Wei Wang¹, Jian Wu¹, Qinqin Yang¹, Jian Han¹, Congbo Cai¹, Shuhui Cai¹, and Zhong Chen^1
1Xiamen University, Xiamen, China

Non-Cartesian spatiotemporally encoded MRI sequence shows obvious advantages in spatial Non-Cartesian spatiotemporally encoded MRI sequence shows obvious advantages in spatial selectivity and sampling efficiency. However, the resulting image is susceptible to severe distortion due to the cumulative effect of the B0 field inhomogeneity. In this work, this issue is addressed by introducing a deep learning-based method that is specifically tailored to inhomogeneous field correction. Simulation and in vivo rat brain experiments show that our method can effectively correct the image distortion and obtain the field map without extra reference scan.

Xinwen Liu¹, Jing Wang^1,2, Fangfang Tang¹, Hongfu Sun¹, Feng Liu¹, and Stuart Crozier^1
1School of Information Technology and Electrical Engineering, the University of Queensland, Brisbane, Australia, ²School of Information and Communication Technology, Griffith University, Brisbane, Australia

In MRI, region-of-interest (ROI) imaging is frequently used in clinical applications. The sub-sampling-based scheme is capable of accelerating the ROI-focused image reconstruction process but degrades the image quality. The degradation could be alleviated by ROI-weighted optimization; however, existing methods mainly focus on the local signal restoration and have no explicit control of the noise from the entire image. In this abstract, we propose to reconstruct the ROI using a non-local U-net method that incorporates contextual information from the whole image. The results show the proposed algorithm improves PSNR and SSIM over conventional methods.

Xiangshun Liu^1,2, Minghui Zhang¹, Qiegen Liu¹, Leslie Ying³, Xin Liu², Hairong Zheng², and Shanshan Wang^2
1Department of Electronic Information Engineering, Nanchang University, Nanchang, China, ²Paul C. Lauterbur Research Center for Biomedical Imaging, SIAT, Chinese Academy of Sciences, Shenzhen, China, ³Department of Biomedical Engineering and Electrical Engineering, The State University of New York, Buffalo, New York, NY, United States

This paper proposes an enhanced denoising autoencoder prior (EDAEP) network learning method for multi-contrast MR reconstruction using deep learning. Specifically, a multi-model autoencoder with various noise levels was developed to capture different features from multi-contrast images. A weighted aggregation strategy was also adopted to balance the impact of multiple model outputs. These designs empower the network to explore the correlations and similarities among multi-contrast images, handle different acceleration trajectories and avoid a lot of cumbersome retraining. Experimental results demonstrate that our method can improve the quality of reconstructed images compared to other classical methods. 

Ali Pour Yazdanpanah^1,2, Onur Afacan^1,2, and Simon K. Warfield^1,2
1Harvard Medical School, Boston, MA, United States, ²Boston Children's Hospital, Boston, MA, United States

Fast data acquisition in MRI is vastly in demand and scan time depends on the number of acquired k-space samples. The data-driven methods based on deep networks have resulted in promising improvements, compared to the conventional methods. The connection between deep network and Ordinary Differential Equation (ODE) has been studied recently. Here, we propose an ODE-based deep network for MRI reconstruction to enable the rapid acquisition of MR images with improved image quality. Our results with undersampled data demonstrate that our method can deliver higher quality images in comparison to the reconstruction methods based on the UNet and Residual networks.

Bo Zhu^1,2,3, Neha Koonjoo^1,2,3, Bragi Sveinsson^1,2,3, and Matthew S Rosen^1,2,3
1Department of Radiology, A.A Martinos Center for Biomedical Imaging/MGH, Charlestown, MA, United States, ²Harvard Medical School, Boston, MA, United States, ³Department of Physics, Harvard University, Cambridge, MA, United States

Here we propose to quantitatively measure the data complexity of training datasets using the Rademacher Complexity metric, and demonstrate its effectiveness in analyzing dataset composition and its effect on neural network training for image reconstruction tasks.

Matthew Muckley*¹, Tullie Murrell*², Suvrat Booshan², Hersh Chandarana¹, Florian Knoll¹, and Daniel K. Sodickson^1
1Radiology, NYU School of Medicine, New York, NY, United States, ²Facebook AI Research, Menlo Park, CA, United States

The introduction of machine learning for medical image reconstruction has opened up new opportunities for reconstruction speed and subsampling; however, acquiring ground truth data is expensive or impossible in the case of dynamic imaging. Here we investigate a technique for optimizing a CNN on continuous radial data by treating the NUFFT-CNN function as an autoencoding deep image prior. Using this method, we are able to reconstruct images that increment over time frames as short as a single spoke. The technique opens up new possibilities for dynamic image reconstruction.

