Yixue Feng¹, Bramsh Qamar Chandio^1,2, Tamoghna Chattopadhyay¹, Sophia I. Thomopoulos¹, Conor Owens-Walton¹, Neda Jahanshad¹, Eleftherios Garyfallidis², and Paul M. Thompson^1
1Imaging Genetics Center, Mark and Mary Stevens Neuroimaging and Informatics Institute, Keck School of Medicine, University of Southern California, Marina Del Rey, CA, United States, ²Department of Intelligent Systems Engineering, School of Informatics, Computing, and Engineering, Indiana University Bloomington, Bloomington, IN, United States

We present a deep generative model to autoencode tractography streamlines into a smooth low dimensional latent distribution, which captures their spatial and sequential information with 1D convolutional layers. Using linear interpolation, we show that the learned latent space translates smoothly into the streamline space and can decode meaningful outputs from sampled points. This allows for inference on new data and direct use of Euclidean distance on the embeddings for downstream tasks, such as bundle labeling, quantitative inter-subject comparisons, and group statistics.

Zechen Zhou¹, Peter Börnert², and Chun Yuan^3
1Philips Research North America, Seattle, WA, United States, ²Philips Research Hamburg, Hamburg, Germany, ³Vascular Imaging Lab, University of Washington, Seattle, WA, United States

Supervised learning is widely used for deep learning based image quality enhancement for improved clinical diagnosis. However, the difficulties to acquire a large number of high-quality reference image for different MR applications can limit its generalization performance. An unsupervised domain adaptation (DA) approach is proposed and incorporated into the deep learning based image enhancement framework, which improves the performance of trained network on new datasets. Preliminary evaluation on point spread function enhanced turbo spin echo imaging has showed that the unsupervised DA approach can provide more stabilized image sharpness improvement without severe amplified noise.

Ozan Genc¹, Pranathi Chunduru², Annette Molinaro¹, Valentina Pedoia¹, Susan Chang¹, Javier Villanueva-Meyer¹, and Janine Lupo Palladino^1
1University of California San Francisco, San Francisco, CA, United States, ²Johnson & Johnson, San Francisco, CA, United States

Missing value imputation is an important concept in statistical analyses. We utilized conditional GAN based deep learning models to learn missing contrasts in MR images. We trained two deep learning models (FSE to FLAIR and T1 post GAD to T1 pre-GAD) for MR image conversion for missing value imputation.  The model performances are evaluated by visual examination and comparing SSIM values. We observed that these models can learn the output contrast.

Yan-Lin Liu¹, Zhongwei Chen², Youfan Zhao², Yang Zhang^1,3, Jiejie Zhou², Jeon-Hor Chen¹, Ke Nie³, Meihao Wang², and Min-Ying Su^1
1Department of Radiological Sciences, University of California, Irvine, CA, United States, ²Department of Radiology, The First Affiliated Hospital of Wenzhou Medical University, Wenzhou, China, ³Department of Radiation Oncology, Rutgers-Cancer Institute of New Jersey, Robert Wood Johnson Medical School, New Brunswick, NJ, United States

268 patients with DCE-MRI and mammography were analyzed to evaluate the diagnostic performance of radiomics models. The dataset was split to 202 (146 malignant 56 benign) for training, and 66 (48 malignant 18 benign) for testing. The MIP of MR contrast enhancement maps was generated to simulate the CC and MLO view as guidance for manual ROI drawing on mammography. The models were built using features extracted by PyRadiomics. Combined MRI and mammography features can reach 89.6% accuracy in training and 83.3% in testing datasets, and the addition of mammography can improve specificity while maintaining high sensitivity of MRI.

Xing Lu¹, Yajun Ma¹, Kody Xu¹, Saeed Jerban¹, Hyungseok Jang¹, Chun-Nan Hsu², Amilcare Gentili^1,3, Eric Y Chang^1,3, and Jiang Du^1
1Department of Radiology, University of California, San Diego, San Diego, CA, United States, ²Department of Neurosciences, University of California, San Diego, San Diego, CA, United States, ³Radiology Service, Veterans Affairs San Diego Healthcare System, San Diego, CA, United States

In this study, we proposed end-to-end deep learning convolutional neural networks to perform simultaneous multi-tissue segmentation and multi-parameter quantification (MSMQ-Net) on the knee without and with physical constraints. The performance robustness of MSMQ-Net was also evaluated using reduced input magnetic resonance images.  Results demonstrated the potential of MSMQ-Net for fast and accurate UTE-MRI analysis of the knee, a “whole-organ” approach which is impossible with conventional clinical MRI.

