Huan Minh Luu¹, Dong-Hyun Kim¹, Seung-Hong Choi², and Sung-Hong Park^1
1Magnetic Resonance Imaging Laboratory, Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology, Daejeon, Korea, Republic of, ²Department of Radiology, Seoul National University Hospital, Seoul, Korea, Republic of

Quantitative magnetization transfer (qMT) imaging provides quantitative measures of magnetization transfer properties, but the method itself suffers from long acquisition and processing time. Previous research has looked into the application of deep learning to accelerate qMT imaging. Specifically, a network called qMTNet was proposed to accelerate both data acquisition and fitting. In this study, we propose qMTNet⁺, an improved version of qMTNet, that accomplishes both acceleration tasks as well as generation of missing data with a single residual network. Results showed that qMTNet⁺ improves the quality of generated MT images and fitted qMT parameters compared to qMTNet.

Quan Chen¹, Huajun She¹, Zhijun Wang¹, and Yiping P. Du^1
1School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China

Acceleration of the myelin water fraction (MWF) mapping at R=6 using a Global and Local Deep Dictionary Learning Network (GLDDL) is demonstrated in this study. The global and the local spatiotemporal correlations in the relaxations are learned simultaneously. The global temporal encode-decoder layers are utilized to reduce the computational complexity of the local DL network and improve the denoising performance. The deep DL network utilizes the merits of traditional DL and deep learning to improve the reconstruction. The high-quality MWF maps obtained from the GLDDL network has demonstrated the feasibility to accelerate the whole brain MWF mapping in 1 minute.

Jing Cheng¹, Yuanyuan Liu¹, Xin Liu¹, Hairong Zheng¹, Yanjie Zhu¹, and Dong Liang^1
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China

Existing deep learning-based methods for rapid MR parametric mapping often use the reference parametric maps fitted from fully sampled images to train the networks. Nevertheless, the fitted parametric map is sensitive to the noise and the fitting algorithms. In this work, we proposed to incorporate the quantitative physical model into the deep learning framework to simultaneously reconstruct the parameter-weighted images and generate the parametric map without the reference parametric maps. Experimental results on the quantitative MR T_1ρ mapping show the promising performance of the proposed framework.

Serge Vasylechko^1,2, Simon K. Warfield^1,2, Sila Kurugol^1,2, and Onur Afacan^1,2
1Boston Children's Hospital, Boston, MA, United States, ²Harvard Medical School, Boston, MA, United States

Deep learning methods have the potential to improve quantitative MRI methods. However, performance of deep learning methods is highly sensitive to the amount of available training data. In this work, we propose generating a substantial amount of 3D synthetic data, and demonstrate its application to myelin water fraction mapping. A parameter sampling model is designed within a naturally occurring range of multi-component T2 distributions to generate a large set of varying synthetic signals. This model is combined with a spatially varying sampling model that generates a multitude of spatial deformations and signal perturbations.

Jinwei Zhang¹, Hang Zhang¹, Pascal Spincemaille², Mert Sabuncu³, Thanh Nguyen², Ilhami Kovanlikaya², and Yi Wang^2
1Cornell University, New York, NY, United States, ²Weill Cornell Medical College, New York, NY, United States, ³Cornell University, Ithaca, NY, United States

To accelerate the acquisition time of quantitative susceptibility mapping (QSM) using a 3D multi-echo gradient echo (mGRE) sequence, an unrolled multi-channel deep ADMM reconstruction network with a LOUPE-ST based 2D variable density sampling pattern optimization module is trained to optimize both the k-space under-sampling pattern and the reconstruction. Prospectively under-sampled k-space data are acquired using a modified mGRE sequence and reconstructed by the trained unrolled network. Prospective study shows the learned sampling pattern achieves better image quality in QSM compared to a manually designed pattern.

Srivathsa Pasumarthi¹, Jon Tamir², Enhao Gong², Greg Zaharchuk², and Tao Zhang^2
1Subtle Medical Inc, Menlo Park, CA, United States, ²Subtle Medical Inc., Menlo Park, CA, United States

Deep learning (DL) has recently become a promising technique to address the safety concerns for Gadolinium-based Contrast Agents (GBCAs) in MRI. Studies have shown that high quality contrast-enhanced images can be generated by DL with only a small fraction of the standard dose, and in some cases no dose at all. To build upon existing research that has heavily focused on qualitative evaluation by radiologists, this work proposes an automated quantitative evaluation scheme for the GBCA dose reduction using DL.

