Colin Hansen¹, Qi Yang¹, Francois Rheault², Bramsh Qamar³, Owen Williams⁴, Susan Resnick⁴, Eleftherios Garyfallidis³, Adam W Anderson^5,6, Maxime Descoteaux², Bennett A Landman^5,6,7,8, and Kurt G Schilling^5
1Computer Science, Vanderbilt University, Nashville, TN, United States, ²Sherbrooke Connectivity Imaging Laboratory (SCIL), Universite de Sherbrooke, Sherbrooke, QC, Canada, ³Intelligent Systems Engineering, Indiana University Bloomington, Bloomington, IN, United States, ⁴Laboratory of Behavioral Neuroscience, National Institute on Aging, Baltimore, MD, United States, ⁵Vanderbilt University Institute of Imaging Science, Nashville, TN, United States, ⁶Biomedical Engineering, Vanderbilt University, Nashville, TN, United States, ⁷Radiology and Radiological Sciences, Vanderbilt University Medical Center, Nashville, TN, United States, ⁸Electrical Engineering, Vanderbilt University, Nashville, TN, United States

Here, we present a tool and reconstruction method to label white matter pathways directly on structural images without the need for diffusion MRI or tractography. A 3D U-net was trained utilizing 1109 scan sessions where fiber pathways were segmented using two different segmentation schemes. Results on testing datasets show anatomically viable segmentations and moderate-to-high volume overlaps with gold-standard pathways, on par with scan-rescan reproducibility of tractography on the same datasets. We envision the use of this tool for visualizing the expected course of white matter pathways when diffusion data are not available. 

Lavanya Umapathy¹, Gloria J Guzman Perez-Carrillo², Mahesh Bharath Keerthivasan^2,3, Maria I Altbach², Blair Winegar², Craig Weinkauf⁴, and Ali Bilgin^1,2,5
1Department of Electrical and Computer Engineering, University of Arizona, Tucson, AZ, United States, ²Department of Medical Imaging, University of Arizona, Tucson, AZ, United States, ³Siemens Healthcare USA, Tucson, AZ, United States, ⁴Department of Surgery, University of Arizona, Tucson, AZ, United States, ⁵Department of Biomedical Engineering, University of Arizona, Tucson, AZ, United States

Detection and quantification of White Matter Hyperintensities (WMH) on T2-FLAIR images can provide valuable information to assess neurological disease progression. We propose a fully automated stacked generalization ensemble of three orthogonal 3D Convolutional Neural Networks (CNNs), StackGen-Net, to detect WMH on 3D FLAIR images. Each orthogonal CNN predicts WMH on axial, sagittal, and coronal orientations. The posteriors are then combined using a Meta CNN. StackGen-Net outperforms individual CNNs in the ensemble, their ensemble combination, as well as some state-of-the-art deep learning-based models. StackGen-Net can reliably detect and quantify WMH in clinically feasible times, with performance comparable to human inter-observer variability.

Vaanathi Sundaresan¹, Mark Jenkinson¹, Giovanna Zamboni², and Ludovica Griffanti^1
1FMRIB, Wellcome Centre for Integrative Neuroimaging (WIN), Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom, ²Centre for prevention of Stroke and Dementia, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom

We propose a triplanar U-Net ensemble network (TrUE-Net) model for detecting white matter hyperintensities (WMHs) on structural brain MR images. The network uses a combination of loss functions based on the anatomical distribution of WMHs. The model takes T1-weighted and FLAIR images as input channels. The U-Nets in three planes (axial, coronal, sagittal) provides three 2D probability maps, which are combined by averaging across planes to obtain the final probability map. When evaluated on three different cohorts from the MICCAI WMH segmentation challenge dataset, TrUE-Net provided better average performance (Dice=0.78, voxel-wise TPR=0.75), when compared to FSL-BIANCA (Dice=0.69, voxel-wise TPR=0.74).

