Shuai Shen^1,2,3,4, Xiao Liu⁵, Zhuyuerong Li⁵, Tao Jiang⁵, Hairong Zheng^1,3,4, Xin Liu^1,3,4, and Na Zhang^1,3,4
1Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhen, China, ²College of Software, Xinjiang University, Urumqi, China, ³Key Laboratory for Magnetic Resonance and Multimodality Imaging of Guangdong Province, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhen, China, ⁴CAS key laboratory of health informatics, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhen, China, ⁵Department of radiology, Beijing Chao-Yang hospital, Capital medical university, beijing, China

At present, deep learning has gradually been applied to the field of plaque segmentation. However, the existing work is mainly used for the processing of 2D images. In this study, we trained a 3D network model to automatically segment the middle cerebral artery plaques based on 3D images and compared the accuracy with 2D network model. Magnetic resonance vessel wall imaging (MR-VWI) data from 102 patients were used for training. The results showed that all quantitative accuracy indicators of V-net were higher than U-net, and experiments showed that V-net was more stable.

Ning Jin¹, Maria Monzon², Teodora Chitiboi³, Aaron Pruitt⁴, Daniel Giese², Matthew Tong⁵, and Orlando P Simonetti^5,6,7
1Cardiovascular MR R&D, Siemens Medical Solutions USA, Inc., Cleveland, OH, United States, ²Siemens Healthcare, Erlangen, Germany, ³Siemens Medical Solutions USA, Inc, Princeton, NJ, United States, ⁴Biomedical Engineering, The Ohio State University, Columbus, OH, United States, ⁵Internal Medicine, The Ohio State University, Columbus, OH, United States, ⁶Davis Heart & Lung Research Institute, The Ohio State University, Columbus, OH, United States, ⁷Radiology, The Ohio State University, Columbus, OH, United States

Phase-contrast (PC) MRI is used to evaluate blood hemodynamics; however, it can be time consuming to process PC-MR data. In this work, we developed a fully automated segmentation algorithm for PC MR images using deep learning (DL). Automated segmentation of aorta and main pulmonary artery from PC MRI scans can be successfully achieved using the DL model.

Xi Xu¹, Yuxin Yang¹, Yuanyuan Liu¹, Dong Liang¹, Hairong Zheng¹, and Yanjie Zhu^1
1Shenzhen Institute of Advanced Technology, ShenZhen, China

We evaluate three different image denoising methods in cardiac diffusion tensor imaging (CDTI) regarding image quality and accuracy of parameter estimates with simulation and ex-vivo experiments. The local principal component analysis (LPCA) performs the best in improving image quality both in simulated and ex-vivo data, and the uncertainty of parameter estimations is reduced by all three algorithms in the ex-vivo experiment.

Dana Peters¹, Jérôme Lamy¹, Felicia Seemann², Einar Heiberg³, and Ricardo Gonzales^1
1Yale Unversity, New Haven, CT, United States, ²National Institutes of Health, Bethesda, MD, United States, ³Lund University, Lund, Sweden

Using a deep-learning tool for tracking the left ventricular valve plane on long-axis cine, markers of diastolic dysfunction could be easily evaluated, including e-prime, a-prime (atrial kick valve velocity), s-prime, and isovolumic relaxation time (IVRT).  This analysis, performed in 23 patients with atrial fibrillation, revealed a dependence of both a-prime and IVRT  on atrial fibrosis.

Chuyu Liu¹, Shuo Chen¹, and Rui Li^1
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, Tsinghua University, Beijing, China

By adapting 3D FuseUnet, CNN fSNAP showed better performance in lumen and IPH depiction compared with traditional fSNAP. The results suggest that deep learning can help fast SNAP scans produce high quality images, which could have great clinical utility.