Patrick Kinz¹, Sebastian Thomas¹, and Lothar R. Schad^1
1Computer Assisted Clinical Medicine, Heidelberg University, Medical Faculty Mannheim, Mannheim, Germany

MRI-based mapping of oxygen extraction fraction with QSM and qBOLD is a non-invasive diagnostic tool with many possible applications. But current reconstruction methods based on quasi-Newton (QN) methods are very dependent on accurate parameter initialization. We compare the standard QN method with a clustering approach, and with a new initialization method in form of an ANN, which we previously developed to replace QN. This leads to reconstructed images with much less noise, compared to the direct output of the ANN and keeps the effect of reduced intersubject variability.

Elisa Moya-Sáez^1,2, Rodrigo de Luis-García¹, and Carlos Alberola-López^1
1University of Valladolid, Valladolid, Spain, ²Fundación Científica AECC, Valladolid, Spain

Synthetic MRI is gaining popularity due to its ability to generate realistic MR images. However, T1, T2, and PD maps are rarely used despite they provide the information needed to synthesize any image modality by applying the appropriate theoretical equations derived from MR physics or through involved simulation procedures. In this work we propose an extension of an state-of-the-art standard CNN to a self-supervised CNN by including MR physics priors to tackle confounding factors not considered in the equations, while bypassing the need of costly simulations. Our approach yields both realistic maps and weighted images from real data.

Xi Lin¹, Qinqin Yang¹, Jianfeng Bao², Shuhui Cai¹, Zhong Chen¹, Congbo Cai¹, and Jingliang Cheng^2
1Department of Electronic Science, Xiamen University, Xiamen, China, ²Department of Magnetic Resonance Imaging, The First Affiliated Hospital of Zhengzhou University, Zhengzhou University, Zhengzhou, China

Quantitative T₂ mapping of abdomen has the potential for many clinical scenarios, such as for distinguishing cirrhosis and characterizing renal lesions. In this study, we applied overlapping echo acquisition together with deep learning-based reconstruction to achieving T₂ mapping of abdomen in single shot for the first time. The resulting abdominal T₂ values are in good agreement with the reported ones. This method makes high temporal resolution quantitative abdominal imaging possible, and has strong motion robustness.

Nikkita Khattar¹, Zhaoyuan Gong¹, Matthew Kiely¹, Curtis Triebswetter¹, Maryam H. Alsameen¹, and Mustapha Bouhrara^1
1Laboratory of Clinical Investigation, National Institute on Aging, Baltimore, MD, United States

MRI mapping of myelin water fraction (MWF), a surrogate of myelin content, has provided important insights into brain maturation and neurodegeneration, with promising potential use to quantify disease progression or therapeutic effect. Besides the complex modeling, MWF imaging, using either conventional or advanced methods such as the BMC-mcDESPOT approach, requires prolonged acquisition times, hampering their integration in clinical investigations. In this proof-of-concept work, we propose an artificial neural network model to derive MWF maps from conventional relaxation times and proton density maps. This work opens a way to further developments for practical and fast MWF imaging.

Neta Stern¹, Dvir Radunsky¹, Tamar Blumenfeld-Katzir¹, Yigal Chechik^2,3, Chen Solomon¹, and Noam Ben-Eliezer^1,4,5
1Department of Biomedical Engineering, Tel Aviv University, Tel Aviv, Israel, ²Department of Orthopedics, Shamir Medical Center, Zerifin, Israel, ³Sackler Faculty of Medicine, Tel Aviv University, Tel Aviv, Israel, ⁴Sagol School of Neuroscience, Tel Aviv University, Tel Aviv, Israel, ⁵Center for Advanced Imaging Innovation and Research (CAI2R), New-York University, Langone Medical Center, NY, United States

This study evaluates the performance of Marchenko-Pastur Principle Component Analysis (MP-PCA) denoising algorithm for improving T₂ mapping precision using the EMC algorithm. Denoising efficiency was successfully validated on in vivo brain and knee scans, showing an increase in T₂ maps’ precision while preserving anatomical features. This enables precise mapping of T₂ values at high-resolutions. The proposed method can benefit a wide range of clinical applications, and carries the potential to enhance the detection of abnormalities on both T₂ weighted images and qualitative T2 maps.