Sherif Abdulatif¹, Karim Armanious^1,2, Anish Rao Bhaktharaguttu¹, Thomas Küstner², Bin Yang¹, and Sergios Gatidis^2
1University of Stuttgart, Stuttgart, Germany, ²University Hospital Tübingen, Tübingen, Germany

Age is an essential clinical parameter. It is often utilized as a risk factor for various disorders with the potential of influencing therapeutic decisions. However, a discrepancy exists between the chronological age (CA) and the biological age (BA) of an individual due to many factors such as medical history, genetics and lifestyle. In this preliminary work, we propose a novel deep-learning architecture for organ-specific CA estimation from 3D MR volumes for the brain and knee. We hypothesize that the introduced organ-specific approach would enable future analysis of the BA as different organs are expected to exhibit different aging characteristics.

Nicola K Dinsdale¹, Emma Bluemke², Mark Jenkinson¹, and Ana IL Namburete^2
1WIN, University of Oxford, Oxford, United Kingdom, ²IBME, University of Oxford, Oxford, United Kingdom

Nonlinear registration forms a part of standard MRI neuroimaging pipelines but leads to suppression of morphological information. Using attention gates within a convolutional neural network, we explore the effect of the nonlinear registration on age prediction, comparing to linear registration. We show that the network is driven by interpolation effects near the ventricles when trained with nonlinear data, whereas when trained with linear data it considers the whole brain volume. The network may, therefore, be missing cortical changes, limiting the utility of the networks in detecting the early stages of neurological disease.

Muheng Li¹, Yi Xiao¹, Tingyin Liu², Junshen Xu³, Esra Turk⁴, Borjan Gagoski^4,5, Karen Ying¹, Polina Golland^2,3, P. Ellen Grant^4,5, and Elfar Adalsteinsson^3,6
1Department of Engineering Physics, Tsinghua University, Beijing, China, ²Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology, Cambridge, MA, United States, ³Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ⁴Fetal-Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Boston, MA, United States, ⁵Harvard Medical School, Boston, MA, United States, ⁶Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA, United States

We propose a three-dimensional convolutional neural network applied to echo planar EPI time series of pregnant women for the automatic segmentation of the uterus (placenta excluded) and fetal body. The segmentation results are utilized to create a dynamic model for the fetus for retrospective analyses. The 3D dynamic fetal-uterine motion model will provide quantitative information of fetal motion characteristics for diagnostic purposes and may guide future fetal imaging strategies where adaptive, online slice prescription is used to mitigate motion artifacts.

Lufan Liao¹, Xin Zhang², Fenqiang Zhao¹, Jingjiao Lou¹, Li Wang¹, Xiangmin Xu², He Zhang³, and Gang Li^1
1Department of Radiology and BRIC, The University of North Carolina at Chapel Hill, Chapel Hill, CA, United States, ²School of Electronic and Information Engineering, South China University of Technology, GUANGZHOU, China, ³Department of Radiology, Obstetrics and Gynecology Hospital, Fudan University, ShangHai, China

In this study, an end-to-end framework, combining deformable convolution and label distribution learning, is developed for fetal brain age prediction based on MRI. Furthermore, a multi-path architecture is proposed to deal with multi-view MRI scenarios. Experiments on the collected dataset demonstrate that the proposed model achieves promising performance.

Logan A Thorneloe¹, Arjun D Desai^2,3, Garry E Gold^3,4,5, Brian A Hargreaves^2,3,4, Neal K Bangerter⁶, and Akshay S Chaudhari^3
1Electrical Engineering, Brigham Young University, Provo, UT, United States, ²Electrical Engineering, Stanford University, Stanford, CA, United States, ³Radiology, Stanford University, Stanford, CA, United States, ⁴Bioengineering, Stanford University, Stanford, CA, United States, ⁵Orthopedic Surgery, Stanford University, Stanford, CA, United States, ⁶Bioengineering, Imperial College London, London, United Kingdom

Recent advances in deep learning and convolutional neural networks (CNNs) have shown promise for automatic segmentation in MR images. However, because of the stochastic nature of the training process, it is difficult to interpret what information networks learn to represent. In this study, we explore how differences in learned weights between networks can be used to express semantic relationships between different tissues. For cartilage and meniscus segmentation in the knee, we show that network generalizability for segmenting tissues can be measured by distances between networks. We also use these findings to motivate robust training policies for fine-tuning with limited data.

