Elvisha Dhamala^1,2, Keith W Jamison¹, Sarah M Dennis³, Raihaan Patel^4,5, M Mallar Chakravarty^4,5,6, and Amy Kuceyeski^1,2
1Radiology, Weill Cornell Medicine, New York, NY, United States, ²Neuroscience, Weill Cornell Medicine, New York, NY, United States, ³Sarah Lawrence College, Bronxville, NY, United States, ⁴Biological and Biomedical Engineering, McGill University, Montreal, QC, Canada, ⁵Cerebral Imaging Centre, Douglas Mental Health University Institute, Montreal, QC, Canada, ⁶Psychiatry, McGill University, Montreal, QC, Canada

Structural connectivity (SC) and functional connectivity (FC) can be independently used to predict cognition and show distinct patterns of variance in relation to cognition.  No work identified has yet investigated whether SC and FC can be combined to better predict cognitive abilities. In this work, we aimed to predict cognitive measures in 785 healthy adults using a hybrid structure-function connectome and quantify the most important connections. We show that: 1) hybrid connectomes explain 15% of the variance in individual cognitive measures, and 2) long-range cortico-cortical functional connections and short-range cortico-subcortical and subcortico-subcortical structural connections are most important for the prediction.

Sangtae Ahn¹, Anne Menini², Graeme McKinnon³, Erin M. Gray⁴, Joshua D. Trzasko⁴, John Huston⁴, Matt A. Bernstein⁴, Justin E. Costello⁵, Thomas K. F. Foo¹, and Christopher J. Hardy^1
1GE Research, Niskayuna, NY, United States, ²GE Healthcare, Menlo Park, CA, United States, ³GE Healthcare, Waukesha, WI, United States, ⁴Department of Radiology, Mayo Clinic College of Medicine, Rochester, MN, United States, ⁵Walter Reed National Military Medical Center, Bethesda, MD, United States

Deep learning-based undersampled MRI reconstructions can result in visible blurring, with loss of fine detail. We investigate here various structural similarity (SSIM) based loss functions for training a compressed-sensing unrolled iterative reconstruction, and their impact on reconstructed images. The conventional unweighted SSIM has been used both as a loss function, and, more generally, for assessing perceived image quality in various applications. Here we demonstrate that using an appropriately weighted SSIM for the loss function yields better reconstruction of small anatomical features compared to L1 and conventional SSIM loss functions, without introducing image artifacts.

Dahan Kim^1,2, Laura Eisenmenger³, and Kevin M. Johnson^3,4
1Department of Medical Physics, University of Wisconsin, Madison, WI, United States, ²Department of Physics, University of Wisconsin, Madison, WI, United States, ³Department of Radiology, University of Wisconsin, Madison, WI, United States, ⁴Department of Medical Physics, University of Wisconsin, Middleton, WI, United States

4D-flow MRI suffers from long scan time due to a minimum of four velocity encodings necessary to solve for three velocity components and the reference background phase. We examine the feasibility of using machine learning (ML) to determine the background phase and hence three velocity components from only three flow encodings. The results show that ML is capable of estimating three-directional velocities from three flow encodings with high accuracy (1.5%-3.8% velocity underestimation) and high precision (R²=0.975). These findings indicate that 4D-flow MRI can be accelerated without requiring a dedicated reference scan, with a scan time reduction of 25%.

Gian Franco Piredda^1,2,3, Virginie Piskin¹, Vincent Dunet², Gibran Manasseh², Mário J Fartaria^1,2,3, Till Huelnhagen^1,2,3, Jean-Philippe Thiran^2,3, Tobias Kober^1,2,3, and Ricardo Corredor-Jerez^1,2,3
1Advanced Clinical Imaging Technology, Siemens Healthcare AG, Lausanne, Switzerland, ²Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ³LTS5, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland

Multiple sclerosis studies following the widely accepted MAGNIMS protocol guidelines might lack non-contrast-enhanced T₁-weighted acquisitions as they are only considered optional. Most existing automated tools to perform morphological brain analyses are, however, tuned to non-contrast T₁-weighted images. This work investigates the use of deep learning architectures for the generation of pre-Gadolinium from post-Gadolinium image volumes. Two generative models were tested for this purpose. Both were found to yield similar contrast information as the original non-contrast T₁-weighted images. Quantitative comparison using an automated brain segmentation on original and synthesized non-contrast T₁-weighted images showed good correlation (r=0.99) and low bias (<0.7 ml).