Divya Varadarajan^1,2, Katie Bouman³, Bruce Fischl*^1,2,4, and Adrian Dalca*^1,5
1Martinos Center for Biomedical Imaging, Charlestown, MA, United States, ²Department of Radiology, Harvard Medical School, Boston, MA, United States, ³Department of Computing and Mathematical Sciences, California Institute of Technology, Pasadena, CA, United States, ⁴Massachusetts General Hospital, Boston, MA, United States, ⁵Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States

Estimating tissue properties and synthesizing contrasts can ease the need for long acquisitions and have many clinical applications.  In this work we propose an unsupervised deep-learning strategy that employs the FLASH MRI model. The method jointly estimates the T1, T2* and PD tissue parameter maps with the goal to synthesize physically plausible FLASH signals. Our approach is additionally trained for random acquisition parameters and generalizes across different acquisition protocols and provides improved performance over fixed acquisition based training methods. We also demonstrate the robustness of our approach by performing these estimation with as low as three input contrast images.

Yuuzo Kamiguchi¹, Sadanori Tomiha², and Masao Yui^3
1Advanced Technology Reserch Dept. Reserch and Development Center, Canon Medical Systems Corporation, Kawasaki, Japan, ²Advanced Technology Reserch Dept. Reserch and Development Center, Canon Medical Systems Corporation, Otawara, Japan, ³Reserch and Development Center, Canon Medical Systems Corporation, Otawara, Japan

We examined the quantitative multi parameter mapping method so called transient state DESPOT (tsDESPOT) which based on conventional DESPOT sequence. From the acquired data, low rank approximated (LRA) images which include transient state information were reconstructed, then T1, T2, B1 and PD maps were estimated by dense neural network. In this study, we proposed fast estimation method of accurate full sampled LRA images using approximate ADMM (alternating direction method of multipliers) which optimize Unet estimation and data consistency. Compared to simple Unet estimation method, the method improved quantitative accuracy of maps and removed artifact that couldn’t be removed.
 

Yuze Li¹, Huijun Chen¹, Haikun Qi², Zhangxuan Hu³, Zhensen Chen¹, Runyu Yang¹, Huiyu Qiao¹, Jie Sun⁴, Tao Wang⁵, Xihai Zhao¹, Hua Guo¹, and Huijun Chen^1
1Center for Biomedical Imaging Research, Medical School, Tsinghua University, Beijing, China, ²School of Biomedical Engineering and Imaging Sciences, King’s College London, London, United Kingdom, ³GE Healthcare, Beijing, China, ⁴Vascular Imaging Lab and BioMolecular Imaging Center, Department of Radiology, University of Washington, Seattle, Seattle, WA, United States, ⁵Department of Neurology, Peking University Third Hospital, Beijing, China

A Deep learning enhAnced T1 parameter mappIng and recoNstruction framework using spatial-Temporal and phYsical constraint (DAINTY) was proposed. DAINTY explicitly imposed low rank and sparsity constraints on the multi-frame T1 weighted images to exploit the spatial-temporal correlation. A deep neural network was used to efficiently perform T1 mapping as well as denoise and reduce under-sampling artifacts. More importantly, smooth and accurate T1 maps generated from the neural network were transformed to T1 weighted images using the physical model, which the transformed T1 weighted images were also refined. Combining refined images and intermediate reconstructed images, the image quality was greatly improved. 

Kuo-Wei Lai^1,2, Jeremias Sulam¹, Manisha Aggarwal³, Peter van Zijl^2,3, and Xu Li^2,3
1Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD, United States, ²F.M. Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Institute, Baltimore, MD, United States, ³Department of Radiology and Radiological Sciences, Johns Hopkins University, Baltimore, MD, United States

We extended a Learned Proximal Convolutional-Neural-Network (LP-CNN) model used in scalar-based Quantitative Susceptibility Mapping (QSM) to tensor-based Susceptibility Tensor Imaging (STI). To improve the accuracy in reconstructed susceptibility anisotropy and tensor eigenvectors, we devised a decomposition loss function to balance training errors in isotropic and anisotropic components. Results using a synthetic dataset demonstrated that, compared to the conventional iterative approach, LP-CNN-STI provides better estimates of susceptibility tensor and smaller errors in anisotropy and eigenvectors. This deep learning-based STI method naturally incorporates the STI physical model, and is a first step toward development of learning-based STI potentially with less acquisition orientations.

Eric Z. Chen¹, Yongquan Ye², Xiao Chen¹, Jingyuan Lyu², Zhongqi Zhang³, Yichen Hu², Terrence Chen¹, Jian Xu², and Shanhui Sun^1
1United Imaging Intelligence, Cambridge, MA, United States, ²UIH America, Inc., Houston, TX, United States, ³United Imaging Healthcare, Shanghai, China

We propose a deep learning framework for undersampled 3D MRI data reconstruction and apply it to a recently developed multi-flip-angle (FA) and multi-echo GRE method (named MULTIPLEX) that can simultaneously acquire multiple parametric images with just one single MRI scan. The proposed deep learning method shows good performance in image quality and reconstruction time.