Nate Tran¹, Jacob Ellison¹, Oluwaseun Adegbite¹, James Golden¹, Yan Li¹, Joanna Phillips², Devika Nair¹, Anny Shai², Annette Molinaro², Valentina Pedoia¹, Javier Villanueva-Meyer¹, Mitchel Berger², Shawn Hervey-Jumper², Aghi Manish², Susan Chang², and Janine Lupo^1
1Radiology & Biomedical Imaging, University of California, San Francisco, SAN FRANCISCO, CA, United States, ²Neurological Surgery, University of California, San Francisco, SAN FRANCISCO, CA, United States

Using spectrum obtained at the spatial location of 549 tissue samples from 261 newly diagnosed patients with glioma, we trained and tested an support vector regression (SVR) model on individual metabolites, and a 1D-CNN model on the whole spectrum, to predict tumor biology such as cellularity, Ki-67, and tumor aggressiveness. A regression based 1D-CNN model using the entire spectrum pre-trained on a similar classification task outperformed the SVR model using metabolite peak heights.

Michael Hirano¹, Anum S. Kazerouni², Mladen Zecevic², Laura C. Kennedy³, Shaveta Vinayak², Habib Rahbar², Matthew J. Nyflot², Suzanne Dintzis², and Savannah C. Partridge^2
1University of Washingon, Seattle, WA, United States, ²University of Washington, Seattle, WA, United States, ³Vanderbilt University, Nashville, TN, United States

Radiomics is an advancing field of medical image analysis based on extracting large sets of quantitative features that can be used for outcome modeling for clinical decision support.  Our study investigated the value of radiomics features extracted from pre-treatment dynamic contrast-enhanced MRI for the prediction of neoadjuvant chemotherapy response in patients with triple-negative breast cancer. In a retrospective cohort of 103 TNBC patients, radiomics-based models using post-contrast images and kinetics maps were moderately predictive of pathologic response, and lesion size and shape features were the most consistent predictors across all image types.

Emmanuelle M. M. Weber¹, Xucheng Zhu², Patrick Koon², Anja Brau², Shreyas Vasanawala¹, and Jennifer A. McNab^1
1Stanford, Stanford, CA, United States, ²GE Healthcare, Menlo Park, CA, United States

To reduce artifacts in free breathing single-shot diffusion MRI of the liver, UNet based convolutional neural networks were trained to predict breath-hold data from free-breathing data using:  1) simulated data based on a digital phantom and 2) 31 scans of a healthy volunteer.  The developed networks successfully reduced motion induced artifacts in DWI images.

Siddharth Srinivasan Iyer^1,2, Christopher M. Sandino³, Mahmut Yurt³, Xiaozhi Cao², Congyu Liao², Sophie Schauman², and Kawin Setsompop^2,3
1Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ²Department of Radiology, Stanford University, Stanford, CA, United States, ³Department of Electrical Engineering, Stanford University, Stanford, CA, United States

Recent developments in spatiotemporal MRI techniques enable whole-brain multi-parametric mapping in incredibly short acquisition times through highly-efficient k-space encoding, subspace reconstruction and carefully-designed regularization. However, this comes at the cost of long reconstruction times making such methods difficult to integrate into clinical practice.

This abstract proposes a framework denoted SMILR (pronounced smile-r) to reduce the reconstruction times of subspace methods from multiple hours to a few minutes through machine learning. To evaluate performance, the framework is applied to multi-axis spiral projection MRF (denoted SPI-MRF) where it achieves improved reconstruction over conventional subspace reconstruction with locally low-rank at ~16-20x faster speed.

Neel Dey¹, Jo Schlemper², Seyed Sadegh Mohseni Salehi², Bo Zhou³, and Michal Sofka^2
1Computer Science and Engineering, New York University, New York City, NY, United States, ²Hyperfine, New York City, NY, United States, ³Yale University, New Haven, CT, United States

We propose an unsupervised contrastive representation learning framework for deformable and diffeomorphic multi-modality MR image registration. The proposed deep network and data-driven objective function yield improved registration performance in terms of anatomical volume overlap over several previous hand-crafted objectives such as Mutual Information and others. For fair comparison, our experiments train all methods over the entire range of a key registration hyperparameter controlling deformation smoothness using conditional registration hypernetworks. T1w and T2w brain MRI registration improvements are presented across a large cohort of 1041 high-field 3T research-grade acquisitions while maintaining comparable deformation smoothness and invertibility characteristics to previous methods.

