Kanghyun Ryu¹, Christopher M. Sandino¹, Zhitao Li¹, Xucheng Zhu², Andrew Coristine³, Martin Janich⁴, and Shreyas S. Vasanawala^1
1Stanford University, Stanford, CA, United States, ²GE Healthcare, Menlo Park, CA, United States, ³GE Healthcare, Montreal, QC, Canada, ⁴GE Healthcare, Munich, Germany

Recently, unrolled neural networks (UNNs) have been shown to improve reconstruction over conventional Parallel Imaging Compressed Sensing (PI-CS) methods for dynamic MR image reconstruction. In this work we propose two methods to improve UNN for Non-Cartesian cardiac cine image reconstruction, namely memory efficient training and Convolutional LSTM based network architecture.The proposed method can significantly improve conventional UNN with higher image quality.

Manuel A. Morales^1,2, Maaike van den Boomen^2,3,4, Christopher Nguyen^2,4, Jayashree Kalpathy-Cramer², Bruce R. Rosen^1,2, Collin Stultz ^1,5,6, David Izquierdo-Garcia^1,2, and Ciprian Catana^2
1Health Sciences & Technology, Massachusetts Institute of Technology, Cambridge, MA, United States, ²Radiology, Athinoula A. Martinos Center for Biomedical Imaging, MGH, HMS, Charlestown, MA, United States, ³Radiology, University Medical Center Groningen, Groningen, Netherlands, ⁴Cardiovascular Research Center, MGH, HMS, Charlestown, MA, United States, ⁵Cardiology, Massachusetts General Hospital, Boston, MA, United States, ⁶Electrical Engineering and Computer Science, MIT, Cambridge, MA, United States

Myocardial strain analysis from cinematic magnetic resonance imaging (cine-MRI) data is a promising technique for earlier detection of subclinical dysfunction prior to reduction in left-ventricular ejection fraction (LVEF), but sources of discrepancies including user-related variations have limited its wide clinical adoption. Using healthy and cardiovascular disease (CVD) subjects (n=150) we developed a fast, user-independent deep-learning-based workflow for strain analysis from cine-MRI data. Relative to a reference tagging-MRI method, there was no significant difference in end-systolic global strain based on subject-paired cine-MRI data from 15 heathy subjects. Applications in CVD subjects without reduced LVEF showed both global and asymmetric strain abnormalities.

Hossam El-Rewaidy^1,2, Rui Guo¹, and Reza Nezafat^1
1Medicine, Beth Israel Deaconess Medical Center and Harvard Medical School, Boston, MA, United States, ²Graduate School of Bioengineering, Department of Computer Science, Technical University of Munich, Munich, Germany

In this work, we developed and evaluated a rapid (4-5 heartbeats) myocardial T₁ mapping approach by estimating voxel-wise T₁ values from one look-locker (LL) experiment of MOLLI sequence using a fully-connected neural network (MyoMapNet). MyoMapNet consists of 5 hidden layers that map the input 4-5 T₁-weighted samplings and their inversion times into T₁ values. MyoMapNet was trained and evaluated on a large dataset of native MOLLI-5(3)3 T₁ in 717 subjects and post-contrast MOLLI-4(1)3(1)2 in 535 subjects. MyoMapNet showed similar T₁ estimations to MOLLI-5(3)3 and MOLLI-4(1)3(1)2 T₁ (mean difference=1±17ms, and -3±18ms, respectively, p-value >0.1 for both).

Thomas Küstner^1,2, Alina Psenicny¹, Camila Munoz¹, Niccolo Fuin³, Aurelien Bustin⁴, Haikun Qi¹, Radhouene Neji^1,5, Karl P Kunze^1,5, Reza Hajhosseiny¹, Claudia Prieto¹, and René M Botnar^1
1School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom, ²Department of Radiology, Medical Image and Data Analysis (MIDAS), University Hospital of Tübingen, Tübingen, Germany, ³Ixico, London, United Kingdom, ⁴IHU LIRYC, Electrophysiology and Heart Modeling Institute, Université de Bordeaux, INSERM, Centre de recherche Cardio-Thoracuique de Bordeaux, Bordeaux, France, ⁵MR Research Collaborations, Siemens Healthcare Limited, Frimley, United Kingdom

3D whole‐heart coronary MR angiography (CMRA) has shown significant potential for the diagnosis of coronary artery disease. Undersampled motion corrected reconstruction approaches have enabled free-breathing isotropic 3D CMRA in ~5-10min scan time. However, spatial resolution is still limited compared to coronary CT angiography and scan time remains relatively long. In this work, we propose a deep-learning based super-resolution (SR) framework, combined with non-rigid respiratory motion compensation (SR-CMRA), to shorten the acquisition time to <1min. A 16-fold increase in spatial resolution is achieved by reconstructing a high-resolution CMRA (1.2mm³) from a low-resolution acquisition (1.2x4.8x4.8mm³, 50s scan).