Eric D Carruth¹, Samuel W Fielden¹, Amro Alsaid¹, Brandon K Fornwalt¹, and Christopher M Haggerty^1
1Geisinger, Danville, PA, United States

As genomic screening initiatives expand, pre-symptomatic identification of genetic risk of disease is increasingly common. For example, Geisinger’s Genomic Screening and Counseling program has identified carriers of rare variants associated with arrhythmogenic cardiomyopathy, but most have no clinical diagnosis. We explored whether myocardial T1 or T2 maps or synthetic extracellular volume (sECV) in these patients could assist in risk stratification by identifying pre-clinical myocardial abnormalities. We compared these quantitative measures to those of controls with a normal cardiac MRI. Native T1 was significantly elevated in asymptomatic variant carriers; however, T2, post-contrast T1, and sECV values were comparable.

Botian Xu¹ and John C. Wood^1,2
1Department of Biomedical Engineering, University of Southern California, Los Angeles, CA, United States, ²Division of Cardiology, Children's Hospital Los Angeles, Los Angeles, CA, United States

Post processing of cardiac tagged MRI has always been challenging because of poor SNR and image artifact. Inspired by unsupervised anomalies detection using variational autoencoder (VAE), we treat tags as anomalies and employ robust variational autoencoder (RVAE), using \(\beta\)-ELBO cost, which is more robust to outliers, to replace the log-likelihood optimization of a prototypical VAE, to generate tag-free results from cardiac tagged images.

Samuel Fielden^1,2, Eric Carruth¹, Brandon Fornwalt^1,3,4, and Christopher Haggerty^1,3
1Translational Data Science and Informatics, Geisinger, Danville, PA, United States, ²Medical and Health Physics, Geisinger, Danville, PA, United States, ³Heart Institute, Geisinger, Danville, PA, United States, ⁴Radiology, Geisinger, Danville, PA, United States

Displacement Encoding with Stimulated Echoes (DENSE) is a powerful technique that has found great utility in accurately measuring cardiac tissue displacement. However, DENSE remains time-consuming to acquire, particularly for 3-dimensionally encoded or higher resolution schemes, so methods to accelerate image acquisition are needed. Deep learning has shown promise to assist with a myriad of reconstruction problems, including DENSE. Here, we explore the reconstruction performance of a non-Cartesian Deep Cascade of Convolutional Neural Networks (DCCNN) when presented with undersampled data generated from multiple spiral trajectory designs and acceleration rates.

Wenjing Xu^1,2,3,4, Xiong Yang⁵, Yikang Li^1,3,4, Jin Fang⁶, Guihua Jiang⁶, Shuheng Zhang⁵, Yanqun Teng⁵, Xiaomin Ren⁵, Lele Zhao⁵, Jiayu Zhu⁵, Qiang He⁵, Hairong Zheng^1,3,4, Xin Liu^1,3,4, and Na Zhang^1,3,4
1Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, ²Faculty of Information Technology, Beijing University of Technology, Beijing, China, ³Key Laboratory for Magnetic Resonance and Multimodality Imaging of Guangdong Province, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, ⁴CAS key laboratory of health informatics, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, ⁵Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China, ⁶Department of Radiology, Guangdong Second Provincial General Hospital, Guangzhou, China

The rupture of unstable or vulnerable atherosclerotic plaque is the major cause of ischemic stroke. Manual analysis of the vessel wall and plaque is labor-intensive and experience dependent. The purpose of this study is to develop an automatic method to segment the vessel wall and lumen contour for quantitative measurement. In this work, a CNN architecture for fully automated segmentation of arterial lumen and vessel wall on MR vessel wall images was developed and evaluated on ischemic stroke patients. In conclusion, we proposed a good performance automatic analysis method for the vessel wall and is important for plaque analysis.

kankan hao^1,2, yanjie zhu^1,2, dong liang^1,2, shihua zhao³, xin liu^1,2, and hairong zheng^1,2
1Paul C. Lauterbur Research Centre for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhen, China, ²University of Chinese Academy of Sciences, Beijing, China, ³Department of Magnetic Resonance Imaging, Fuwai Hospital and National Center for Cardiovascular Diseases, Chinese Academy of Medical Sciences, Peking Union Medical College, beijing, China

Cardiac magnetic resonance(CMR) is a highly reliable measurement to assess and predict cardiovascular events. The traditional regression model need a linear assumption but it can not be guaranteed. We use a ML method to predict cardiovascular events in HCM patients. According to our result, the C statistic for the ML model (0.804 [95% CI, 0.757-0.850]) was higher than Cox regression model (0.764, [95% CI, 0.671-0.857]). With the random sample, the ROC for the ML model(0.96 in the training set, 0.83 in the test set) was higher than the regression model(0.80 in the training set, 0.79 in the test set).  

