Thomas Hadler¹, Clemens Ammann¹, Jan Gröschel^1,2, and Jeanette Schulz-Menger^1,2,3
1Charité - Universitätsmedizin Berlin, Working Group on CMR, Experimental and Clinical Research Center, a joint cooperation between the Max-Delbrück-Center for Molecular Medicine and the Charité – Universitätsmedizin Berlin, Berlin, Germany, ²DZHK (German Centre for Cardiovascular Research), partner site Berlin, Berlin, Germany, ³Department of Cardiology and Nephrology, Helios Hospital Berlin-Buch, Berlin, Germany

Three popular convolutional neural networks were trained on short-axis cine MR images of the heart acquired by a prototype 2-shot 2D Compressed Sensing sequence. Network performance was evaluated on the level of clinical results and segmentation quality. Analysis revealed high correlation for quantitative results between all networks and a human expert. Automatic segmentation of the right ventricle is significantly more difficult than the left ventricle and shows more outliers. Segmentation decision errors concentrate in basal and apical slices, with the largest millilitre differences in the basal slices. Fast acquisition and automated image analysis promise high efficiency in CMR.

Thomas Joyce¹, Nico Schulthess¹, Gloria Wolkerstorfer¹, Stefano Buoso¹, and Sebastian Kozerke^1
1University and ETH Zurich, Zurich, Switzerland

We propose to use a combination of the StyleGAN2, ADA and DatasetGAN methods to produce synthetic short- and long-axis view cardiac magnetic resonance (CMR) images accompanied with corresponding 11-class tissue masks. The image generator networks are trained on datasets of approximately 1850 and 5000 unlabelled images, for short- and long-axis images respectively.  The segmentation networks are trained on only 30 manually annotated synthetic images in total. We further demonstrate a proof-of-concept method for generating coherent long- and short-axis images of the same synthetic patient.

Sina Amirrajab¹, Cristian Lorenz², Jürgen Weese², Josien Pluim¹, and Marcel Breeuwer^1,3
1Biomedical Engineering Department, Eindhoven University of Technology, Eindhoven, Netherlands, ²Philips Research Laboratories, Hamburg, Germany, ³MR R&D - Clinical Science, Philips Healthcare, Best, Netherlands

We propose a method for synthesizing cardiac MR images with plausible heart shape and realistic appearance. It breaks down the synthesis into labels deformation and label-to-image translation. The former is achieved via latent space interpolation in a VAE model, while the latter is accomplished via a conditional GAN model. We synthesize 32 short-axis slices within each subject (intrasubject), as well as eight intermediary generated subjects between two dissimilar real subjects (intersubject) that have different anatomies. Such method could provide a solution to enrich a database and to pave the way for the development of generalizable DL based algorithms.

Andrew Phair¹, Gastao Cruz¹, René Botnar¹, and Claudia Prieto^1
1School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom

Iterative reconstructions for multi-contrast multidimensional non-Cartesian MRI can require considerable computation time, with a large fraction dedicated to non-uniform fast Fourier transforms (NUFFTs). The Toeplitz approach has been previously shown to allow NUFFT operations to be replaced by computationally efficient fast Fourier transforms (FFTs) without loss of accuracy. Herein, we combine this approach with a 3D multi-contrast low-rank cardiac-motion-corrected radial reconstruction with low-rank patch-based regularisation to achieve a ~13.3-fold computational speed-up.

Sherine Brahma¹, Tobias Schäffter^1,2,3, Christoph Kolbitsch^1,3, and Andreas Kofler^1
1Physikalisch-Technische Bundesanstalt, Braunschweig and Berlin, Germany, ²Department of Biomedical Engineering, Technical University of Berlin, Berlin, Germany, ³School of Imaging Sciences and Biomedical Engineering, King’s College London, London, United Kingdom

Due to the black-box nature of Deep Learning (DL) algorithms, uncertainty quantification (UQ) is a promising approach to assess their risk in medical applications. However, UQ is challenging in imaging techniques like non-Cartesian multi-coil CINE MRI because the data is high-dimensional, and the acquisition process is computationally demanding. In this work, i) we propose to utilize spatio-temporal (ST) networks, demonstrating efficient UQ in a high-dimensional setting, and ii) we show a reduction in uncertainty by adopting the forward model of radial multi-coil 2D cine MR in the reconstruction process. UQ is performed using MCMC dropout with additional aleatoric loss terms.

Daniel Amsel¹, Seung Su Yoon^1,2, Jens Wetzl², and Andreas Maier^1
1Friedrich-Alexander University Erlangen-Nuernberg, Erlangen, Germany, ²Siemens-Healthineers, Erlangen, Germany

Diagnostic applications often require the estimation of organ motion. Image registration enables motion estimation by computing deformation fields for an image pair. In this work, voxelmorph, a framework for deep learning-based diffeomorphic image registration is used to register CINE cardiac MR images in four-chamber view. Additionally, the framework is extended to a one-to-many registration to also utilize temporal information within a time-resolved MR scan. Registration performance as well as the performance of a valve tracking application using this approach are evaluated. The results are comparable to a state-of-the-art registration method, while noticeably reducing the computation time.