Marcia Sahaya Louis¹, Eduardo Coello², Huijun Liao², Ajay Joshi³, and Alexander Lin^2
1ECE, Boston University, Boston, MA, United States, ²Radiology, Brigham and Women's hospital, Boston, MA, United States, ³Boston University, Boston, MA, United States

Water is present in the brain tissue at a concentration that is at least four orders of magnitude higher than metabolites of interest. As a result, it is necessary to suppress the water resonance so that the brain metabolites of interest can be better visualized and quantified. This work presents a neural network model for extracting the metabolites spectrum from non-water-suppressed proton magnetic resonance spectra. The autoencoder model learns a vector field for mapping the water signal to a lower-dimensional manifold and accurately reconstructs the metabolite spectra as compared to water-suppressed spectra from the same subject.

Berkin Bilgic^1,2, Long Wang¹, Enhao Gong¹, Greg Zaharchuk^1,3, and Tao Zhang^1
1Subtle Medical Inc, Menlo Park, CA, United States, ²Martinos Center for Biomedical Imaging, MGH/Harvard, Charlestown, MA, United States, ³Stanford University, Stanford, CA, United States

The long scan time of 3D isotropic MRI (often 5 minutes or longer) has limited the wide clinical adoption despite the apparent advantages. For many clinical sites, shorter 2D sequences are used routinely in brain MRI exams instead. The latest development of deep learning (DL) has demonstrated the feasibility of significant resolution improvement from low resolution acquisitions. In this work, we propose a deep learning method to synthesize 3D isotropic FLAIR images from 2D FLAIR acquisition with 5mm slice thickness. To demonstrate the generalizability, the proposed method is validated on both simulated and real 2D FLAIR datasets.

Yi-Tien Li^1,2, Yi-Wen Chen¹, David Yen-Ting Chen^1,3, and Chi-Jen Chen^1,4
1Department of Radiology, Taipei Medical University - Shuang Ho Hospital, New Taipei, Taiwan, ²Institute of Biomedical Engineering, National Taiwan University, Taipei, Taiwan, ³Department of Radiology, Stanford University, Palo Alto, CA, United States, ⁴School of Medicine, College of Medicine, Taipei Medical University, Taipei, Taiwan

A huge amount of T2-FLAIR images with appearance of white matter hyperintencities (WMH) were used. 1368 cases from one hospital were selected as the training set. Another 100 cases from the same hospital and 200 cases from the other 2 different hospitals were treated as the independent test set. Based on multi-class U-SegNet approach, it can achieve the highest F1 score (same hospital: 90.01%; different hospital: 86.52%) in the test set compared with other approaches.  The result suggested that the multi-class segmentation approach is more resilient to the variations of MR image parameters than the single label segmentation approach.

Hualei Shen¹, Chenyu Wang^1,2, Kain Kyle², Chun-Chien Shieh^2,3, Lynette Masters⁴, Fernando Calamante^1,5, Dacheng Tao⁶, and Michael Barnett^1,2
1Brain and Mind Centre, the University of Sydney, Sydney, Australia, ²Sydney Neuroimaging Analysis Centre, Sydney, Australia, ³Sydney Medical School, the University of Sydney, Sydney, Australia, ⁴I-MED Radiology Network, Sydney, Australia, ⁵Sydney Imaging and School of Biomedical Engineering, the University of Sydney, Sydney, Australia, ⁶School of Computer Science, the University of Sydney, Sydney, Australia

Brain extraction from 3D MRI datasets using existing 3D U-Net convolutional neural networks suffers from limited accuracy. Our proposed method overcame this challenge by combining a 3D U-Net with voxel-wise spatial information. The model was trained with 1,615 T1 volumes and tested on another 601 T1 volumes, both with expertly segmented labels. Results indicated that our method significantly improved the accuracy of brain extraction over a conventional 3D U-Net. The trained model extracts the brain from a T1 volume in ~2 minutes and has been deployed for routine image analyses at the Sydney Neuroimaging Analysis Centre.