Jack R. Warren¹, Matthew J. Middione², Julio A. Oscanoa^2,3, Christopher M. Sandino⁴, Shreyas S. Vasanawala², and Daniel B. Ennis^2,5
1Department of Computing + Mathematical Sciences, California Institute of Technology, Pasadena, CA, United States, ²Department of Radiology, Stanford University, Stanford, CA, United States, ³Department of Bioengineering, Stanford University, Stanford, CA, United States, ⁴Department of Electrical Engineering, Stanford University, Stanford, CA, United States, ⁵Cardiovascular Institute, Stanford University, Stanford, CA, United States

Phase Contrast MRI (PC-MRI) measures the flow of blood. In order to obtain high-quality measurements, patients must hold their breath for ~20 seconds, which oftentimes can be difficult. Advances in deep learning (DL) have allowed for the reconstruction of highly undersampled MRI data. A 2D PC-MRI DL-ESPIRiT network was recently proposed to undersample the data acquisition by up to 8x without compromising clinically relevant measures of flow accuracy within ±5%. This work uses k-fold cross validation to evaluate the DL-ESPIRiT network on 2D PC-MRI data in terms of accuracy and variability for pixel velocity, peak velocity, and net flow.

Julio A. Oscanoa^1,2, Matthew J. Middione², Christopher M. Sandino³, Shreyas S. Vasanawala², and Daniel B. Ennis^2,4
1Department of Bioengineering, Stanford University, Stanford, CA, United States, ²Department of Radiology, Stanford University, Stanford, CA, United States, ³Department of Electrical Engineering, Stanford University, Stanford, CA, United States, ⁴Cardiovascular Institute, Stanford University, Stanford, CA, United States

We propose a novel Deep Learning (DL) based reconstruction framework for accelerated 2D Phase Contrast MRI (PC-MRI) datasets. We extend a previously developed DL method based on ESPIRiT reconstruction for cardiac cine and combine it with a direct Complex Difference estimation approach. We tested the DL methods using retrospectively undersampled 2D PC-MRI data and compared it with conventional Compressed Sensing (CS) reconstruction. Our method outperformed CS and enabled higher acceleration factors up to 8x while maintaining error metrics within a targeted accuracy of ±5%.

Vadim Malis¹, Won Bae¹, Asako Yamamoto¹, Yoshimori Kassai², Marin A McDonald¹, and Mitsue Miyazaki^1
1Radiology, UC San Diego, San Diego, CA, United States, ²Canon Medical, Tochigi, Japan

This study demonstrated that a mild exercise increases length and depiction of discernible lenticulostriate artery branches in 3D susceptibility weighted imaging, facilitated by applying denoising deep learning reconstruction (dDLR), and post-processed using vascular structure extraction algorithm (VSEA). Additionally, exercise increased intracranial perfusion, determined using 3D pCASL.

Haben Berhane¹, Michael Scott¹, Justin Baraboo¹, Cynthia Rigsby², Joshua Robinson², Bradley Allen³, Chris Malaisrie³, Patrick McCarthy³, Ryan Avery³, and Michael Markl^1
1Biomedical Engineering, Northwestern University, Chicago, IL, United States, ²Lurie Childrens Hospital of Chicago, Chicago, IL, United States, ³Northwestern Radiology, Evanston, IL, United States

While 4D flow MRI is capable of providing extensive hemodynamic quantifications, it requires cumbersome and time-consuming pre-processing. In order to accelerate the process, we developed and validated a multi-label convolutional neural network (CNN) for automatic aortic 3D segmentation and regional-labeling (ascending, AAo; arch; and descending aorta, DAo). Utilizing 320 4D flow MRI datasets, we used a 10-fold cross validation for training and testing the CNN. The Dice scores across each region were AAo: 0.95 [0.93-0.98], arch: 0.90 [0.89-0.95], and DAo: 0.95 [0.94-0.98]. Across all flow metrics, Bland-Altman comparisons showed moderate to excellent agreement between the manual and automated regional segmentations.