Jing Cheng¹, Yiling Liu¹, Qiegen Liu², Ziwen Ke¹, Haifeng Wang¹, Yanjie Zhu¹, Leslie Ying³, Xin Liu¹, Hairong Zheng¹, and Dong Liang^1
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, ²Nanchang University, Nanchang, China, ³University at Buffalo, The State University of New York, Buffalo, Buffalo, NY, United States

We introduce an effective strategy to maximize the potential of deep learning and model-based reconstruction based on the network of ISTA-net, which is the unrolled version of iterative shrinkage-thresholding algorithm for compressed sensing reconstruction. By relaxing the constraints in the reconstruction model and the algorithm, the reconstruction quality is expected to be better. The prior of the to-be-reconstructed image is obtained by the trained networks and the data consistency is also maintained through updating in k-space for the reconstruction. Brain data shows the effectiveness of the proposed strategy.

Sizhuo Liu¹, Ning Jin², Philip Schniter³, and Rizwan Ahmad^1
1Biomedical Engineering, Ohio State University, Columbus, OH, United States, ²Siemens Medical Solutions Inc, Columbus, OH, United States, ³Electrical and Computer Engineering, Ohio State University, Columbus, OH, United States

The recently proposed Plug-and-Play (PnP) methods provide an avenue to combine physics-driven MR models with sophisticated, learned image models instantiated by image denoising subroutines. The performance of PnP methods, however, is sensitive to changes in the measurement signal-to-noise ratio (SNR) and algorithmic parameters that balance the contributions from the data fidelity and denoising terms.  We propose a discrepancy-principle-based scheme that mediates the impact of the denoising subroutine, leading to more consistent performance across different measurement SNRs without manual intervention. For validation, the proposed scheme is applied to cine images collected at 3T, 1.5T, and 0.35T scanners.

Shohei OUCHI¹ and Satoshi ITO^1
1Utsunomiya University, Utsunomiya, Japan

A novel single image domain learning CNN based reconstruction method for phase-varied images is proposed in which real and imaginary part of complex image are reconstructed independently. Proposed method uses symmetrical sub-sampling which enable reconstruction for real and imaginary part of complex images independently of each other without estimating phase distribution on the image. Reconstruction experiments showed that higher PSNR images are obtained in proposed method compared to phase estimating CNN or ADMM-CSNet. Proposed method is highly practical since it is robust to phase variation and is easy for training because of its simple CNN structure

Sunghun Seo¹, Seung Hong Choi², and Sung-Hong Park^1
1Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology, Daejeon, Korea, Republic of, ²Department of Radiology, Seoul National University College of Medicine, Seoul, Korea, Republic of

It is known that optimizing a deep learning model based on best validation loss achieves best quantitative results in image reconstruction, but resulting images are often blurry. In this study we propose an alternative way of optimization in which convolutional neural network (CNN) is trained beyond best validation loss to produce realistic MR images by monitoring Fréchet Inception Distance. The new approach generated sharper and more realistic images than the conventional optimization, providing a new insight into optimization for MR image reconstruction.

Chengyan Wang¹, Yan Li², Jun Lv³, Bo Li⁴, Fei Dai⁵, Weibo Chen⁶, and He Wang^1,5
1Human Phenome Institute, Fudan University, Shanghai, China, ²Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, ³Department of Computer Science, Yantai University, Yantai, China, ⁴The Third Affiliated Hospital of Nanchang University, Nanchang, China, ⁵Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, ⁶Philips Healthcare, Shanghai, China

Conventional SENSE requires accurate estimation of coil sensitivity maps, which remains to be a challenge in practical scenarios. This study aims to apply CNN to extract coil sensitivity information from the undersampled center k-space, and use the estimated sensitivity maps for parallel imaging. Results show that no obvious residual signal can be seen in the reconstructed images for all cases, which indicates the efficacy of the proposed method. Besides, the CNN based SENSE image without ACS appears to be less noisy than conventional SENSE results with ACS, which may benefit from the denoising effect from CNN on the sensitivity maps.

Jun Lyu¹, Peng Wang¹, and Chengyan Wang^2
1Yantai University, Yantai, China, ²Fudan University, Shanghai, China

This study aims to use GAN architecture to remove the g-factor artifacts in SENSE reconstruction. The proposed method outperforms SENSE and ZF+GAN in terms of the measured quality metrics (decreases of NMSE and increases of PSNR and SSIM). Besides, our method performs well in preserving images details with under-sampling factor of up to 6-fold, which is promising to be applied in clinical applications.