Maxime Bertrait¹, Pascal Ceccaldi¹, Boris Mailhé¹, Youngjin Yoo¹, and Mariappan S. Nadar^1
1Digital Technology and Innovation, Siemens Healthineers, Princeton, NJ, United States

In Magnetic Resonance Imaging, acquisition protocol may varies from one clinical task to another affecting the resulting reconstructed scan in terms of field of view and resolution. In research, 3D acquired MRI scans are widely available providing high quality isotropic medical images but is far from what can exist in clinical environment such as 2D multi-slices with thick slices acquisition that can provide anisotropic medical images. We then present a framework, through a brain extraction task, called Dual Recursive Refiner able to work with both acquisitions. The presented framework outperforms baseline architectures for segmentation on both isotropic and anisotropic data.

Theodor Rumetshofer¹, Tor Olof Strandberg², Peter Mannfolk³, Andreas Jönsen⁴, Markus Nilsson¹, Johan Mårtensson¹, and Pia Maly Sundgren^1,5
1Department of Clinical Sciences Lund/Diagnostic Radiology, Lund University, Lund, Sweden, ²Clinical Memory Research Unit, Department of Clinical Sciences, Malmö, Lund University, Lund, Sweden, ³Clinical Imaging and Physiology, Skåne University Hospital, Lund, Sweden, ⁴Department of Reumatology, Skåne University Hospital, Lund, Sweden, ⁵Department of Clinical Sciences/Centre for Imaging and Function, Skåne University Hospital, Lund, Sweden

Evaluating white matter hyperintensities (WMHs) in neuropsychiatric systemic lupus erythematosus (NPSLE) is a challenging task. Multimodal MRI images in combination with unsupervised machine characterization can provide a powerful tool to investigate the spatial WHM distribution of relevant phenotypes. Automatically segmented WMH maps were spatially allocated to a white matter tract atlas. Cluster analysis was applied on this tract-wise lesion-load map to obtain subtypes with a distinct WMH damage profile. This approach on microstructural changes could help to identify specific progression pattern which may improve the accuracy of NPSLE classification.

Po-Jui Lu^1,2,3, Reza Rahmanzadeh^1,2, Riccardo Galbusera^1,2, Matthias Weigel^1,2,4, Youngjin Yoo³, Pascal Ceccaldi³, Yi Wang⁵, Jens Kuhle², Ludwig Kappos^1,2, Philippe Cattin⁶, Benjamin Odry⁷, Eli Gibson³, and Cristina Granziera^1,2
1Translational Imaging in Neurology (ThINk) Basel, Department of Medicine and Biomedical Engineering, University Hospital Basel and University of Basel, Basel, Switzerland, ²Neurologic Clinic and Policlinic, Departments of Medicine, Clinical Research and Biomedical Engineering, University Hospital Basel and University of Basel, Basel, Switzerland, ³Digital Technology and Innovation, Siemens Healthineers, Princeton, NJ, United States, ⁴Radiological Physics, Department of Radiology, University Hospital Basel, Basel, Switzerland, ⁵Department of Radiology, Weill Cornell Medical College, New York, NY, United States, ⁶Center for medical Image Analysis & Navigation, Department of Biomedical Engineering, University of Basel, Basel, Switzerland, ⁷Covera Health, New York, NY, United States

White matter lesions in multiple sclerosis patients exhibit distinct characteristics depending on their locations in the brain. Multiple quantitative MR sequences sensitive to white matter micro-environment are necessary for the assessment of those lesions; but how to judge which sequences contain the most relevant information remains a challenge. In this abstract, we are proposing a convolutional neural network with a gated attention mechanism to quantify the importance of MR metrics in classifying juxtacortical and periventricular lesions. The results show the statistically significant order of quantitative importance of metrics, one step closer to combining more relevant metrics for better interpretation.