Haoran Sun¹, Xueqing Liu¹, Xinyang Feng¹, Chen Liu², Nanyan Zhu³, Sabrina Josefina Gjerswold-Selleck¹, Hong-Jian Wei^4,5, Pavan Shankar Upadhyayula^5,6, Angeliki Mela^5,6, Cheng-Chia Wu^4,5, Peter Canoll^5,6, Andrew F. Laine¹, John Thomas Vaughan¹, Scott A. Small^7,8,9, and Jia Guo^7,10
1Department of Biomedical Engineering, Columbia University, New York, NY, United States, ²Department of Electrical Engineering, Columbia University, New York, NY, United States, ³Department of Biological Science, Columbia University, New York, NY, United States, ⁴Department of Radiation Oncology, Columbia University, New York, NY, United States, ⁵Columbia University Irving Medical Center, Columbia University, New York, NY, United States, ⁶Department of Pathology and Cell Biology, Columbia University, New York, NY, United States, ⁷Department of Psychiatry, Columbia University, New York, NY, United States, ⁸Departments of Neurology, Columbia University, New York, NY, United States, ⁹Departments of Radiology, Columbia University, New York, NY, United States, ¹⁰Mortimer B. Zickerman Mind Brain Behavior Institute, Columbia University, New York, NY, United States

Cerebral blood volume (CBV) is a hemodynamic correlate of oxygen metabolism and reflects brain activity and function. High-resolution CBV maps can be generated using the steady-state gadolinium-enhanced MRI technique. Recent studies suggest that the exogenous gadolinium based contrast agent (GBCA) can accumulate in the brain after frequent use. Here, we develop and optimize a deep learning algorithm, DeepContrast, which performs equally well as exogenous GBCA in mapping CBV of the normal brain tissue and enhancing glioblastoma. Together, these studies validate our hypothesis that a deep learning approach can potentially replace the need for GBCAs in brain MRI.

Chen Liu^1,2, Nanyan Zhu^1,3, Xinyang Feng^4,5, Frank A Provenzano⁶, John T Vaughan^4,7,8, Scott A Small^6,7,9, and Jia Guo^8,9
1These authors contribute equally to this work and are joint first authors, New York, NY, United States, ²Electrical Engineering, Columbia University, New York, NY, United States, ³Biological Science, Columbia University, New York, NY, United States, ⁴Biomedical Engineering, Columbia University, New York, NY, United States, ⁵Facebook, San Francisco, NY, United States, ⁶Neurology, Columbia University, New York, NY, United States, ⁷Radiology, Columbia University, New York, NY, United States, ⁸Mortimer B. Zuckerman Mind Brain Behavior Institute, Columbia University, New York, NY, United States, ⁹Psychiatry, Columbia University, New York, NY, United States

MRI estimation of cerebral blood volume (CBV) is useful in mapping potential brain function. To obtain high-resolution CBV maps, it typically requires intravenous (IV) injections of Gadolinium-based contrast agents (GBCAs), the use of which has come under new scrutiny. Here, we design and implement a deep learning algorithm, DeepContrast, to estimate GBCA contrast directly from T1-weighted (T1W) structural MRI. The predicted contrast performs equally well as the GBCA-enhanced CBV map even in mapping subtle age-related functional changes in the human brain. Therefore, our study demonstrates the feasibility of substituting GBCA in human brain MRI using DeepContrast.

Jiaren Zou¹, James Balter^1,2, and Yue Cao^1,2,3
1Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, United States, ²Department of Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States, ³Department of Radiology, University of Michigan, Ann Arbor, MI, United States

Conventional nonlinear least squares (LS) methods to fit DCE-MRI to a pharmacokinetic (PK) model are time-consuming. We propose a long Short-Term Memory (LSTM) network that is capable of efficiently learning temporal dependency in sequence data to map PK parameters from single-voxel DCE signals with their corresponding AIFs. The LSTM model showed 90 folds of computation time reduction with comparable performance to LS fitting, while outperforming it for temporally sparsely sampled DCE-MRI. The proposed model can potentially accelerate the data acquisition and PK parameter inference of DCE-MRI.      