Yudu Li^1,2, Yue Guan³, Ziyu Meng^2,3, Fanyang Yu^2,4, Rong Guo^1,2, Yibo Zhao^1,2, Tianyao Wang⁵, Yao Li³, and Zhi-Pei Liang^1,2
1Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign, Urbana, IL, United States, ²Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, Urbana, IL, United States, ³Institute for Medical Imaging Technology (IMIT), School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China, ⁴Department of Bioengineering, University of Illinois at Urbana-Champaign, Urbana, IL, United States, ⁵Department of Radiology, The Fifth People's Hospital of Shanghai, Shanghai, China

Machine learning (ML) based MR image reconstruction leverages the great power and flexibility of deep networks in representing complex image priors. However, ML image priors are often inaccurate due to limited training data and high dimensionality of image functions. Therefore, direct use of ML-based reconstructions or treating them as statistical priors can introduce significant biases. To address this limitation, we treat ML-based reconstruction as an initial estimate and use an information theoretical framework to incorporate it into the final reconstruction, which is optimized to capture novel image features. The proposed method may provide an effective framework for ML-based image reconstruction.

Yihang Zhou¹, Hongyu Li², Jing Yuan¹, Leslie Ying², Kin Yin Cheung¹, and Siu Ki Yu^1
1Medical Physics & Research Department, Hong Kong Sanatorium & Hospital, Hong Kong, China, ²Department of Biomedical Engineering, Department of Electrical Engineering, The State University of New York at Buffalo, Buffalo, NY, United States

MR-guided radiotherapy (MRgRT) is creating new perspectives towards an individualized precise radiation therapy solution. However, spatial resolution of fractional MRI can be much restricted, in order to shorten scan time, by patient tolerance of immobilization, intra-fractional anatomical motion and complicated MRgRT workflow. We hypothesized that the quality of low-resolution daily MRI could be greatly restored to generate super-resolution MRI, whose quality should be comparable of high-resolution planning MRI, by applying deep learning techniques. In this study, we aimed to investigate the feasibility of deep learning super-resolution MRI generation in the head-and-neck for adaptive MRgRT purpose.

Lea Bogensperger^1,2, Oliver Maier¹, and Rudolf Stollberger^1,3
1Institute of Medical Engineering, Graz University of Technology, Graz, Austria, ²Institute for Computer Graphics and Vision, Graz University of Technology, Graz, Austria, ³Biotechmed, Graz, Austria

A deep learning approach is proposed to estimate M0 and T1 maps from undersampled variable flip angle (VFA) data to explore the potential of this method for acceleration and rapid reconstruction even without parallel imaging. A U-Net was implemented with a model consistency term containing the signal equation to ensure the physical validity and to include prior knowledge of B₁₊. Training is performed on numerical brain phantoms and by means of transfer-learning on retrospectively undersampled in-vivo data. Qualitative and quantitative results show the acceleration potential for both numerical and in-vivo data for acceleration factors R=1.89, 3.43, and 5.84.

Yiling Liu¹, Jing Cheng², Yanjie Zhu², Haifeng Wang², Ziwen Ke¹, Qiegen Liu³, Xin Liu², Hairong Zheng², Leslie Ying⁴, and Dong Liang^1,2
1Research center for Medical AI, Shenzhen Institutes of Advanced Technology, shenzhen, China, ²Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, shenzhen, China, ³Department of Electronic Information Engineering, Nanchang University, Nanchang, China, ⁴Departments of Biomedical Engineering and Electrical Engineering, University at Buffalo,the State University of New York, Buffalo, NY, United States

ADMM is a popular algorithm for Compressed sensing (CS) MRI. ADMM-based deep networks have also achieved a great success by unrolling the ADMM algorithm into deep neural networks. Nevertheless, ADMM-Nets only make the components in the regularization term learnable. In this work, we propose a relaxed version of ADMM-Net (i.e. Relax-ADMM-Net) to further improve its performance for fast MRI, where the additional data consistency term and variable combinations in the updating rules are all freely learned by the network. Experiments reveal the effectiveness of the proposed network compared with several competing model-driven networks.

Chen Qin¹, Jo Schlemper^1,2, Kerstin Hammernik¹, Jinming Duan³, Ronald M Summers⁴, and Daniel Rueckert^1
1Imperial College London, London, United Kingdom, ²Hyperfine Research Inc., Guilford, CT, United States, ³School of Computer Science, University of Birmingham, Birmingham, United Kingdom, ⁴NIH Clinical Center, Bethesda, MD, United States

We present a deep network interpolation strategy for accelerated parallel MR image reconstruction. In particular, we examine the network interpolation in parameter space between a source model that is formulated in an unrolled scheme with L1 and SSIM losses and its counterpart that is trained with an adversarial loss. We show that by interpolating between the two different models of the same network structure, the new interpolated network can model a trade-off between perceptual quality and fidelity.