Samantha J Ma¹, Mona Sharifi Sarabi¹, Kai Wang¹, Soroush Heidari Pahlavian¹, Wenli Tan¹, Madison Lodge¹, Lirong Yan¹, Yonggang Shi¹, and Danny JJ Wang^1
1University of Southern California, Los Angeles, CA, United States

Cerebral small vessels are largely inaccessible to existing clinical in vivo imaging technologies. As such, early cerebral microvascular morphological changes in small vessel disease (SVD) are difficult to evaluate. A deep learning (DL)-based algorithm was developed to automatically segment lenticulostriate arteries (LSAs) in 3D black blood images acquired at 3T. Using manual segmentations as supervision, 3D segmentation of LSAs is demonstrated to be feasible with relatively high performance and can serve as a useful tool for quantitative morphometric analysis in patients with cerebral SVD.

Jonathan Doucette^1,2, Christian Kames^1,2, and Alexander Rauscher^1,3,4
1UBC MRI Research Centre, Vancouver, BC, Canada, ²Department of Physics & Astronomy, University of British Columbia, Vancouver, BC, Canada, ³Department of Pediatrics, Faculty of Medicine, University of British Columbia, Vancouver, BC, Canada, ⁴Division of Neurology, Faculty of Medicine, University of British Columbia, Vancouver, BC, Canada

In this work we use Bayesian learning methods to investigate data-driven approaches to parameter inference of multi-exponential white matter signals. Multi spin-echo (MSE) signals are simulated by solving the Bloch-Torrey on 2D geometries containing myelinated axons, and a conditional variational autoencoder (CVAE) model is used to learn to map simulated signals to posterior parameter distributions. This approach allows for the mapping of MSE signals directly to physical parameter vectors without expensive post-processing. We demonstrate the effectiveness of this model through the simultaneous inference of the myelin water fraction, flip angle, intra-/extracullar water $$$T_2$$$, myelin water $$$T_2$$$, and myelin g-ratio.
 

Huanyu Luo¹, Jing Xue¹, Yunyun Duan¹, Cheng Xu¹, Jonathan Tamir², Srivathsa Pasumarthi Venkata², and Yaou Liu^1
1Radiology, Beijing Tiantan Hospital, Beijing, China, ²Subtle Medical Inc, Menlo Park, CA, United States

The reported gadolinium deposition phenomenon has caused extensive concern in the radiology community. This study focuses on validating the clinical performance of a proposed deep learning architecture which can significantly reduce the dosage of gadolinium-based contrast agents (GBCA) in brain MRI. The results suggest that the synthesized contrast images using deep learning with reduced GBCA dose can maintain its diagnostic quality under certain clinical circumstances.

Keerthi Sravan Ravi^1,2, Sairam Geethanath², Girish Srinivasan³, Rahul Sharma⁴, Sachin R Jambawalikar⁴, Angela Lignelli-Dipple⁴, and John Thomas Vaughan Jr.^2
1Biomedical Engineering, Columbia University, New York, NY, United States, ²Columbia Magnetic Resonance Research Center, Columbia University, New York, NY, United States, ³MediYantri Inc., Palatine, IL, United States, ⁴Columbia University Irving Medical Center, New York, NY, United States

A 2015 survey indicates that burnout of radiologists was seventh highest among all physicians. In this work, two neural networks are designed and trained to generate text-based first read radiology reports. Existing tools are leveraged to perform registration and then brain tumour segmentation. Feature vectors are constructed utilising the information extracted from the segmentation masks. These feature vectors are fed to the neural networks to train against a radiologist’s reports on fifty subjects. The neural networks along with image statistics are able to characterise tumour type, mass effect and edema and report tumour volumetry; compiled as a first-read radiology report.