Dave Van Veen^1,2, Arjun Desai¹, Reinhard Heckel^3,4, and Akshay S. Chaudhari^1
1Stanford University, Stanford, CA, United States, ²University of Texas at Austin, Austin, TX, United States, ³Rice University, Houston, TX, United States, ⁴Technical University of Munich, Munich, Germany

We investigate untrained convolutional neural networks for accelerating both 2D and 3D MRI scans of the knee. Machine learning has demonstrated great potential to accelerate scans while maintaining high quality reconstructions. However, these methods are often trained over a large number of fully-sampled scans, which are difficult to acquire. Here we demonstrate MRI acceleration with untrained networks, achieving similar performance to a trained baseline. Further, we use undersampled k-space measurements as regularization priors to increase the robustness of untrained methods.

Yilmaz Korkmaz^1,2, Salman Ul Hassan Dar^1,2, Mahmut Yurt^1,2, and Tolga Çukur^1,2,3
1Department of Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey, ²National Magnetic Resonance Research Center, Bilkent University, Ankara, Turkey, ³Neuroscience Program, Aysel Sabuncu Brain Research Center, Bilkent University, Ankara, Turkey

A popular framework for reconstruction of undersampled MR acquisitions is deep neural networks (DNNs). DNNs are typically trained in a supervised manner to learn mapping between undersampled and fully sampled acquisitions. However, this approach ideally requires training a separate network for each set of contrast, acceleration rate, and sampling density, which introduces practical burden. To address this limitation, we propose a style generative model that learns MR image priors given fully sampled training dataset of specific contrast. Proposed approach is then able to efficiently recover undersampled acquisitions without any training, irrespective of the image contrast, acceleration rate or undersampling pattern.

Burhaneddin Yaman^1,2, Seyed Amir Hossein Hosseini^1,2, Steen Moeller², Jutta Ellermann², Kâmil Uğurbil², and Mehmet Akçakaya^1,2
1University of Minnesota, Minneapolis, MN, United States, ²Center for Magnetic Resonance Research, Minneapolis, MN, United States

Self-supervised physics-guided deep learning (PG-DL) approaches enable training neural networks without fully-sampled data. These methods split the available k-space measurements into two sets. One is used in the data consistency units of the unrolled network, while the other is used to define the loss. Although self-supervised learning performs well at moderately high acceleration rates, scarcity of acquired data at high acceleration rates degrades the reconstruction performance. In this work, we propose a multi-mask self-supervised learning approach, which retrospectively splits acquired measurement into multiple 2-tuples of disjoint sets. Proposed multi-mask self-supervised learning method outperforms its single-mask counterpart at high acceleration rates.

Jun Lyu¹, Chengyan Wang², and Guang Yang^3,4
1School of Computer and Control Engineering, Yantai University, Yantai, China, ²Human Phenome Institute, Fudan University, Shanghai, China, ³Cardiovascular Research Centre, Royal Brompton Hospital, London, United Kingdom, ⁴National Heart and Lung Institute, Imperial College London, London, United Kingdom

To demonstrate the feasibility of combining parallel imaging (PI) with the generative adversarial network (GAN) for accelerated multi-channel MRI reconstruction. In our proposed PIC-GAN framework, we used a progressive refinement method in both frequency and image domains, which can not only help to stabilize the optimization of the network, but also make full use of the complementarity of the two domains. More specifically, the loss function in the image domain ensures to reduce aliasing artifacts between the reconstructed images and their corresponding ground truth. This enables the model to ensure high-fidelity reconstructions can be obtained even at high acceleration factors. 

Tianming Du¹, Yuemeng Li¹, Honggang Zhang², Stephen Pickup¹, Rong Zhou¹, Hee Kwon Song¹, and Yong Fan^1
1Radiology, University of Pennsylvania, Philadelphia, PA, United States, ²Beijing University of Posts and Telecommunications, Beijing, China

Deep learning in k-space has demonstrated great potential for image reconstruction from undersampled k-space data. However, existing deep learning-based image reconstruction methods typically apply weight-sharing convolutional neural networks to k-space data without taking into consideration the k-space data’s spatial frequency properties, leading to ineffective learning of the image reconstruction models. To improve image reconstruction performance, we develop a residual Encoder-Decoder network architecture with self-attention layers to adaptively focus on k-space data at different spatial frequencies and channels for interpolating the undersampled k-space data. Experimental results demonstrate that our method achieves significantly better image reconstruction performance than current state-of-the-art techniques.