Peizhou Huang¹, Chaoyi Zhang², Hongyu Li², Sunil Kumar Gaire², Ruiying Liu², Xiaoliang Zhang¹, Xiaojuan Li³, Liang Dong⁴, and Leslie Ying^1,2
1Biomedical Engineering, State University of New York at Buffalo, Buffalo, NY, United States, ²Electrical Engineering, State University of New York at Buffalo, Buffalo, NY, United States, ³Program of Advanced Musculoskeletal Imaging (PAMI), Cleveland Clinic, Cleveland, OH, United States, ⁴Paul C. Lauterbur Research Center for Biomedical Imaging, SIAT, CAS, Shenzhen, China

Deep learning has been applied to MRI image reconstruction successfully. Most existing works require labeled ground-truth images to learn network parameters for image reconstruction, which is not practical in some MR applications where acquisition of fully sampled images takes too long. In this abstract, we propose a novel unsupervised deep neural network for reconstruction from undersampled data. The proposed network, named URED-net, is built upon conventional ADMM algorithm for compressed sensing reconstruction, but incorporating noise2noise, an unsupervised deep denoising network. The experimental results demonstrate proposed URED-net is superior to the standard noise2noise network with and without ground-truth images for training.

Zhiyang Fu¹, Maria I Altbach², Diego R Martin², and Ali Bilgin^1,2,3
1Electrical and Computer Engineering, University of Arizona, Tucson, AZ, United States, ²Department of Medical Imaging, University of Arizona, Tucson, AZ, United States, ³Biomedical Engineering, University of Arizona, Tucson, AZ, United States

MR images are often reconstructed first and then used for medical image analysis tasks such as segmentation or classification. This sequential procedure can compromise the performance of the image analysis task. In this work, we propose a multi-task learning framework that jointly reconstructs underlying images and detects multiple sclerosis lesions. This framework outperforms the conventional sequential processing pipeline. We also introduce a multi-objective optimization as an effective and automated approach to balance the trade-off among multi-task losses. Experimental results suggest that taking into account subsequent detection tasks during image reconstruction may lead to enhanced detection performance.

Xinwen Liu¹, Jing Wang^1,2, Fangfang Tang¹, Shekhar S. Chandra¹, Feng Liu¹, and Stuart Crozier^1
1School of Information Technology and Electrical Engineering, the University of Queensland, Brisbane, Australia, ²School of Information and Communication Technology, Griffith University, Brisbane, Australia

MRI images of the same subject in different contrasts contain shared information, such as the anatomical structure.  Utilizing the redundant information amongst the contrasts to sub-sample and faithfully reconstruct multi-contrast images could greatly accelerate the imaging speed, improve image quality and shorten scanning protocols. We propose an algorithm that generates the optimized sampling pattern and reconstruction scheme of one contrast (e.g. T2-weighted image) when images with different contrast (e.g. T1-weighted image) have been acquired. The proposed algorithm achieves increased PSNR and SSIM with the resulting optimal sampling pattern compared to other acquisition patterns and single contrast methods.

Hangfei Liu¹, Jingjing Li¹, Qing Tang¹, and Tao Zhang^1,2,3
1School of Life Science and Technology, University of Electronic Science and Technology of China, chengdu, China, ²High Field Magnetic Resonance Brain Imaging Laboratory of Sichuan, Chengdu, China, ³Key Laboratory for Neuro Information, Ministry of Education, Chengdu, China

Recently parallel imaging reconstruction based on deep learning has made lots of progresses, however, there still exist several common challenges, i.e. generalization, transferability and robustness. On the contrary, SENSE reconstruction has been routinely used in clinical scans due to its high robustness and excellent image quality. A high-quality coil sensitivity map (HQCSM) is the key to achieve good SENSE reconstruction. We proposed a hybrid SENSE reconstruction frame, combining the SENSE reconstruction algorithm with a deep convolutional neural network to learn HQCSM from a few automatic calibration lines (ACS), which shows good generalization for different under-sampling ratio and enhanced robustness.

Maarten L Terpstra^1,2, Federico d'Agata^1,2,3, Bjorn Stemkens^1,2, Jan JW Lagendijk¹, Cornelis AT van den Berg^1,2, and Rob HN Tijssen^1,2
1Department of Radiotherapy, Division of Imaging & Oncology, University Medical Center Utrecht, Utrecht, Netherlands, ²Computational Imaging Group for MR diagnostics & therapy, Center for Image Sciences, University Medical Center Utrecht, Utrecht, Netherlands, ³Department of Neurosciences, University of Turin, Turin, Italy

Recently, dAUTOMAP has been presented to perform deep learning-based image reconstruction. dAUTOMAP uses gridded k-space points and so far it has only been used to reconstruct Cartesian acquisitions. In this work, we demonstrate that dAUTOMAP can produce high-quality reconstructions on radial and spiral non-Cartesian acquisitions and can resolve artifacts beyond those introduced by the undersampled acquisition.

Mingdong Fan¹, Brendan Eck², Nicole Seiberlich³, Michael Martens¹, and Robert Brown^1
1Physics, Case Western Reserve University, Cleveland, OH, United States, ²Cardiovascular and Metabolic Sciences, Cleveland Clinic, Cleveland, OH, United States, ³Radiology, University of Michigan, Ann Arbor, Ann Arbor, MI, United States

There are two major challenges in MRF reconstruction, the aliasing artifacts that results from the largely under-sampled k-space, and the very long MRF sequence used in practice to improve the reconstruction accuracy. In this study, we propose an end-to-end deep learning based reconstruction model that aims to address the issue of the spatial aliasing artifacts and provide accurate reconstruction with ultra-short MRF signals.

Hidenori Takeshima¹ and Hideaki Kutsuna^2
1Advanced Technology Research Department, Research and Development Center, Canon Medical Systems Corporation, Kanagawa, Japan, ²MRI Systems Development Department, MRI Systems Division, Canon Medical Systems Corporation, Kanagawa, Japan

The authors propose a new model using a convolutional neural network (CNN) named k-t CNN for approximating non-linear spatio-temporal mappings used in various applications. Existing studies imply that spatio-temporal mappings are non-linear. Most existing studies developed various methods using linear models for spatio-temporal applications. Meanwhile, the effectiveness of non-linear models was shown for spatial-domain applications.
As an application of k-t CNN, the effectiveness of the proposed method is shown experimentally in the case of reconstruction of stack-of-spirals k-space data.
 

Shigeru Kiryu¹, Yasutaka Sugano², Tomoyuki Ohta³, and Kuni Ohtomo^4
1Radiology, International University of Health, School of medicine, Chiba, Japan, ²Canon Medical Systems Corporation, Kanagawa, Japan, ³International University of Health and Welfare Hospital, Tochigi, Japan, ⁴International University of Health and Welfare, Tochigi, Japan

We assessed the performance of the Deep Learning-based Reconstruction (dDLR) technique in improving 1.5T MR images. Eleven volunteers underwent MR imaging at 3T and 1.5T on the same day with the same imaging parameters. We applied the dDLR to the 1.5T image data (dDLR-1.5T), and then compared the 1.5T and dDLR-1.5T datasets with reference to the 3T dataset. The structure similarity of dDLR-1.5T was higher than that of 1.5T and dDLR increased SNR at 1.5T. The dDLR technique improves the image quality of MR images obtained at 1.5T.

Wei Yu^1,2, Lixin Wang^1,2, Xianchang Zhang³, Zhentao Zuo⁴, Rong Xue^4,5,6, and Tianyi Qian^1,2
1Sinovation Ventures AI Institute, Beijing, China, ²QuantMind, Beijing, China, ³MR Collaboration, Siemens Healthcare, Beijing, China, ⁴State Key Laboratory of Brain and Cognitive Science, Beijing MRI Center for Brain Research, Institute of Biophysics, Chinese Academy of Sciences, Beijing, China, ⁵University of Chinese Academy of Sciences, Beijing, China, ⁶Beijing Institute for Brain Disorders, Beijing, China

This study investigates a reconstruction method designed to generate 7T magnetic resonance images (MRI) from 3T MRI based on a generative adversarial network. Compared with current reconstruction methods, this method can simultaneously reconstruct well-defined anatomical details and salient blood vessels. This reconstruction method may be useful for increasing the efficiency of brain MRI examinations.

Ruiyang Zhao¹, Tao Wang², Kang Yan¹, Chengcheng Zhang³, Zhipei Liang⁴, Yiping Du¹, Dianyou Li³, Bomin Sun³, and Yuan Feng^1
1Institute for Medical Imaging Technology, Shanghai Jiao Tong University, Shanghai, China, ²Functional Neurosurgery,Ruijin Hospital affiliated to Shanghai Jiao Tong University, Shanghai, China, ³Shanghai Jiao Tong University Medical School Affiliated Ruijin Hospital, Shanghai, CHINA, Shanghai, China, ⁴Beckman Institute for Advanced Science & Technology, Department of Electrical & Computer Engineering，University of Illinois at Urbana-Champaign, Urbana-Champaign, IL, United States

Real –time MR image-guided neurosurgery could greatly improve the surgery accuracy and outcome. However, real-time guidance requires highly accelerated imaging. In this study, we proposed a Convolutional Long Short-term Memory (Conv-LSTM) based U-net to reconstruct consecutive image frames with golden-angle sampling. The Conv-LSTM based architecture was developed to explore time coherence information. Training and test datasets were generated from MR images of patients treated with Deep Brain Stimulation (DBS). Results showed that our model could achieve an acceleration rate ~80x, which provided great potentials for application in MR-guided interventional therapy.