Johnathan Le^1,2, Jason Mendes², Mark Ibrahim³, Brent Wilson³, Edward DiBella^1,2, and Ganesh Adluru^1,2
1Department of Biomedical Engineering, University of Utah, Salt Lake City, UT, United States, ²Utah Center for Advanced Imaging Research, University of Utah, Salt Lake City, UT, United States, ³Department of Cardiology, University of Utah, Salt Lake City, UT, United States

Cardiac T1 mapping has been shown to be a promising method for assessing different cardiomyopathies. Recently, native T1 mapping has been used to identify ischemic regions in coronary artery disease without the use of gadolinium-based contrast agents. Most cardiac T1 mapping methods require long breath holds during the acquisition sequence which can be difficult for patients particularly during exercise or pharmacologically induced stress. Here we proposed using attention-gated neural networks to reduce the acquisition time of native and post-contrast cardiac T1 mapping sequences without significant loss of quality. 

Aniket Tolpadi^1,2, Francesco Caliva¹, Misung Han¹, Valentina Pedoia¹, and Sharmila Majumdar^1
1Radiology and Biomedical Imaging, UCSF, San Francisco, CA, United States, ²Bioengineering, University of California, Berkeley, Berkeley, CA, United States

Degeneration of intervertebral discs (IVDs) is correlated with low back pain, but conventional MRI fails to capture early signs of degeneration. Quantitative MRI (qMRI) is sensitive to early degenerative biochemical changes but suffers from long acquisition times. We present a recurrent encoder-decoder architecture that predicts fully sampled IVD T2 maps from spatially and temporally undersampled qMRI echos. The network allows for up to eightfold reduction in acquisition time while exhibiting strong correlation to ground truth maps, maintaining fidelity to T2 values, and retaining textures. With further development, this network can make qMRI a more regular part of lumbar spine imaging.

Ingo Hermann^1,2,3, Alena-Kathrin Golla^1,3, Eloy Martinez-Heras⁴, Ralf Schmidt¹, Elisabeth Solana⁴, Sara Llufriu⁴, Achim Gass⁵, Lothar Schad¹, Sebastian Weingärtner², and Frank Zöllner^1,3
1Computer Assisted Clinical Medicine, Medical Faculty Mannheim, University Heidelberg, Mannheim, Germany, ²Magnetic Resonance Systems Lab, Department of Imaging Physics, Delft University of Technology, Delft, Netherlands, ³Mannheim Institute for Intelligent Systems in Medicine, Medical Faculty Mannheim, University Heidelberg, Mannheim, Germany, ⁴Center of Neuroimmunology, Laboratory of Advanced Imaging in Neuroimmunological Diseases, Hospital Clinic Barcelona, Universitat de Barcelona, Barcelona, Spain, ⁵Department of Neurology, Medical Faculty Mannheim, University Heidelberg, Mannheim, Germany

A single deep learning regression network was trained for unified reconstruction, distortion correction and denoising of T₁, T₂*, WM, GM and lesion probability maps from MRF-EPI acquisition. The network was trained with binary lesion masks and the WM, GM probability maps generated with SPM. The training T₁ and T₂* maps were reconstructed using dictionary matching. The relative deviation was 7.6% for the 5 output mask in the whole brain between the proposed deep learning network and the conventional processing. Dice coefficients were 0.85 for WM and GM and 0.67 for the lesions with a lesion detection rate of 0.83.

Veronica Ravano^1,2,3, Gian Franco Piredda^1,2,3, Tom Hilbert^1,2,3, Bénédicte Maréchal^1,2,3, Reto Meuli², Jean-Philippe Thiran^2,3, Tobias Kober^1,2,3, and Jonas Richiardi^2
1Advanced Clinical Imaging Technology, Siemens Healthineers, Lausanne, Switzerland, ²Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ³LTS5, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland

Post-acquisition data harmonization promises to unlock multi-site data for deep learning applications. In turn, this rests on measuring image similarity. Here, we investigate the sensitivity of several similarity measures to fourteen acquisition protocols of a 3D T1-weighted (MPRAGE) contrast. Standard similarity metrics, a deep perceptual loss and a segmentation loss are extracted between image pairs and compared. The perceptual loss is highly correlated with L1 distance and outperforms other metrics in detecting acquisition parameter changes. The segmentation loss, however, is poorly correlated with other metrics, suggesting that these image similarity metrics alone aren't sufficient to harmonize data for clinical applications.