Zi Wang¹, Yihui Huang¹, Zhangren Tu², Di Guo², Vladislav Orekhov³, and Xiaobo Qu^1
1Department of Electronic Science, National Institute for Data Science in Health and Medicine, Xiamen University, Xiamen, China, ²School of Computer and Information Engineering, Xiamen University of Technology, Xiamen, China, ³Department of Chemistry and Molecular Biology, University of Gothenburg, Gothenburg, Sweden

Multi-dimensional nuclear magnetic resonance (NMR) spectroscopy is an invaluable biophysical tool but often suffers from long measurement. Several methods have been established for spectra reconstruction from undersampled data, two of which are model-based optimization and data-driven deep learning. Combining the main merits of them, we present a model-inspired flexible deep learning framework, for reliable, robust, and ultra-fast spectra reconstruction. Besides, we demonstrate that the model-inspired network needs very few parameters and is not sensitive to training datasets, which greatly reduces the demand for memory footprints and can work effectively in a wide range of scenarios without re-training.

Andreas Kofler¹, Tobias Schaeffter^1,2,3, and Christoph Kolbitsch^1,2
1Physikalisch-Technische Bundesanstalt, Berlin and Braunschweig, Berlin, Germany, ²School of Imaging Sciences and Biomedical Engineering, King's College London, London, United Kingdom, ³Department of Biomedical Engineering, Technical University of Berlin, Berlin, Germany

In this work, we combine Convolutional Neural Networks (CNN)- with Dictionary Learning (DL)- and Sparse Coding (SC)-based regularization for dynamic cardiac MR image reconstruction. The regularization on the image is imposed by patch-wise sparsity with respect to a learned overcomplete dictionary and closeness to a CNN-based image-prior which is obtained from a pre-trained CNN. We compare the proposed method to two iterative methods which incorporate the different components separately. We demonstrate the combination of CNNs with DL and SC leads to improved image quality and faster convergence compared to DL+SC only.

Eric Z. Chen¹, Xiao Chen¹, Jingyuan Lyu², Qi Liu², Zhongqi Zhang³, Yu Ding², Shuheng Zhang³, 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

Retrospectively gated cine (retro-cine) MRI is the clinical standard for cardiac functional analysis. To the best of our knowledge, this is the first work to evaluate the cine MRI with deep learning reconstruction for cardiac function analysis and compare it with other conventional methods. The cardiac functional values obtained from cine MRI with deep learning reconstruction are consistent with values from clinical standard retro-cine MRI.

Xiaoxin Li^1,2, Xinjie Lou¹, Junwei Yang³, Yong Chen⁴, and Dinggang Shen^2,5
1College of Computer Science and Technology, Zhejiang University of Technology, Hangzhou, China, ²School of Biomedical Engineering, ShanghaiTech University, Shanghai, China, ³Department of Computer Science and Technology, University of Cambridge, Cambridge, United Kingdom, ⁴Case Western Reserve University, Cleveland, OH, United States, ⁵Shanghai United Imaging Intelligence Co., Ltd., Shanghai, China

Recent studies explored the merits of using T1-weighted image (T1WI) as a guidance to reconstruct other undersampled MRI modalities via convolutional neural networks (CNNs). However, even aided by T1WI, the reconstruction of highly undersampled MRI data is still suffering from aliasing artifacts, due to the one-upsampling style of existing CNN architectures and not fully using the T1 information. To address these issues, we propose a deep Laplacian pyramid MRI reconstruction framework (LapMRI), which performs progressive upsampling while integrating multiscale prior of T1WI. We show that LapMRI consistently outperforms state-of-the-art methods and can preserve anatomical structure faithfully up to 12-fold undersampling.