Peibei Cao^1,2, Linfang Xiao^1,2, Yilong Liu^1,2, Yujiao Zhao^1,2, Yanqiu Feng³, Alex T Leong^1,2, and Ed X Wu^1,2
1Laboratory of Biomedical Imaging and Signal Processing, The University of Hong Kong, Hong Kong, China, ²Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong, China, ³School of Biomedical Engineering, Southern Medical University, Guangzhou, China

Convolutional neural network (CNN) has emerged as a powerful tool for medical image reconstruction. In this study, we designed and implemented a CNN model for partial Fourier MRI reconstruction, and compared its performance with the existing projection onto convex sets (POCS) method. The results demonstrated that our proposed deep learning approach could effectively recovered the high frequency components and outperformed the POCS method especially when partial Fourier fraction is close to 50%.

Yan Wu¹, Yajun Ma², Jiang Du², and Lei Xing^1
1Radiation Oncology, Stanford University, Stanford, CA, United States, ²Radiology, University of California San Diego, La Jolla, CA, United States

The application of quantitative MRI is limited by additional data acquisition for variable contrast images. Leveraging from the unique ability of deep learning, we propose a data-driven strategy to derive quantitative T1 map and proton density map from a single qualitative MR image without specific requirements on the weighting of the input image. The quantitative parametric mapping tasks are accomplished using self-attention deep convolutional neural networks, which make efficient use of local and non-local information. In this way, qualitative and quantitative MRI can be attained simultaneously without changing the existing imaging protocol.

Rita G. Nunes¹, Santiago Sanz-Estébanez², Joseph V. Hajnal³, Lucilio Cordero-Grande³, and Carlos Alberola-López^2
1ISR-Lisbon/LARSyS and Department of Bioengineering, Instituto Superior Técnico – University of Lisbon, Lisbon, Portugal, Lisbon, Portugal, ²Laboratorio de Procesado de Imagen, Universidad de Valladolid, Valladolid, Spain, Valladolid, Spain, ³Centre for the Developing Brain and Department of Biomedical Engineering, School of Biomedical Engineering and Imaging Sciences, King’s College London,London U.K, London, United Kingdom

Incorporation of machine learning (ML) approaches in MR reconstruction is currently a hot topic because of the enormous potential that deep learning solutions have shown in vision and imaging communities. Recently, a procedure known as NAMER has been proposed; this procedure incorporates a ML module into an iterative reconstruction for multishot acquisitions with inter-shot motion estimation and correction (referred to as aligned reconstruction). In this abstract we provide some insight on the benefits and limitations associated to NAMER by analyzing its behavior both with a steady and a discontinued use of the ML artifact cleaning step.

Ziwen Ke^1,2, Yanjie Zhu³, Jing Cheng^2,3, Leslie Ying⁴, Xin Liu³, Hairong Zheng³, and Dong Liang^1,3
1Research Center for Medical AI, Shenzhen Institutes of Advanced Technology，Chinese Academy of Sciences, Shenzhen, China, ²Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen, China, ³Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology，Chinese Academy of Sciences, Shenzhen, China, ⁴Department of Biomedical Engineering and Department of Electrical Engineering, The State University of New York, Buffalo, NY, United States

In our previous work, we proposed an unsupervised deep learning method for parallel MR cardiac imaging via time interleaved sampling. The comparisons with classical methods on in vivo data have shown that this method can achieve improved reconstruction results. However, the proposed unsupervised framework is based on the time interleaved sampling scheme. Does the model trained with time interleaved undersampling pattern have good generalization to other sampling patterns? In this paper, we will explore the generalization performance of the learned unsupervised deep learning method under different sampling patterns.

Gaspar Delso¹, Suryanarayanan Kaushik², Graeme C McKinnon², Daniel Lorenzatti³, Julián Vega³, Teresa M. Caralt³, Adelina Doltra³, José T. Ortiz-Pérez³, Rosario J. Perea³, Susanna Prat³, Marta Sitges³, and Martin A. Janich^4
1ASL MR, GE Healthcare, Barcelona, Spain, ²GE Healthcare, Waukesha, WI, United States, ³Hospital Clínic de Barcelona, Barcelona, Spain, ⁴GE Healthcare, Munich, Germany

We present the evaluation results of a novel Deep Learning reconstruction framework, applied to clinical Delayed Myocardial Enhancement datasets acquired with a 3D inversion-recovery T1-weighted gradient echo sequence.

