Sciences, University of California Los Angeles, Los Angeles, CA, United States, ²Department of Physics and Biology in Medicine, University of California Los Angeles, Los Angeles, CA, United States, ³Department of Bioengineering, University of California Los Angeles, Los Angeles, CA, United States, ⁴MR Application Predevelopment, Siemens Healthcare GmbH, Erlangen, Germany, ⁵MR R&D Collaborations, Siemens Medical Solutions USA, Inc., Cary, NC, United States, ⁶MR R&D Collaborations, Siemens Medical Solutions USA, Inc., Los Angeles, CA, United States

Undersampling is desired to reduce scan time but can cause streaking artifacts in stack-of-radial imaging. State-of-the-art deep neural networks such as the U-Net can be trained in a supervised manner to remove streaking artifacts but produce blurred images and loss of image details. Therefore, we developed and trained a 3D generative adversarial network to preserve perceptual image sharpness while removing streaking artifacts. The network used a combination of adversarial loss, L2 loss and structural similarity index loss. We demonstrated the feasibility of the proposed network for removing streaking artifacts and preserving perceptual image sharpness.

Zhiyuan Li^1,2, Hailong Li^1,3, Nehal Parikh^1,4, Lili He^1,4, Adebayo Braimah¹, and Jonathan Dillman^1,5
1Imaging Research Center, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States, ²Electrical Engineering and Computer Science, University of Cincinnati, Cincinnati, OH, United States, ³The Perinatal Institute, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States, ⁴Pediatrics, University of Cincinnati, Cincinnati, OH, United States, ⁵Radiology, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States

Between 35-40% of very preterm infants (≤32 weeks’ gestational age) develop cognitive deficits, thereby increasing their risk for poor educational, health, and social outcomes. Timely and accurate identification of infants at risk soon after birth is desirable for early intervention allocation. We proposed a novel Ontology-guided Attribute Partitioning Ensemble Learning (OAP-EL) model using quantitative structural MRI data obtained soon after birth to predict cognitive deficits at 2 years corrected age in very preterm infants.    

Alyssa Bird¹, Francesco Caliva¹, Sharmila Majumdar¹, Valentina Pedoia¹, and Richard B Souza^1
1University of California, San Francisco, San Francisco, CA, United States

The aim of this study was to reduce time-to-segmentation for analysis of muscle quality across a large volume of the hip abductors in individuals with hip osteoarthritis by developing an automatic segmentation network. The gluteus medius (GMED), gluteus minimus (GMIN), and tensor fasciae latae (TFL) were manually segmented on fat-water separated IDEAL MR images in 44 subjects. A 3D V-Net was trained, validated, and tested using these manually segmented image volumes and resulted in a mean Dice coefficient of 0.94, 0.87, and 0.91 for GMED, GMIN, and TFL. The automatic segmentation network demonstrated strong performance and provides drastically reduced time-to-segmentation.

Eric Y. Pierre^1,2, Kieran O'Brien³, Thorsten Feiweier⁴, Josef Pfeuffer⁴, and Daniel Staeb^2
1The Florey Institute of Neuroscience, Melbourne, Australia, ²MR Research Collaborations, Siemens Healthcare Pty Ltd, Bayswater, Australia, ³MR Research Collaborations, Siemens Healthcare Pty Ltd, Brisbane, Australia, ⁴MR Applications Predevelopment, Siemens Healthcare GmbH, Erlangen, Germany

We propose an efficient, densely connected network to synthesize unacquired DW-volumes from other acquired DW-directions and b=0 mm²/s volumes, allowing acceleration of high-angular DWI shell acquisition by skipping some DW-directions altogether. 

 For training, we used a high-quality dataset of 20 HCP subjects with 90 DW-directions per subject at both b=1000 mm²/s and 3000 mm²/s. 40 HCP subjects were used for validation. 30 DW-directions were selected for input to reconstruct the other 60 missing target DW-directions.

Comparison with a linear-interpolation benchmark show improved fidelity of synthesized DW-volumes and FOD maps to a gold standard acquisition, for both b-values.

 

Nataliia Molchanova^1,2, Bénédicte Maréchal^1,2,3, Jean-Philippe Thiran^2,3, Tobias Kober^1,2,3, Jonas Richiardi³, and Till Huelnhagen^1,2,3
1Advanced Clinical Imaging Technology, Siemens Healthcare AG, Lausanne, Switzerland, ²Signal Processing Laboratory (LTS 5), Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland, ³Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland

We propose a novel refacing technique employing a 3D conditional generative adversarial network to allow protecting subject privacy while maintaining consistent post-processing results. We evaluate the method compared to current refacing techniques using brain morphometry as an example. Results show that the proposed method compromises brain morphometry results to a lesser extent than existing methods while showing lower similarity of the final image to the original one, hence suggesting an improved privacy protection. We conclude that the proposed method represents a fast and viable alternative for image data de-identification compared to currently existing methods.