Elizabeth Cole¹, Shreyas Vasanawala¹, and John Pauly^1
1Stanford University, Palo Alto, CA, United States

Deep learning (DL)-based image reconstruction methods have achieved promising results across multiple MRI applications. However, most approaches require large-scale fully-sampled ground truth data for supervised training. Acquiring fully-sampled data is often either difficult or impossible, particularly for dynamic datasets. We present a DL framework for MRI reconstruction which does not use fully-sampled data. We test the proposed method in two scenarios: retrospectively undersampled cine and prospectively undersampled abdominal DCE. Our unsupervised method can produce faster reconstructions which are non-inferior to compressed sensing. Our novel proposed method can enable accelerated imaging and accurate reconstruction in applications where fully-sampled data is unavailable.

Ziwen Ke¹, Zhuo-Xu Cui¹, Jing Cheng¹, Leslie Ying², Xin Liu¹, Hairong Zheng¹, Yanjie Zhu¹, and Dong Liang^1
1Shenzhen Institutes of Advanced Technology, Shenzhen, China, ²University at Buffalo, The State University of New York, Buffalo, NY, United States

Manifold learning has achieved success in cardiac MRI. It models the dynamic images as points on a smooth, low dimensional manifold in high dimensional space. The low dimensional assumption is extracted as a regularizer, but corresponding algorithms are not performed along with the manifold's nonlinear structure. In this paper, we propose a deep manifold learning for dynamic MR imaging. The manifold assumption is no longer taken as the regularization term in the proposed method, but the deep optimization model is directly developed on the nonlinear manifold. The validation on in vivo data shows that our method can achieve improved reconstruction.

Ziwen Ke¹, Wenqi Huang¹, Jing Cheng¹, Leslie Ying², Xin Liu¹, Hairong Zheng¹, Yanjie Zhu¹, and Dong Liang^1
1Shenzhen Institutes of Advanced Technology, Shenzhen, China, ²University at Buffalo, The State University of New York, Buffalo, NY, United States

Accelerated dynamic MR Imaging is a significant but challenging task. Our previous work demonstrated that deep low-rank priors could achieve improved reconstruction performance by unrolling a sparse and low-rank-based optimization algorithm. However, the optimization algorithm is highly customized, and currently, no deep learning methods exist to apply low-rankness as prior to general inverse problems. In this paper, we propose a plug-and-play low-rank network module in dynamic MR imaging. The low-rank network module can be easily embedded into other deep learning models. The embedding of the LR module can effectively improve the reconstruction results. 

Zhehong Zhang¹, Yuze Li², and Huijun Chen^2
1Department of Engineering Physics, Tsinghua University, Beijing, China, ²Department of Biomedical Engineering, School of Medicine, Tsinghua University, Beijing, China

The concatenation of several-element U-nets operating in both k-space and image domains is a deep learning network model that has been used for magnetic resonance image (MRI) reconstruction. Here, we present a new method that incorporates deformable 2D convolution kernels into the model. The proposed method leverages motion information of dynamic MRI and thus deformable convolution kernel naturally adapts to image structures. We demonstrate the improved performance of the proposed method using CINE dataset.

Jiaren Zou^1,2 and Yue Cao^1,2,3
1Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, United States, ²Department of Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States, ³Department of Radiology, University of Michigan, Ann Arbor, MI, United States

Joint optimization of deep learning based undersampling pattern and the reconstruction network has shown to improve the reconstruction accuracy for a given acceleration factor in static MRI. Here, we investigate the joint training of a reconstruction network, sampling pattern and data sharing for dynamic contrast-enhanced MRI. By adding a degree of freedom in the temporal direction to the sampling pattern, better reconstruction quality can be achieved. Jointly learned data sharing can further improve the reconstruction accuracy.

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 (UMRAM), Bilkent University, Ankara, Turkey, ³Neuroscience Program, Aysel Sabuncu Brain Research Center, Bilkent University, Ankara, Turkey

Deep neural networks (DNNs) have recently found emerging use in accelerated MRI reconstruction. DNNs typically learn data-driven priors from large datasets constituting pairs of undersampled and fully-sampled acquisitions. Acquiring such large datasets, however, might be impractical. To mitigate this limitation, we propose a few-shot learning approach for accelerated MRI that merges subject-driven priors obtained via physical signal models with data-driven priors obtained from a few training samples. Demonstrations on brain MR images indicate that the proposed approach requires just a few samples to outperform traditional parallel imaging and DNN algorithms.

Beliz Gunel¹, Morteza Mardani¹, Akshay Chaudhari², Shreyas Vasanawala², and John Pauly^1
1Electrical Engineering, Stanford University, Stanford, CA, United States, ²Radiology, Stanford University, Stanford, CA, United States

Untrained neural networks such as ConvDecoder have emerged as a compelling MR image reconstruction method. Although ConvDecoder does not require any training data, it requires tens of minutes to reconstruct a single MR slice at inference time, making the method impractical for clinical deployment. In this work, we propose using ConvDecoder to construct "weak labels" from undersampled MR scans at training time. Using limited supervised pairs and constructed weakly supervised pairs, we train an unrolled neural network that gives strong reconstruction performance with fast inference time, significantly improving over supervised and self-training baselines in the low data regime.

Abhinav Saksena¹, Makarand Parigi¹, Nicole Seiberlich², and Yun Jiang^2
1EECS, University of Michigan, Ann Arbor, MI, United States, ²Department of Radiology, University of Michigan, Ann Arbor, MI, United States

The purpose of this work is to test and evaluate a number of candidate loss functions for the reconstruction of diagnostic quality brain MRI images using undersampled k-space data and CNNs. We investigate both per-pixel (L1) and perceptual based (SSIM) loss functions, before developing a custom loss function that incorporates elements of both. We train these loss functions implemented in a UNet architecture on both 4x and 8x undersampled 16-coil MRI data. The custom loss function is shown to produce both the best quantitative results and also sharper and more detailed reconstructions across a number of image contrasts.

Aowen Liu¹, Meiling Ji², Xiaoqian Huang¹, Yawei Zhao², Renkuan Zhai², Guobin Li², Dinggang Shen¹, and Shu Liao^1
1United Imaging Intelligence, Shanghai, China, ²United Imaging Healthcare, Shanghai, China

Many deep learning models for MR image restoration have high computational cost, which raises significant hardware cost and also restricts their usage. To address it, we propose a lightweight network based on the encoder-decoder architecture which integrates image features of different scales and levels to improve the representation capability. A novel loss function is also designed to constrain the model in both image domain and frequency domain. The experimental results show that our model efficiently reduces computational burden while maintaining high performance compared to other conventional models.

Batu Ozturkler¹, Arda Sahiner¹, Mert Pilanci¹, Shreyas Vasanawala², John Pauly¹, and Morteza Mardani^1
1Electrical Engineering, Stanford University, Stanford, CA, United States, ²Radiology, Stanford University, Stanford, CA, United States

Deep-learning based reconstruction methods have shown great promise for undersampled MR reconstruction. However, their lack of interpretability, and the nonconvex nature impedes their utility as they may converge to undesirable local minima. Moreover, training deep networks in high-dimensional imaging applications such as DCE, and 4D flow requires large amounts of memory that may overload GPUs. Here, we advocate a layer-wise training method amenable to convex optimization, and scalable for training 3D-4D datasets. We compare convex layer-wise training to traditional end-to-end training. The proposed method matches the reconstruction quality of end-to-end training while it is interpretable, convex, and demands less memory.

Zilin Deng^1,2, Burhaneddin Yaman^1,2, Chi Zhang^1,2, Steen Moeller², and Mehmet Akçakaya^1,2
1University of Minnesota, Minneapolis, MN, United States, ²Center for Magnetic Resonance Research, Minneapolis, MN, United States

Unrolled neural networks have been shown to improve the reconstruction quality for accelerated MRI. While they have been widely applied in 2D settings, 3D processing may further improve reconstruction quality for volumetric imaging with its ability to capture multi-dimensional interactions. However, implementation of 3D unrolled networks is generally challenging due to GPU-memory limitations and lack of availability of large databases of 3D data. In this work, we tackle both these issues by an augmentation approach that generates smaller sub-volumes from large volumetric datasets. We then compare the 3D unrolled network to its 2D counterpart, showing the improvement from 3D processing. 

Ke Wang¹, Michael Kellman², Christopher M. Sandino³, Kevin Zhang¹, Shreyas S. Vasanawala⁴, Jonathan I. Tamir⁵, Stella X. Yu⁶, and Michael Lustig^1
1Electrical Engineering and Computer Sciences, University of California, Berkeley, Berkeley, CA, United States, ²Pharmaceutical Chemistry, University of California, San Francisco, Berkeley, CA, United States, ³Electrical Engineering, Stanford University, Stanford, CA, United States, ⁴Radiology, Stanford University, Stanford, CA, United States, ⁵Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX, United States, ⁶International Computer Science Institute, University of California, Berkeley, Berkeley, CA, United States

High-dimensional Deep learning (DL) reconstructions (e.g. 3D, 2D+time, 3D+time) can exploit  multi-dimensional information and achieve improved results over lower dimensional ones. However, the size of the network and its depth for these large-scale reconstructions are currently limited by GPU memory. Here, we use a memory-efficient learning (MEL) framework, which favorably trades off storage with minimal increased computation and enables deeper high-dimensional DL reconstruction on a single GPU. We demonstrate improved image quality with learned high-dimensional reconstruction enabled by MEL for in-vivo 3D MRI and 2D cardiac cine imaging applications. MEL uses much less GPU memory while minimally increasing training time.

Sebastian Rosenzweig^1,2 and Martin Uecker^1,2
1Diagnostic and Interventional Radiology, University Medical Center Göttingen, Göttingen, Germany, ²Partner Site Göttingen, German Centre for Cardiovascular Research (DZHK), Göttingen, Germany

Cardiac MRI is challenging because of respiratory and cardiac motion. Current clinical approaches try to bypass motion-related issues by ECG-triggering and breath-holds, which comes with several drawbacks. Alternatively, self-gating techniques can be used to determine respiratory and cardiac motion from the acquired raw-data itself. We present novel insights on the quadrature-pair self-gating signals estimated by SSA-FARY: We show that one element of each pair is certain to be in-phase with the motion it represents, as it is the result of a filtering process with a zero-phase filter. This enables the use of less respiratory bins, which decreases the computational demand.

Ibtisam Aslam^1,2, Fariha Aamir², Lindsey A CROWE¹, Miklos KASSAI¹, Hammad Omer², and Jean-Paul VALLEE^1
1Service of Radiology, Geneva University Hospitals and Faculty of Medicine, University of Geneva, Geneva, Switzerland, ²Medical Image Processing Research Group (MIPRG), Deptt. of Electrical & Computer Engineering, COMSATS University Islamabad, Islamabad, Pakistan

In a clinical setting, multiple breath-hold, multi-slice, ECG-gated cine MR (CMR) Cartesian acquisition is a gold standard. Multiple breath-holds in standard CMR acquisition can result in slice-misalignment due to inconsistent breath-hold positions and it forces long exam time. To reduce CMR scan time and to avoid slice-misalignment, under-sampled non-Cartesian (NC) trajectories are useful but lead to artifacts. This paper proposes U-Net based transfer-learning approach with NUFFT (NUFFT TL-Net) to reconstruct artifact-free whole heart, radial CMR images. The preliminary experiments show improved performance of the proposed NUFFT TL-Net both visually and in terms of evaluation parameters than contemporary NUFFT U-Net.

Qing Zou¹, Abdul Haseeb Ahmed¹, Prashant Nagpal¹, Rolf Schulte², and Mathews Jacob^1
1University of Iowa, Iowa City, IA, United States, ²GE Global Research, Munich, Germany

The main focus of this work is to introduce an unsupervised deep generative manifold model for the alignment and joint recovery of the slices in free-breathing and ungated cardiac cine MRI. The main highlights are

(1)   the ability to align multi-slice data and capitalize on the redundancy between the slices.

(2)   The ability to estimate the gating information directly from the k-t space data.

(3)   The unsupervised learning strategy that eliminates the need for extensive training data.

The joint recovery facilitates the acquisition of data from the whole heart in around 2 minutes of acquisition time.

Weijie Gan¹, Yu Sun¹, Cihat Eldeniz², Jiaming Liu³, Hongyu An², and Ulugbek S. Kamilov^1,3
1Department of Computer Science and Engineering, Washington University in St. Louis, St. Louis, MO, United States, ²Mallinckrodt Institute of Radiology, Washington University in St. Louis, St. Louis, MO, United States, ³Department of Electrical and Systems Engineering, Washington University in St. Louis, St. Louis, MO, United States

One of the key limitations in conventional deep learning-based MR image reconstruction is the need for registered pairs of training images, where fullysampled ground truth images are required as the target. We address this limitation by proposing a novel registration-augmented image reconstruction method that trains a CNN by directly mapping pairs of unregistered and undersampled MR measurements. The proposed method is validated on a single-coil MRI data set by training a model directly on pairs of undersampled measurements from images that have undergone nonrigid deformations.

Hemant Kumar Aggarwal¹ and Mathews Jacob^1
1University of Iowa, Iowa City, IA, United States

Deep learning image reconstruction algorithms often suffer from model mismatches when the acquisition scheme differs significantly from the forward model used during training. We introduce a Generalized Stein's Unbiased Risk Estimate (GSURE) loss metric to adapt or fine-tune the network to the measured k-space data, thus minimizing the impact of model misfit. Unlike current methods that rely on the mean square error in k-space, the proposed metric accounts for noise in the measurements. This makes the approach less vulnerable to overfitting, thus offering improved reconstruction quality compared to schemes that rely on mean-square error.

Peter Dawood^1,2, Martin Blaimer³, Peter M. Jakob¹, and Johannes Oberberger^2
1Department of Experimental Physics 5, University of Würzburg, Würzburg, Germany, ²Department of Internal Medicine I, University Hospital Würzburg, Würzburg, Germany, ³Magnetic Resonance and X-Ray Imaging Department, