Haikun Qi¹, Gastao Cruz¹, Thomas Kuestner¹, Karl Kunze², Radhouene Neji², René Botnar¹, and Claudia Prieto^1
1School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom, ²MR Research Collaborations, Siemens Healthcare Limited, Frimley, United Kingdom

A non-rigid respiratory motion-corrected reconstruction technique (non-rigid PROST) has achieved high-quality coronary MRA (CMRA). However, non-rigid PROST requires respiratory-resolved (bin) image reconstruction, bin-to-bin non-rigid registration and regularized reconstruction, leading to long computation time. In this study, we propose an end-to-end deep learning non-rigid motion-corrected reconstruction technique for highly undersampled free-breathing CMRA. It consists of a diffeomorphic motion estimation network and a motion-informed model-based deep learning reconstruction network that were trained jointly for motion-corrected undersampled reconstruction. Compared with non-rigid PROST, the proposed technique achieved better reconstruction performance in both retrospectively and prospectively 9x-accelerated CMRA, while operating orders of magnitude faster.

Javier Montalt-Tordera¹, Michael Quail¹, Jennifer Anne Steeden¹, and Vivek Muthurangu^1
1Centre for Cardiovascular Imaging, UCL Institute of Cardiovascular Science, University College London, London, United Kingdom

Contrast-enhanced MR angiography (CE-MRA) is often used in cardiovascular MRI, but contrast agents can have adverse effects. This work proposes to use deep learning to reduce contrast dose by 80%. A deep neural network was trained to enhance low-dose images using a synthetically generated dataset and validated with both synthetic and real low-dose images. The method was assessed for image quality, diagnostic accuracy and confidence and accuracy of vessel measurements. Enhanced low-dose images were found to be comparable to reference high-dose data. Therefore, the method could enable a reduction in contrast dose for CMR.

Tommaso Di Noto¹, Guillaume Marie¹, Sebastien Tourbier¹, Yasser Alemán-Gómez^1,2, Oscar Esteban¹, Guillaume Saliou¹, Meritxell Bach Cuadra^1,3, Patric Hagmann¹, and Jonas Richiardi^1
1Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ²Center for Psychiatric Neuroscience, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ³Medical Image Analysis Laboratory (MIAL), Centre d’Imagerie BioMédicale (CIBM), Lausanne, Switzerland

Machine learning challenges serve as a benchmark for determining state-of-the-art results in medical imaging. They provide direct comparisons between algorithms, and realistic estimates of generalization capability. By participating in the Aneurysm Detection And segMentation (ADAM) challenge, we learnt the most effective deep learning design choices to adopt when tackling automated brain aneurysm detection on multi-site data. Adjusting patch overlap ratio during inference, using a hybrid loss, resampling to uniform voxel spacing, using a 3D neural network architecture, and correcting for bias field were the most effective. We show that, when adopting these expedients, our model drastically improves detection performances. 

Yu Wang¹, Sona Ghadimi¹, Changyu Sun¹, and Frederick H. Epstein^1,2
1Biomedical Engineering, University of Virginia, Charlottesville, VA, United States, ²Radiology, University of Virginia, Charlottesville, VA, United States

As cine DENSE provides myocardial contours and intramyocardial displacement data, we investigated the use of DENSE to train deep networks to predict intramyocardial motion from contour motion. FlowNet2, an optical-flow convolutional neural network, was used as a comparator/reference, and as the starting point for a DENSE-trained network (DT-FlowNet2). Further, we added a correction network with convolution along time, resulting in a through-time-corrected DENSE-trained network (TC-DT-FlowNet2). TC-DT-FlowNet2 outperformed other methods, providing accurate intramyocardial displacements from myocardial contours. DENSE-based learning of intramyocardial displacements from contours holds promise as a new method for computing strain from the contours of standard cine MRI. 

Xiao Chen¹, Masoud Edalati², Qi Liu², Xingxian Shou², Abhishek Sharma¹, Mary P. Watkins³, Daniel J. Lenihan³, Linzhi Hu², Gregory M. Lanza³, Terrence Chen¹, and Shanhui Sun^1
1United Imaging Intelligence, Cambridge, MA, United States, ²UIH America, Inc., Houston, TX, United States, ³Cardiology, Washington University School of Medicine, St. Louis, MO, United States

Myocardium strain measures myocardial deformation and has been demonstrated a comprehensive, sensitive and early indicator of cardiac dysfunction. Feature tracking can assess myocardium strain from cine CMR images with no special acquisitions but requires observer intervention and expertise. We propose a deep-learning-based fully-automated myocardium strain assessment system that requires zero human intervention for accurate strain assessment.

Matthew J. Middione¹, Julio A. Oscanoa^1,2, Michael Loecher¹, Christopher M. Sandino³, Shreyas S. Vasanawala¹, and Daniel B. Ennis^1,4
1Department of Radiology, Stanford University, Palo Alto, CA, United States, ²Department of Bioengineering, Stanford University, Palo Alto, CA, United States, ³Department of Electrical Engineering, Stanford University, Palo Alto, CA, United States, ⁴Cardiovascular Institute, Stanford University, Stanford, CA, United States

2D PC-MRI Compressed sensing (CS) and deep learning (DL) reconstruction techniques may introduce a reconstruction induced phase bias, distinct from eddy current-induced background phase offsets, which may impact the accuracy of flow measurements if not corrected. Herein, we analyzed this reconstruction induced phase bias to determine the maximum acceleration factor that could be used with CS and DL reconstruction frameworks for 2D PC-MRI while minimizing errors in peak velocity and total flow within ±5%.

Kaiyue Tao¹, Li Chen², Niranjan Balu³, Gador Canton³, Wenjin Liu³, Thomas S. Hatsukami⁴, and Chun Yuan^3
1University of Science and Technology of China, Hefei, China, ²Department of Electrical and Computer Engineering, University of Washington, Seattle, WA, United States, ³Department of Radiology, University of Washington, Seattle, WA, United States, ⁴Department of Surgery, University of Washington, Seattle, WA, United States

Popliteal vessel wall features hidden in the Osteoarthritis Initiative (OAI) dataset warrant further investigation. However, if the number of annotations is insufficient, deep learning-based feature map analysis from MRI images may overfit and fail to generate a meaningful feature space. We designed a metric learning network combined with an episodic training method to overcome the problem of limited annotations, and demonstrated its ability to learn a meaningful feature embedding. Based on our feature map, we proposed an iterative workflow and identified vessel wall images with high probability of diseases from 1974 cases.

Ioannis Valasakis¹, Haikun Qi¹, Kerstin Hammernik², Gastao Lima da Cruz¹, Daniel Rueckert^2,3, Claudia Prieto¹, and Rene Botnar^1
1King's College London, London, United Kingdom, ²Technical University of Munich, Munich, Germany, ³Imperial College London, London, United Kingdom

3D whole-heart coronary MR angiography (CMRA) is limited by long scan times. Undersampled reconstruction approaches, such as compressed sensing or low-rank methods show promise to significantly accelerate CMRA but are computationally expensive, require careful parameter optimisation and can suffer from residual aliasing artefacts. A 2D multi-scale variational network (VNN) as recently proposed to improve image quality and significantly shorten the reconstruction time. We propose to extend the VNN reconstruction to 3D to fully capture the spatial redundancies in 3D CMRA. The 3D-VNN is compared against conventional and 3D model-based U-Net reconstruction techniques, showing promising results while shortening the reconstruction time.

Kevin Blansit¹, Tara Retson¹, Naeim Bahrami², Phillip Young^3,4, Christopher Francois³, Lewis Hahn¹, Michael Horowitz¹, Seth Kligerman¹, and Albert Hsiao^1
1UC San Diego, La Jolla, CA, United States, ²GE Healthcare, Menlo Park, CA, United States, ³Mayo Clinic, Rochester, MN, United States, ⁴Mayo, Rochester, MN, United States

We show that an automated system of deep convolutional neural networks can effectively prescribe double-oblique imaging planes necessary for acquisition of cardiac MRI. To examine its clinical potential, we assess its component-wise performance by comparing simulated imaging planes predicted by DCNNs against ground truth imaging planes defined by a cardiac radiologist. Performance was assessed on 280 exams including 1252 image series obtained on ten 1.5T and 3T MRI scanners from three academic institutions. We further compare imaging planes acquired by technologists at the time of original acquisition against the same ground truth as an additional external reference.

Félix Dumais¹, Marco Perez Caceres¹, Noémie Arès-Bruneau², Christian Bocti^2,3,4, and Kevin Whittingstall^5
1Médecine nucléaire et radiobiologie, Université de Sherbrooke, Sherbrooke, QC, Canada, ²Faculté de Médecine et des Sciences de la Santé, Université de Sherbrooke, Sherbrooke, QC, Canada, ³Clinique de la Mémoire et Centre de Recherche sur le Vieillissement, CIUSSS de l’Estrie-CHUS, Sherbrooke, QC, Canada, ⁴Service de Neurologie, Département de Médecine, CHUS, Sherbrooke, QC, Canada, ⁵Radiologie diagnostique, Université de Sherbrooke, Sherbrooke, QC, Canada

The Circle of Willis (CW) is a system of arteries located at the base of the brain. Its structure is a key component in different cerebrovascular diseases, though quantifying this on MR images is meticulous and labor intensive. We developed an analysis pipeline that uses a convolutional neural network to do multilabel segmentation of the CW as viewed through TOF-MRA images. Results show that this approach can automatically locate and label the CW with the same accuracy as expert human annotators.