Samuel W Fielden^1,2, Eric D Carruth¹, Christopher D Nevius¹, Christopher M Haggerty¹, and Brandon K Fornwalt^1,3
1Imaging Sciences & Innovation, Geisinger, Danville, PA, United States, ²Medical & Health Physics, Geisinger, Danville, PA, United States, ³Radiology, Geisinger, Danville, PA, United States

Displacement Encoding with Stimulated Echoes (DENSE) is a powerful technique that has found great utility in accurately measuring cardiac tissue displacement. However, DENSE remains time-consuming to acquire, particularly for 3-dimensionally encoded or higher resolution schemes, and so methods to accelerate image acquisition are needed. Here, we apply the Deep Cascade of Convolutional Neural Networks (DCCNN) to the complex-valued, non-Cartesian data of DENSE to show that accelerated imaging via k-space undersampling is feasible using a deep learning-based reconstruction.

Mohammad Abdishektaei¹, Xue Feng¹, Craig H Meyer¹, and Frederick H Epstein^1
1Biomedical Engineering, University of Virginia, Charlottesville, VA, United States

Cine displacement encoding with stimulated echoes (DENSE) is an accurate and reproducible method of strain imaging. The stimulated echo (STE), which carries the tissue displacement information in it’s phase, is simultaneously acquired with two artifact-generating echoes. A combination of phase-cycled acquisitions and through-plane dephasing are typically used to suppress the artifact-generating echoes. The limitations of these methods are longer acquisition times, susceptibility to breathing motion and the loss of signal-to-noise ratio due to intravoxel dephasing. To potentially overcome these limitations, the use of a deep convolutional neural network to suppress the undesired echoes from a single acquisition was investigated.

Jing Cheng¹, Ziwen Ke¹, Haifeng Wang¹, Yanjie Zhu¹, Leslie Ying², Xin Liu¹, Hairong Zheng¹, and Dong Liang^1
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, ²University at Buffalo, The State University of New York, Buffalo, Buffalo, NY, United States

Most deep learning methods for MR reconstruction heavily rely on the large number of training data pairs to achieve best performance. In this work, we introduce a simple but effective strategy to handle the situation where collecting lots of fully sampled rawdata is impractical. By defining a CS-based loss function, the deep networks can be trained without ground-truth images or full sampled data. In such an unsupervised way, the MR image can be reconstructed through the forward process of deep networks. This approach was evaluated on in vivo MR datasets and achieved superior performance than the conventional CS method.

Jinwei Zhang^1,2, Hang Zhang^2,3, Cagla Deniz Bahadir⁴, Alan Wang³, Mert Rory Sabuncu^1,2,3, Pascal Spincemaille², Thanh D. Nguyen², and Yi Wang^1,2
1Biomedical Engineering, Cornell University, Ithaca, NY, United States, ²Weill Cornell Medicine, New York, NY, United States, ³Electrical and Computer Engineering, Cornell University, Ithaca, NY, United States, ⁴Siemens Healthineers, Princeton, NJ, United States

In this work, we propose LOUPE-ST, which extends the previously introduced optimal k-space sampling pattern learning framework called LOUPE by employing a straight-through estimator to better handle the gradient back-propagation in the binary sampling layer and incorporating an unrolled optimization network (MoDL) to reconstruct T2w images from under-sampled k-space data with high fidelity.  Our results indicate that, compared with the variable density under-sampling pattern at the same under-sampling ratio (10%), superior reconstruction performance can be achieved with LOUPE-ST optimized under-sampling pattern. This was observed for all reconstruction methods that we experimented with.

Zechen Zhou¹, Christophe Schülke², Chun Yuan³, and Peter Börnert^2
1Philips Research North America, Cambridge, MA, United States, ²Philips Research Hamburg, Hamburg, Germany, ³Department of Radiology, University of Washington, Seattle, WA, United States

Recently, the convolutional neural network (CNN) based reconstruction concept has emerged as a promising implementation of compressed sensing tailored for specific fast imaging applications. The reconstruction performance of such data-driven models may depend on the CNN structure which determines the feature extraction process for sparse representation. In this study, a locally and globally concatenated network is proposed and compared with the residual network as well as the traditional L1-wavelet ESPIRiT. Preliminary experiments on a public knee imaging database showed that the proposed approach provided improved fine structure (e.g. vessel wall) restoration and background noise reduction.

Ruiying Liu¹, Hongyu Li¹, Dong Liang², Xiaojuan Li³, Chaoyi Zhang¹, Peizhou Zhou¹, Leslie Ying¹, and Xiaoliang Zhang^4
1Department of Biomedical Engineering, Department of Electrical Engineering, The State University of New York at Buffalo, Buffalo, NY, United States, ²Paul C. Lauterbur Research Center for Biomedical Imaging, SIAT, CAS,, Shenzhen, China, ³Program of Advanced Musculoskeletal Imaging (PAMI), Cleveland Clinic, Cleveland, OH, United States, ⁴Department of Biomedical Engineering, The State University of New York at Buffalo, Buffalo, NY, United States

Deep learning methods have demonstrated great potential in image reconstruction due to its ability to learn the non-linearity relationship between the undersampled k-space data and the corresponding desired image. Among these methods, Generative Adversarial Networks (GANs) is known to reconstruct images that are sharper and more realistic-looking. In this abstract, we study whether an MR-specific feature map that is trained on a large number of MRI images and used in the loss function can improve the GAN-based reconstruction. We demonstrate that the MR-specific feature map is superior to the pre-trained feature map typically used for GAN-based reconstruction.

Qing Tang¹, Ye Li¹, Hangfei Liu¹, and Tao Zhang^1,2,3
1School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China, ²High Field Magnetic Resonance Brain Imaging Laboratory of Sichuan, Chengdu, China, ³Key Laboratory for Neuro Information, Ministry of Education, Chengdu, China

T2 weighted image (T2WI) usually takes more time and thus is more vulnerable to motion artifacts. With the recent development of applying deep learning to MR imaging, many neural networks are proposed to synthesize high-quality T2 images from under-sampled T2 or other modalities (such as T1). Here we develop a Simple-ResNet network to synthesize high-quality T2 images based on multi-modality information and followed by a k-space correction module. Results show that our model is very easy to train and the synthesized T2 images can achieve comparable image quality as the fully-sampled T2 images.

Wallace Souza Loos¹, Roberto Souza¹, Mariana Bento¹, Robert Marc Lebel^1,2, and Richard Frayne^1
1University of Calgary, Calgary, AB, Canada, ²General Electric Healthcare, Calgary, AB, Canada

Magnetic resonance (MR) imaging still has a high acquisition time due to inherent sequential procedure required to fill k-space. Deep-cascade networks have been used to reconstruct MR images from an under-sampled k-space in order to reduce acquisition time. In this work we investigate a deep-cascade to reconstruct MR images of the brain. We trained the network with 14 different acceleration factors (R). Relevant brain structures were preserved until R = 7x. For R ≥ 8x, MR images presented noticeable blurring artifact. The quality of the segmentation of the brain structures were similar to the reference MR image until R=9.

Hui Xue¹, James C Moon², and Peter Kellman^1
1NHLBI, NIH, Bethesda, MD, United States, ²Barts Heart Centre, Barts Health NHS Trust, London, United Kingdom

In this abstract, we proposed a novel self-calibration AI based MR reconstruction algorithm to utilize the power of a deep neural network. Unlike most deep learning MR reconstruction, this algorithm does not require extra training data and only works on the auto-calibration kspace lines. This algorithm is integrated to run on MR scanner via the Gadgetron InlineAI toolbox. We demonstrated this algorithm on cardiac cine imaging, showing improved image quality without introduced unrealistic anatomical structures.

Taohui Xiao¹, Cheng Li¹, Haoyun Liang¹, Hairong Zheng¹, and Shanshan Wang^1
1Paul C. Lauterbur Research Center for Biomedical Imaging, SIAT, Chinese Academy of Sciences, Shenzhen, P.R.China, Shenzhen, China

  This paper investigates the robustness of deep learning MR reconstruction models for adversarial attacks like new lesions, different anatomy and noise pollutions. Specifically, three popular MR reconstruction algorithms were selected to investigate this issue. Experimental results show that model-based deep learning MR reconstruction method is relatively more robust than end-to-end data-driven reconstruction networks when transfer to other organs or face new lesions. Data-driven approaches can achieve better results when the testing images follow similar distributions as the training images. Severe noise can be a big issue for both deep learning methods and the traditional method.

Kamlesh Pawar¹, Gary Egan¹, and Zhaolin Chen^1
1Monash Biomedical Imaging, Monash University, Melbourne, Australia

Deep learning (DL) methods are superior to the conventional method of accelerated imaging such as parallel imaging and compressed sensing but the integration of DL methods into the MR scanners is still in its infancy. The integration of the DL methods into the MR scanner requires the design of new pulse sequences with a modified sampling patterns/trajectories and development of DL reconstruction framework within the MR scanner. In this work, we present an effective plug-and-play approach of integrating the DL reconstruction into the MR scanner that eliminates the need to modify the existing image acquisition and reconstruction pipeline.

Yue Hu¹, Peng Li¹, and Dong Nan^2
1Harbin Institute of Technology, Harbin, China, ²The First Affiliated Hospital of Harbin Medical University, Harbin, China

We propose a novel algorithm for the super-resolution of brain MR images based on feature regularized DIP network, where no prior training pairs are required. We formulate the network by including the total variation (TV) term as the sparsity regularization and the Laplacian as the sharpness regularization. The network is iteratively updated using the image feature regularizations and the measured image. Numerical experiments demonstrate the improved performance offered by the proposed method.