Pierre Daudé¹, Patricia Ancel², Sylviane Confort-gouny¹, Anne Dutour², Bénédicte Gaborit², and Stanislas Rapacchi^1
1Aix-Marseille Univ, CNRS, CRMBM, Marseille, France, ²APHM, Hôpital Universitaire Timone, Service d’Endocrinologie, Marseille, France

Evaluation of epicardial adipose tissue (EAT) burden holds potential as a biomarker for CHD diagnosis. EAT volume is challenging to assess using MRI due to its curved shape susceptible to partial volume effect. As a substitute, 4-chamber EAT surface can be reliably measured and has shown good correlation with EAT volume (r²=0.62). Two fully convolutional neural networks (FCN) were investigated for the segmentation of EAT surface on a database of 126 subjects. Promising results were obtained with DICE values of 0.71.

Pamela Franco^1,2,3, Julio Sotelo^1,3,4, Lydia Dux-Santoy⁵, Andrea Guala⁵, Aroa Ruiz-Muñoz⁵, Arturo Evangelista⁵, José Rodríguez-Palomares⁵, and Sergio Uribe^1,3,6
1Biomedical Imaging Center, School of Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile, ²Electrical Engineering Department, School of Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile, ³Millennium Nucleus for Cardiovascular Magnetic Resonance, Santiago, Chile, ⁴School of Biomedical Engineering, Universidad de Valparaíso, Valparaíso, Chile, ⁵Department of Cardiology, Hospital Universitari Vall d’Hebron, Vall d’Hebron Institut de Recerca (VHIR), Universitat Autònoma de Barcelona, Barcelona, Spain, ⁶Radiology Department, School of Medicine, Pontificia Universidad Católica de Chile, Santiago, Chile

The clinical significance and economic burden of bicuspid aortic valve (BAV) disease justify the need for improved clinical guidelines and more robust therapeutic modalities. Recent advances in medical imaging have demonstrated the existence of altered hemodynamics in these patients. To identify hemodynamic biomarkers for BAV patients, we present a machine learning method consisting of a feature selection mechanism to classify healthy volunteers and BAV patients accurately.

Stefano Buoso¹, Thomas Joyce¹, and Sebastian Kozerke^1
1ETH Zurich, Zurich, Switzerland

We propose a parametric physics-informed neural network that can be personalized to left-ventricular anatomies from cardiac MRI data. The model combines a left-ventricular anatomical shape model derived from cardiac MRI data, and a functional model derived from synthetic cardiac deformations. The network is trained with a label-free approach using a physics-based cost-function in less than 5 minutes on a single CPU. Network inputs are endocardium pressure and myocyte activation. A complete cardiac cycle can be simulated in less than a minute. This approach is 30 times faster than the corresponding finite element simulation even when including training time.

Zhongsen Li¹, Hanyu Wei¹, Chuyu Liu¹, Yichen Zheng¹, Shuo Chen¹, and Rui Li^1
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, Tsinghua University, Beijing, China

In this study, we combine classical “partial separable” model with deep-learning framework “variational network” for accelerated reconstruction of real-time cardiac MR imaging. The proposed PS-VN architecture achieves comparable reconstruction accuracy with baseline algorithm and reduce computational time to around 10 seconds for the reconstruction of over 4 thousand dynamic frames.

Markus J Ankenbrand¹, Liliya Shainberg¹, Michael Hock¹, David Lohr¹, and Laura Maria Schreiber^1
1Chair of Cellular and Molecular Imaging, Comprehensive Heart Failure Center (CHFC), University Hospital Würzburg, Würzburg, Germany

Deep learning-based segmentation models play an important role in cardiac magnetic resonance imaging. While their performance is good on the training and validation data the models themselves are hard to interpret. Sensitivity analysis helps to estimate the effect of different data characteristics on segmentation performance. We demonstrate that a published model exhibits higher sensitivity to basic transformations like rotation for pathology classes than for tissue classes in general.