Michael B Scott¹, Haben Berhane¹, Justin Baraboo¹, Cynthia K Rigsby², Joshua D Robinson², Patrick M McCarthy¹, S Chris Malaisrie¹, Ryan J Avery¹, Bradley D Allen¹, Alexander Barker³, and Michael Markl^1
1Northwestern University, Chicago, IL, United States, ²Lurie Children's Hospital of Chicago, Chicago, IL, United States, ³University of Colorado, Anschutz Medical Campus, Aurora, CO, United States

A fully automated pipeline using four convolutional neural networks was designed to perform analysis of aortic 4D flow MRI, including preprocessing (eddy current correction, noise masking, and antialiasing), 3D segmentation of the aorta, quantification of mean flow-time curves and peak velocities. The analysis pipeline was run on a total of 2084 4D flow MRI studies and compared against manual analysis in a subset of 69 studies. Median segmentation Dice score for the ascending aorta was 0.93 [0.90 – 0.95]. Pipeline-based quantification of ascending aortic peak velocities demonstrated bias of -0.05 m/s versus manual analysis [LOA: -0.26 to 0.15 m/s].

Seung Su Yoon^1,2, Michaela Schmidt², Manuela Rick², Teodora Chitiboi³, Puneet Sharma³, Tilman Emrich^4,5, Christoph Tilmanns⁶, Ralph Waßmuth⁶, Jens Wetzl², and Andreas Maier^1
1Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany, ²Magnetic Resonance, Siemens Healthcare GmbH, Erlangen, Germany, ³Siemens Medical Solutions USA, Inc., Princeton, NJ, United States, ⁴Department of Radiology, University Medical Center, Johannes Gutenberg-University Mainz, Mainz, Germany, ⁵Division of Cardiovascular Imaging, Department of Radiology and Radiological Science, Medical University of South Carolina, Charleston, SC, United States, ⁶Diagnostikum Berlin, Berlin, Germany

In cardiac MRI using the Late Gadolinium Enhancement technique, inversion recovery sequences are acquired for the correct myocardial nulling for optimal image contrast. In clinical practice, the selection of the proper inversion time to null healthy myocardium is manually performed by visual inspection. To standardize the process, we propose an automated deep-learning-based system which selects the “null inversion time” where the myocardium signal is darkest, and “contrast inversion time” where the contrast between the myocardium and blood pool is highest. We validated the system on a prospective study on different scanners. The system achieved high accuracy in observers’ annotation range.

Michael Loecher^1,2, Luigi E Perotti³, and Daniel B Ennis^1,2,4,5
1Radiology, Stanford University, Stanford, CA, United States, ²Radiology, Veterans Affairs Health Care System, Palo Alto, CA, United States, ³Mechanical and Aerospace Engineering, University of Central Florida, Orlando, FL, United States, ⁴Maternal & Child Health Research Institute, Stanford University, Stanford, CA, United States, ⁵Cardiovascular Institute, Stanford University, Stanford, CA, United States

This work introduces a neural network for tracking myocardial motion in cine grid tagged MR images on a voxel-wise basis. This is achieved with the use a synthetic training dataset that includes comprehensive motion patterns. Synthetic training allows for a known ground truth motion to be included in training. The network was tested against a previous network that tracked only tag line intersections. Displacements and strain maps were generated and compared. The voxel tracking network shows qualitatively better spatial localization of strain, and better radial strain values compared to tracking only tag lines.

Miaoqi Zhang¹, Mingzhu Fu¹, Hanyu Wei¹, Shuo Chen¹, and Rui Li^1
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, Tsinghua University, Beijing, China

The rupture of intracranial aneurysm (IA) is the most common cause of subarachnoid hemorrhage (SAH), resulting in patient death and disability. Recently, the morphology of the aneurysm, wall condition as well as hemodynamic factors were found to have kind of relationship with the stability of the aneurysm. In this project, we extracted clinical characteristics, morphology parameters, wall condition and hemodynamic parameters together to predict aneurysm stability using a machine learning model based on 4D-Flow MRI and black blood MRI. Among the two models, the Support Vector Machines model performed well, and the multi-parameter prediction result was greater than 95%.

Hanyue Zhou¹, Jiayu Xiao², Debiao Li^1,2, Dan Ruan^1,3, and Zhaoyang Fan^2,4,5
1Bioengineering, University of California, Los Angeles, Los Angeles, CA, United States, ²Cedars-Sinai Medical Center, Los Angeles, CA, United States, ³Radiation Oncology, University of California, Los Angeles, Los Angeles, CA, United States, ⁴Radiology, University of Southern California, Los Angeles, CA, United States, ⁵Radiation Oncology, University of Southern California, Los Angeles, CA, United States

Intracranial vessel wall segmentation is an essential step for the intracranial atherosclerosis quantification. We have developed an automated intracranial vessel wall segmentation method based on deep learning that utilized a 2.5D UNet++ network structure with a loss function consists of both soft Dice coefficient loss and the approximated Hausdorff distance loss. We show that we have achieved significant improvements over our previous segmentation model based on a 2D UNet structure across various quantitative measures, as well as a better visual resemblance to the ground truth segmentation.

Grzegorz Tomasz Kowalik¹, Javier Montalt-Tordera¹, Jennifer Steeden¹, and Vivek Muthurangu^1
1Institute of Cardiovascular Science, University College London, London, United Kingdom

The Machine Learning aided k-t SENSE for the reconstruction of highly undersampled GAS_perturbed PCMR data is validated. We introduce a modified version of the u-net Convolutional Neural Network (u-net M) that utilises the spatial signal distribution information to improve removal of the MR image magnitude aliases. The high resolution magnitude predictions enable creation of regularisation priors used in the k-t SENSE for the final reconstruction of the PCMR data. 20 patients were scanned in the in-vivo validataion. The technique enabled ~3.6x faster processing than the CS reconstruction with no statistical difference in the measured peak mean velocity and stroke volumes.