Ioannis Koktzoglou^1,2, Rong Huang¹, William J Ankenbrandt^1,2, Matthew T Walker^1,2, and Robert R Edelman^1,3
1Radiology, NorthShore University HealthSystem, Evanston, IL, United States, ²University of Chicago Pritzker School of Medicine, Chicago, IL, United States, ³Northwestern University Feinberg School of Medicine, Chicago, IL, United States

Deep machine learning approaches offer the potential for improved super-resolution (SR) reconstruction which could be useful in many clinical applications. Patients with suspected stroke often undergo MRI, which often includes magnetic resonance angiography (MRA) of the head and neck arteries with scan times of ≈10 to 15 minutes using standard nonenhanced methods. With the aim of shortening scan times, we evaluated the feasibility and performance of four deep neural network (DNN)-based SR reconstructions for restoring the image quality and spatial resolution of thin slab stack-of-stars quiescent interval slice-selective (QISS) head and neck MRA with degraded slice resolution.

Maria Monzon^1,2, Seung Su Yoon^1,2, Carola Fischer², Andreas Maier¹, Jens Wetzl², and Daniel Giese^2
1Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany, ²Magnetic Resonance, Siemens Healthcare GmbH, Erlangen, Germany

The analysis of mitral valve motion is known to be relevant in the diagnosis of cardiac dysfunction. Dynamic motion parameters can be extracted from Cardiac Magnetic Resonance (CMR) images. We propose two chained Convolutional Neural Networks for automatic tracking of mitral valve-annulus landmarks on time-resolved 2-chamber and 4-chamber CMR images. The first network is trained to detect the region of interest and the second to track the landmarks along the cardiac cycle. We successfully extracted several motion-related parameters with high accuracy as well as analyzed unlabeled datasets,  thereby overcoming time-consuming annotation and allowing statistical analysis over large number of datasets.

Miaoqi Zhang¹, Qingchu Jin², Mingzhu Fu¹, Hanyu Wei¹, and Rui Li^1
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, Tsinghua University, Beijing, China, ²Johns Hopkins University, Baltimore, MD, United States

Intracranial aneurysms are abnormal dilations of the cerebral arteries that have a prevalence of 5-8% in the general population. In this study, we successfully segmented IAs from dual inputs (TOF-MRA and T1-VISTA) using the hyperdense net with higher accuracy than a single input. The maximum diameter measurements for IAs derived from our segmentation was consistent with the maximum diameters obtained from the criterion standard (DSA). We showed that larger aneurysms were easier to segment by the deep learning model. In the future, we will test other deep learning models on aneurysm segmentation and compare these results with the hyperdense net.

Yin Guo¹, Li Chen², Dongxiang Xu³, Rui Li⁴, Xihai Zhao⁴, Thomas S. Hatsukami⁵, and Chun Yuan^1,3
1Bioengineering, University of Washington, Seattle, WA, United States, ²Electrical Engineering, University of Washington, Seattle, WA, United States, ³Radiology, University of Washington, Seattle, WA, United States, ⁴Biomedical Engineering, Tsinghua University, Beijing, China, ⁵Surgery, University of Washington, Seattle, WA, United States

The tissue composition of carotid atherosclerotic plaques is crucial for cardiovascular risk assessment and can be quantified with high-resolution multi-contrast MRI by expert reviewers. The purpose of this work is to develop AI-CASCADE, a fully automated solution for quantitative analysis of carotid MRI, including artery localization, vessel wall segmentation, artery registration and plaque component segmentation. Results in our preliminary study show that AI-CASCADE achieves good agreements with manual results and has great potential as an efficient and reliable clinical tool. 

KOHEI YUDA¹, Takashige Yoshida¹, Yuki Furukawa¹, Masami Yoneyama², Jihun Kwon², Nobuo Kawauchi¹, Johannes M. Peeters ³, and Marc Van Cauteren^3
1Radiology, Tokyo Metropolitan Police Hospital, nakanoku, Japan, ²Philips Japan, Tokyo, Japan, shinagwaku, Japan, ³Philips Healthcare, Best, Netherlands, Netherlands, Netherlands

In high-resolution dual inversion recovery myocardial T2-weighted black-blood (T2W-BB) imaging, using very high acceleration factors with very high resolution can result in degradation of image quality due to insufficient noise removal. In this study, we applied the Compressed-SENSE Artificial Intelligence (CS-AI) framework to further increase the spatial resolution and reduce the scan time. The purpose of this study was to acquire high-resolution myocardial T2W-BB with reduced scan time and compare the image quality between images reconstructed with CS-AI and conventional C-SENSE. 

Gaspar Delso¹, Marc Lebel², Suryanarayanan Kaushik², Graeme McKinnon², Paz Garre³, Pere Pujol³, Daniel Lorenzatti³, José T Ortiz³, Susanna Prat³, Adelina Doltra³, Rosario J Perea³, Teresa M Caralt³, Lluis Mont³, and Marta Sitges^3
1GE Healthcare, Barcelona, Spain, ²GE Healthcare, Waukesha, WI, United States, ³Hospital Clínic de Barcelona, Barcelona, Spain

In this study we evaluate a Deep Learning reconstruction framework with adjustable noise reduction on a database of clinical cases, including Delayed Myocardial Enhancement (MDE), Phase-Sensitive MDE (PSIR) and 3D Heart datasets.

Suvai Gunasekaran¹, Julia Hwang¹, Daming Shen^1,2, Aggelos Katsaggelos^1,3, Mohammed S.M. Elbaz¹, Rod Passman⁴, and Daniel Kim^1,2
1Radiology, Northwestern University, Feinberg School of Medicine, Chicago, IL, United States, ²Biomedical Engineering, Northwestern University, Evanston, IL, United States, ³Electrical and Computer Engineering, Northwestern University, Evanston, IL, United States, ⁴Cardiology, Northwestern University, Feinberg School of Medicine, Chicago, IL, United States

Left atrial (LA) late gadolinium enhancement (LGE) imaging is essential for detecting fibrosis in patients with atrial fibrillation. Unfortunately, slow manual segmentation of LA LGE limits its use in the clinic. The purpose of this study was to develop a fully automated segmentation method for LA LGE images with deep learning. We tested two different U-net architectures that used either 2D or 3D image inputs for training. Our results demonstrate that 3D inputs are superior to 2D, and the 3D U-Net is a promising method to explore further for clinical translation of LA LGE fibrosis quantification.

Mohamad Abdi¹, Daniel S Weller^1,2, and Frederick H Epstein^1,3
1Biomedical Engineering, University of Virginia, Charlottesville, VA, United States, ²Electrical and Computer Engineering, University of Virginia, Charlottesville, VA, United States, ³Radiology, University of Virginia, Charlottesville, VA, United States

Conventionally in MRI, respiratory motion leads to shifts of tissue position in the image domain that correspond to linear phase errors in the k-space domain. For DENSE, in addition to position shifts, respiratory motion is displacement-encoded in the stimulated echo, leading to a constant phase error in the k-space domain. We show that in segmented DENSE acquisitions, motion compensation can be applied using per-segment linear and constant phase corrections. As constant phase corrections using image-based navigators are challenging, we show that deep leaning is potentially an effective solution using simulated training data.

Ina Hanninger¹, Eylem Levelt^2,3, Jennifer J Rayner², Christopher T Rodgers^2,4, Stefan Neubauer², Vicente Grau¹, Oliver J Rider², and Ladislav Valkovic^2,5
1Oxford Institute of Biomedical Engineering, Oxford, United Kingdom, ²Radcliffe Department of Medicine, University of Oxford Centre for Clinical Magnetic Resonance Research, Oxford, United Kingdom, ³University of Leeds, Leeds, United Kingdom, ⁴Wolfson Brain Imaging Centre, Cambridge, United Kingdom, ⁵Slovak Academy of Sciences, Institute of Measurement Science, Bratislava, Slovakia

In this study, Random Forest classification was used on data from 197 subjects to discriminate between non-diabetic, diabetic, and obese patients using 31P-MRS and 1H-MRS measurements of cardiac energetics, along with MRI measures of cardiac function. Achieving 91.67%, 73.08% and 88.89% test accuracies, SHAP feature importances indicate a higher predictive impact of metabolic metrics for classifying the diabetic heart compared to global function metrics. Bayesian networks generated through structure learning of the data further suggests a potential causal association of increased visceral fat, increased LVMass resulting in decreased PCr/ATP, and increased cardiac lipid levels attributed to these disease states.

Shan Huang¹, Yuan Li¹, Ke Shi¹, Yi Zhang¹, Ying-kun Guo², and Zhi-gang Yang^1
1Radiology, West China Hospital, Chengdu, China, ²Radiology, West China Second University Hospital, Chengdu, China

We retrospectively included 100 cardiac amyloidosis (CA) and 217 hypertrophic cardiomyopathy (HCM) patients, aiming to elucidate the value of texture analysis (TA) in non-contrast T2-weighted CMR images of these patients. After the texture features were extracted, machine learning algorithms were used to select the optimal features. The results showed that TA was feasible and reproducible for detecting myocardial tissue alterations and differentiating CA from HCM, even in patients with similar hypertrophy. The radiomics model achieved a comparable diagnostic capacity to late gadolinium enhancement (LGE). Thus, TA might help eliminate the use of contrast agent in the diagnosis of these patients.