Sangtae Ahn¹, Uri Wollner², Graeme McKinnon³, Rafi Brada², John Huston⁴, J. Kevin DeMarco⁵, Robert Y. Shih^5,6, Joshua D. Trzasko⁴, Dan Rettmann⁷, Isabelle Heukensfeldt Jansen¹, Christopher J. Hardy¹, and Thomas K. F. Foo^1
1GE Research, Niskayuna, NY, United States, ²GE Research, Herzliya, Israel, ³GE Healthcare, Waukesha, WI, United States, ⁴Mayo Clinic College of Medicine, Rochester, MN, United States, ⁵Walter Reed National Military Medical Center, Bethesda, MD, United States, ⁶Uniformed Services University of the Health Sciences, Bethesda, MD, United States, ⁷GE Healthcare, Rochester, MN, United States

Three-dimensional (3D) MRI can achieve higher spatial resolution and signal-to-noise ratio than 2D MRI at the expense of long scan times. Recently, deep-learning (DL) techniques have been applied to reconstruction from highly undersampled data, resulting in significant scan accelerations. To assess clinical acceptability, we evaluated DL-based reconstruction on 3D MPRAGE data, using scores from image evaluation by neuroradiologists. Our DCI-Net method with reduction factor R=10 received scores higher than or equal to those of conventional parallel imaging with R=2.1. This implies the DL method can accelerate scans by an additional factor of 5 while maintaining comparable diagnostic image quality.

satoshi ITO¹ and Shun UEMATSU^1
1Utsunomiya University, Utsunomiya, Japan

It has been reported that a quadratic phase scrambling of the spin system in advance to Fourier encoding (PSFT) is effective to the improvement of image quality. In this paper, an CNN-based image reconstruction using PSFT signal was examined. Simulation studies showed that proposed method allows equi-spaced under-sampling and that preservation of structure and image contrast were improved compared to standard Fourier transform based CS-CNN or iterative image reconstruction method. These studies indicate that PSFT has the possibility to reconstruct higher quality images in deep learning image reconstruction as well as iterative reconstruction.

Mahmut Yurt^1,2, Muzaffer Ozbey^1,2, Salman Ul Hassan Dar^1,2, Berk Tinaz^1,2,3, Kader Karlı Oğuz^2,4, and Tolga Çukur^1,2,5
1Department of Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey, ²National Magnetic Resonance Research Center, Bilkent University, Ankara, Turkey, ³Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, CA, United States, ⁴Department of Radiology, Hacettepe University, Ankara, Turkey, ⁵Neuroscience Program, Aysel Sabuncu Brain Research Center, Bilkent University, Ankara, Turkey

The gold-standard recovery models for accelerated multi-contrast MRI either involve volumetric or cross-sectional processing. Volumetric models offer elevated capture of global context, but may yield suboptimal training due to expanded model complexity. Cross-sectional models demonstrate improved training with reduced complexity, yet may suffer from loss of global consistency in the longitudinal dimension. We propose a novel progressively volumetrized generative model (ProvoGAN) for contextual learning of image recovery in accelerated multi-contrast MRI. ProvoGAN empowers capture of global and local context while maintaining lower model complexity by performing aimed volumetric mappings via a cascade of cross-sectional mappings task-optimally ordered across rectilinear orientations.

Linshan Xie^1,2, Yilong Liu^1,2, Linfang Xiao^1,2, Peibei Cao^1,2, Alex T. L. 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

In this work, we proposed a residual network based reconstruction method for multi-slice partial Fourier acquisition, where adjacent slices are sampled in a complementary way. The anatomical structure and phase similarity of multi-slice MR data can be exploited to provide complementary information from adjacent slices with different sampling patterns. The proposed method enables highly partial Fourier imaging without noise amplification.

Yuan Lian¹, Xinyu Ye¹, Yajing Zhang², and Hua Guo^1
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, School of Medicine, Tsinghua University, Beijing, China, ²MR Clinical Science, Philips Healthcare, Suzhou, China

Compressed Sensing theory is often applied to accelerate the acquisition of multi-contrast MR images. When highly undersampled, CS-MRI suffers from non-negligible reconstruction error. Here we propose an unrolled iterative deep-learning model to further utilize the group sparsity property for multi-contrast MRI reconstruction at high acceleration factor, named Joint-ISTA-Net, to reduce reconstruction error and aliasing. Our method adds a joint-shrinkage-thresholding model into ISTA-Net to generate a better reconstruction for multi-contrast image pairs. Experiments show the effectiveness of the proposed strategy.

Andrew Palmera Leynes^1,2 and Peder E.Z. Larson^1,2
1Department of Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States, ²UC Berkeley - UC San Francisco Joint Graduate Program in Bioengineering, Berkeley and San Francisco, CA, United States

Deep learning-based MR image reconstruction can provide greater scan time reduction than previously possible. However, this is limited only to MR acquisitions that have large training datasets with fixed hardware and acquisition configurations. We introduce a training- and database-free deep MR image reconstruction technique that may unlock acceleration factors beyond the limits of current state-of-the-art reconstruction methods while being generalizable to any hardware and acquisition configuration. We demonstrate Deep Non-Linear Inversion (DNLINV) on different anatomies and sampling patterns and show high quality reconstructions at higher acceleration factors than previously achievable.

Guangyuan Li¹, Chengyan Wang², Weibo Chen³, and Jun Lyu^1
1School of Computer and Control Engineering, Yantai University, Yantai, China, ²Human Phenome Institute, Fudan University, Shanghai, China, ³Philips Healthcare, Shanghai, China

At present, many deep learning-based methods have been proposed to solve the problems of traditional CS-MRI, but the reconstruction effect under highly under-sampling has not been well resolved. We proposed a modified GAN architecture for accelerating CS-MRI reconstruction, namely RSCA-GAN. The generator in the proposed architecture is composed of two residual U-net block, in which we added spatial and channel-wise attention (SCA). Each encoding-decoding block is composed of two residual blocks with short skip connections. SCA are added to the decoding block and residual block.

Yuxiang Dai¹, Cheng yan Wang², and He Wang^1
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, ²Human Phenome Institute, Fudan University, Shanghai, China

Compressed sensing (CS) has been employed to accelerate magnetic resonance imaging (MRI) by sampling fewer measurements. We proposed a fusion model based on the optimization method to integrate the image and gradient-based priors into CS-MRI for better reconstruction results via convolutional neural network models. In addition, the proposed fusion model exhibited effective reconstruction performance in magnetic resonance angiography (MRA).

Muheng Li¹, Jie Xiang², Fuyixue Wang^3,4, Zijing Dong^3,5, and Kui Ying^2
1Department of Automation, Tsinghua University, Beijing, China, ²Department of Engineering Phycics, Tsinghua University, Beijing, China, ³A. A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA, United States, ⁴Harvard-MIT Health Sciences and Technology, MIT, Cambridge, MA, United States, ⁵Department of Electrical Engineering and Computer Science, MIT, Cambridge, MA, United States

A machine learning based reconstruction framework for Echo Planar Time-resolved Imaging(EPTI) is proposed. This work utilized the special data acquisition trajectory of EPTI, a highly-accelerated spatiotemporal CAIPI sampling, to divide the k-space recovery task into a multi-process program. The missing data is filled within an indicated small kernel with a fully connected neural network. Through image reconstruction tests on human brain data set acquired by EPTI, we demonstrated the high efficiency of this algorithm by shortening the reconstruction time of 216×216×48×32 k-data from over 10 minutes to about 20 seconds.

Jaejin Cho^1,2, Qiyuan Tian^1,2, Robert Frost^1,2, Itthi Chatnuntawech³, and Berkin Bilgic^1,2,4
1Athinoula A. Martinos Center for Biomedical Imaging, Charlestown, MA, United States, ²Havard Medical School, Cambridge, MA, United States, ³National Nanotechnology Center, Pathum Thani, Thailand, ⁴Harvard/MIT Health Sciences and Technology, Cambridge, MA, United States

We propose a wave-encoded model-based deep learning (wave-MoDL) strategy for simultaneous image reconstruction and segmentation. We use CNN-based regularizers in both k- and image-space to employ features in both domains and successfully incorporated wave-encoding strategy into MoDL reconstruction. Wave-MoDL enables R_yxR_z=4x4-fold accelerated 3D imaging using a 32-channel array while reducing NRMSE by 1.45-fold compared to wave-CAIPI. Further, we jointly train wave-MoDL and a U-net for simultaneous reconstruction and segmentation to get additional gain.