Marili Niglas¹, Nicoleta Baxan², Ali Ashek¹, Lin Zhao¹, Jinming Duan³, Declan O'Regan⁴, Timothy JW Dawes^1,4, Wenjia Bai⁵, and Lan Zhao^1
1National Heart and Lung Institute, Imperial College London, London, United Kingdom, ²Biological Imaging Centre, Imperial College London, London, United Kingdom, ³School of Computer Science, University of Birmingham, London, United Kingdom, ⁴MRC London Institute of Medical Sciences, Imperial College London, London, United Kingdom, ⁵Department of Brain Sciences, Imperial College London, London, United Kingdom

Multiple cardiac AI-based image processing pipelines exist for clinical application, yet the equivalent is lacking for rodents. We utilized a fully convolutional network combined with 3D-atlas registration to auto-segment cine images from pulmonary hypertension (PH) rats and produce 3D contraction maps. The auto-segmentations were equivalent to manual (Dice overall >0.7). The volumetric parameters did not differ between methods, except a minor underestimation for RVESV in PH rat (8.2%). 3D contraction maps indicated moderately increased basal wall motion at early (adaptive) stage followed by a 36% reduction at later (maladaptive) stage of PH. This regional motion remodelling correlates with PAH patients.

Alexander Fyrdahl¹ and Nicole Seiberlich^1
1Radiology, University of Michigan, Ann Arbor, MI, United States

Through-time non-Cartesian GRAPPA can be used to enable real-time imaging for evaluation of left-ventricular function. However, it requires arrays with many receiver coils for the parallel imaging reconstruction and gradient systems capable of playing out the specified non-Cartesian trajectories. In this work, we explore the use of non-Cartesian GRAPPA for real-time cardiac imaging on a 0.55T system with lower performance gradients, reduced receiver coils compared to conventional MRI systems.

Mohsen Farzi¹, Darryl McClymont², Hannah Whittington², Marie-Christine Zdora³, Leah Khazin¹, Craig A. Lygate², Christoph Rau³, Erica Dall’Armellina¹, Irvin Teh¹, and Jürgen E. Schneider^1
1Biomedical Imaging Science Department, University of Leeds, Leeds, United Kingdom, ²Cardiovascular Medicine, University of Oxford, Oxford, United Kingdom, ³Diamond Light Source Ltd., Oxfordshire, United Kingdom

Biophysical modelling is a valuable technique for interrogating cardiac microstructure. We utilise a two-compartment model where intra- and extra-cellular space were represented by impermeable cylinders and an oblate diffusion tensor, respectively. Data was acquired ex vivo using a bespoke diffusion scheme with five diffusion times, six b-values and ten directions. Cardiomyocyte radius (r = 7.6 versus 8.9 μm) and volume fraction (ICV = 0.56 versus 0.67) were sensitive markers of cardiomyopathy validated in one healthy and one disease heart with transverse aortic constriction (TAC).

Jaykumar Patel^1,2, Calder Sheagren^1,2, Saqeeb Hassan², Fatemeh Rastegar Jouybari^1,3, Christopher Macgowan^1,3, and Graham Wright^1,2
1Medical Biophysics, University of Toronto, Toronto, ON, Canada, ²Schulich Heart Centre, Sunnybrook Research Institute, Toronto, ON, Canada, ³The Hospital for Sick Children, Toronto, ON, Canada

During MR-guided cardiac catheter intervention, the catheter may be misaligned due to the respiratory motion of the heart. To improve the myocardial border alignment, GPU-based edge-preserving image registration is developed to accurately align highly undersampled 3D cone-trajectory image-based navigators for motion characterization. The edge-preserving registration technique improved myocardial border alignment across differing spatial resolutions and acceleration factors in-silico.

Kateřina Škardová¹, Radek Galabov^1,2, Kateřina Fricková¹, Tomáš Pevný³, Jaroslav Tintěra², Tomáš Oberhuber¹, and Radomír Chabiniok^1,4
1Department of Mathematics, FNSPE CTU in Prague, Prague, Czech Republic, ²Department of Radiology, Institute for clinical and experimental medicine, Prague, Czech Republic, ³Artificial Intelligence Center, CTU in Prague, Prague, Czech Republic, ⁴Department of Pediatrics, UT Southwestern Medical Center, Dallas, TX, United States

A method for estimating tissue parameters using cardiovascular MRI and biophysical model by combining neural network (NN) and numerical optimization (NO) is illustrated on estimating $$$T_1$$$ relaxation time from MOLLI. Compared to the estimation obtained from MOLLI by the scanner, the proposed method provided $$$T_1$$$ closer to turbo spin-echo pseudo-ground in 7 out of 8 phantoms and higher or comparable myocardial and blood $$$T_1$$$ in 6 out of 7 patiens’ datasets. Including the NN-based initial guess accelerated the subsequent NO. NO initialized by NN, trained using simulated data, showed the potential to increase the efficiency and robustness of parameter estimation.

Ahmed S Negm¹, Jianing Pang², Kiaran Mcgee¹, Holly Iverson¹, Maria Halverson¹, Rachel Davids², Prabhakar Rajiah¹, Alexander Bratt¹, Eric Williamson¹, Phillip Young¹, Thomas Foley¹, Tim Leiner¹, Christopher Francois¹, Xiaoming Bi², and Jeremy Collins^3
1Radiology, Mayo Clinic, Rochester, MN, United States, ²Siemens Healthineers, Chicago, IL, United States, ³Mayo Clinic, Rochester, MN, United States

CMR is the gold standard for assessing biventricular size and function. bSSFP Seg Cine is a standard cine method, but requires breath-holding which prolongs table time and limits utility in dyspneic patients. A novel real-time technique RTCSCineMoCo combines free-breathing acquisition with compressed sensing and motion correction. In 100 subjects RTCSCineMoCo generated similar results to Seg Cine for LVEF, LV end-diastolic mass, and RVESV with agreement including no difference at Bland-Altman analysis for all LV and RV measures. The shorter table time for RTCSCineMoCo and ability to robustly image dyspneic patients are important enhancements for clinical practice.

Marina Awad¹, Junyu Wang¹, Xue Feng¹, Ruixi Zhou¹, and Michael Salerno^1,2,3,4
1Biomedical Engineering, University of Virginia, Charlottesville, VA, United States, ²Medicine, University of Virginia, Charlottesville, VA, United States, ³Radiology, University of Virginia, Charlottesville, VA, United States, ⁴Medicine, Stanford University, Stanford, CA, United States

Cardiac real-time cine imaging is useful for patients who cannot hold their breath or have irregular heart rhythms. Free breathing high-resolution real-time cardiac cine images are acquired efficiently using spiral acquisitions, and rapidly reconstructed using our DESIRE framework. However, quantifying the EF from free-breathing real-time imaging is limited by the lack of an EKG signal to define the cardiac cycle, and through-plane cardiac motion resulting from free-breathing. We developed a DL-based segmentation technique to determine the end-expiratory phase and determine end-systole and end-diastole on a slice by slice basis to accurately quantify EF from spiral real-time cardiac cine imaging.

Elizabeth Walker Thompson¹, Abhijit Bhattaru¹, Phuong Vu¹, Elizabeth Donnelly², Elizabeth Goldmuntz², Mark Fogel², and Walter Witschey^1
1University of Pennsylvania, Philadelphia, PA, United States, ²Children's Hospital of Philadelphia, Philadelphia, PA, United States

Tetralogy of Fallot (ToF) is a congenital heart disease that is typically repaired with surgery early in life, but right ventricular remodeling results in adverse events for many patients. This preliminary analysis of 8 patients investigated the feasibility of training a convolutional neural network to segment the right and left ventricles from 2-dimensional cardiovascular magnetic resonance images, resulting in Dice scores ranging from 0.73-0.91 for the left ventricular blood pool, left ventricular myocardium, and right ventricular blood pool. Machine learning shows promise to enable large-scale longitudinal studies of ToF.

Alexander Paul Neofytou¹, Radhouene Neji^1,2, James Wong³, Anastasia Fotaki¹, Joana Ferreira¹, Carl Evans¹, Filippo Bosio¹, Nabila Mughal¹, Reza Razavi¹, Kuberan Pushparajah¹, and Sébastien Roujol^1
1School of Biomedical Engineering and Imaging Sciences, Faculty of Life Sciences and Medicine, King's College London, London, United Kingdom, ²MR Research Collaborations, Siemens Healthcare Limited, Frimley, United Kingdom, ³Department of Paediatric Cardiology, Evelina London Children’s Hospital, London, United Kingdom

A novel reconstruction technique based on iterative rejection of segmented k-space was developed for retrospective correction of respiratory motion in multiple average, free-breathing cine images. In comparison to standard signal averaging reconstruction, it provides higher sharpness, SNR, CNR, image quality, and rate of diagnostic quality images. 

Vivian S. Nguyen¹, Jacob P. Goes¹, Donovan Gorre², Hui Wang^3,4, Marcella K. Vaicik⁵, Amit R. Patel⁶, and Keigo Kawaji^1,6
1Biomedical Engineering, Illinois Institute of Technology, Chicago, IL, United States, ²Radiology, University of Chicago Medical Center, Chicago, IL, United States, ³Philips, Gainesville, FL, United States, ⁴Department of Radiology, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States, ⁵Illinois Institute of Technology, Chicago, IL, United States, ⁶Medicine - Cardiology, University of Chicago Medical Center, Chicago, IL, United States

SENC and SPAMM tagging Cardiac Magnetic Resonance (CMR) are specifically optimized for acquisition before contrast injection and tuned to the end-systolic phase.  While SPAMM tagging is a non-quantitative approach that indirectly derives functional motion from the displacement of the saturated tags in the myocardium, SENC directly computes myocardial tissue compression (strain) in percentage units. Furthermore, SENC quantitation pipeline bypasses the signal intensity map that exhibits the decay of the embedded strain measurements. In this study, we examine both SPAMM and SENC post-Gd measurements to specifically indicate the effect of the variable RF pulse train tuned to target tissue T1.