Jiahao Lin^1,2, Miao Qi^1,3, Chuthaporn Surawech^1,4,5, Steven S Raman¹, Holden Wu¹, and Kyunghyun Sung^1
1Department of Radiology, University of California, Los Angeles, Los Angeles, CA, United States, ²Department of Electrical and Computer Engineering, University of California, Los Angeles, Los Angeles, CA, United States, ³Department of Radiology, The First Affiliated Hospital of China Medical University, Shenyang, China, ⁴Department of Radiology, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand, ⁵Division of Diagnostic Radiology, Department of Radiology, King Chulalongkorn Memorial Hospital, Bangkok, Thailand

Multi-slice two-dimensional turbo spin-echo (2D TSE) imaging is commonly used for its excellent in-plane resolution. However, multiple 2D TSE scans are acquired due to their poor through-plane resolution. In this study, we propose a novel slice-profile-based transformation super-resolution (SPTSR) framework for through-plane super-resolution (SR) of multi-slice 2D TSE imaging. We utilized a large clinical MRI dataset in this study. We demonstrate the effectiveness of our proposed SPTSR framework in 5.5x through-plane SR by both the visual comparison results and reader study results.

Sihao Chen¹, Weijie Gan¹, Cihat Eldeniz¹, Ulugbek S. Kamilov¹, Tyler J. Fraum¹, and Hongyu An^1
1Washington University in St. Louis, Saint Louis, MO, United States

Dynamic contrast-enhanced MRI (DCE-MRI) of the liver offers structural and functional information for assessing the contrast uptake visually. However, respiratory motion and the requirement of high temporal resolution make it difficult to generate high-quality DCE-MRI. In this study, we proposed a novel deep learning based motion transformation integrated forward-Fourier (DL-MOTIF) reconstruction using motion fields derived from a deep learning Phase2Phase (P2P) network and deep learning priors from a residual network on severely undersampled DCE. This approach reconstructs sharp motion-free DCE images with artifacts removal by incorporating deep learning motion fields for motion integration and deep learning priors for regularization.

Quan Dou¹, Xue Feng¹, and Craig H. Meyer^1
1Biomedical Engineering, University of Virginia, Charlottesville, VA, United States

Fast imaging techniques can speed up MRI acquisition but can also be corrupted by noise, reconstruction artifacts, and motion artifacts in a clinical setting. A deep learning-based method was developed to reduce imaging noise and artifacts. A network trained with the supervised approach improved the image quality for both simulated and in vivo data.

Jiehua Li¹, Pan Su^1,2, Rao Gullapalli¹, and Jiachen Zhuo^1
1Diagnostic Radiology and Nuclear Medicine, University of Maryland School of Medicine, Baltimore, MD, United States, ²Siemens Medical Solutions USA, Inc., Malvern, PA, United States

The perivascular space (PVS) plays a major role in brain waste clearance and brain metabolic homeostasis. Enlarged PVS (ePVS) is associated with many neurological disorders. ePVS is best depicted as hyper-intensities T2w images and can be reliable quantified with both the 3D T1w and T2w images. However many studies opt to acquire 3D FLAIR images instead of T2w due to its high specificity to white matter abnormalities (e.g. the ADNI study). Here we show that deep learning techniques can be used to synthesize T2w images from T1w & FLAIR images and improve the ePVS quantification in absence of T2w images.

Richard James Adams¹, Yong Chen¹, Pew-Thian Yap², and Dan Ma^1
1Case Western Reserve University, Cleveland, OH, United States, ²University of North Carolina, Chapel Hill, NC, United States

Magnetic resonance fingerprinting (MRF) simultaneously quantifies multiple tissue properties. Deep learning accelerates MRF’s tissue mapping time; however, previous deep learning MRF models are supervised, requiring ground truths of tissue property maps. It is challenging to acquire quality reference maps, especially as the number of tissue parameters increases. We propose a self-supervised model informed by the physical model of the MRF acquisition without requiring ground truth references. Additionally, we construct a forward model that directly estimates the gradients of the Bloch equations. This approach is flexible for modeling MRF sequences with pseudo-randomized sequence designs where an analytical model is not available.