Jonathan M Scott¹, Arvin Arani², Armando Manduca², Joshua D Trzasko², John Huston III², Richard L Ehman², and Matthew C Murphy^2
1Medical Scientist Training Program, Mayo Clinic, Rochester, MN, United States, ²Radiology, Mayo Clinic, Rochester, MN, United States

Magnetic Resonance Elastography stiffness estimates in small focal lesions are often inaccurate. The assumption of material homogeneity made by most inversion algorithms likely contributes to these errors. Here we describe a machine-learning based inversion algorithm trained on wave simulations of materials with piecewise smooth stiffness variations (Inhomogeneous Learned Inversion, ILI). We show that ILI offers improved delineation of tumor boundaries over two inversions assuming material homogeneity in a series of 17 patients with stiff meningiomas.

Shahrzad Moinian¹, Viktor Vegh^1,2, and David Reutens^1,2
1Centre for Advanced Imaging, The University of Queensland, Brisbane, Australia, ²ARC Centre for Innovation in Biomedical Imaging Technology, The University of Queensland, Brisbane, Australia

The human cerebral cortex may be divided into functionally different, microarchitectonically distinct areas. While quantitative multi-modal MRI methods can reveal microstructural characteristics of cortical tissue, accurate microarchitectural parcellation of the entire cortex is yet to be attained. Here, we examine a novel method of automated in vivo voxel-wise cortical parcellation which exploits the area-specific microstructural information present in MR fingerprinting (MRF) signals. A Radial Basis Function Support Vector Machine (RBF-SVM) classifier, trained with a volume-based feature representation, achieved a macro-average area under the Receiver Operating Characteristic curve (ROC-AUC) of 0.83.
 

Francesco Grussu^1,2, Stefano B. Blumberg², Marco Battiston¹, Andrada Ianuș³, Saurabh Singh⁴, Fiona Gong⁴, Hayley Whitaker⁴, David Atkinson⁴, Claudia A. M. Gandini Wheeler-Kingshott^1,5,6, Shonit Punwani⁴, Eleftheria Panagiotaki², Thomy Mertzanidou², and Daniel C. Alexander^2
1Queen Square MS Centre, Queen Square Institute of Neurology, Faculty of Brain Sciences, University College London, London, United Kingdom, ²Centre for Medical Image Computing, Department of Computer Science, University College London, London, United Kingdom, ³Champalimaud Research, Champalimaud Centre for the Unknown, Lisbon, Portugal, ⁴Centre for Medical Imaging, University College London, London, United Kingdom, ⁵Department of Brain and Behavioural Sciences, University of Pavia, Pavia, Italy, ⁶Brain MRI 3T Center, IRCCS Mondino Foundation, Pavia, Italy

This work introduces the “Select and retrieve via direct up-sampling” network (SARDU-Net), a new method for model-free, data-driven quantitative MRI (qMRI) experiment design. SARDU-Net identifies informative measurements within lengthy acquisitions and reconstructs fully-sampled signals from a sub-protocol, without prior information on the MRI contrast. It combines two deep networks: a selector, which selects a signal sub-sample, and a predictor, which retrieves input signals. SARDU-Net can be run with standard computational resources and can increase the clinical appeal of qMRI.  Here we demonstrate its potential on qMRI of prostate and spinal cord, two areas where fast acquisitions are key.

Cristian Crisosto^1,2, Andreas Voskrebenzev^1,2, Marcel Gutberlet^1,3, Filip Klimeš^1,2, Frank Wacker^1,2, Till Kaireit^1,2, Gesa Poeler^1,2, Lea Behrendt^1,2, Christopher Korz¹, and Jens Vogel-Claussen^1,2
1Institute of Diagnostic and Interventional Radiology, Hanover Medical School, Hannover, Germany, ²Biomedical Research in Endstage and Obstructive Lung Disease Hannover (BREATH), German Center for Lung Research (DZL), Hannover, Germany, ³Biomedical Research in Endstage and 3 Obstructive Lung Disease Hannover (BREATH), German Center for Lung Research (DZL), Hannover, Germany

Translation and establishment of complex pulmonary magnetic resonance (MR) imaging techniques in the clinics requires a reliable, fully automated and fast calculation. In this work we present a semantic convolutional neural network (CNN) model for lung parenchyma and vessel segmentation in combination with parallelized computation on a high-performance computer to design an end-to-end pipeline for phase-resolved functional lung (PREFUL) MRI. The CNN was trained (n=1118) and validated (n=1064) with manually segmented images by a trained radiologist. Automatic segmentation of lung parenchyma was achieved for all tested images.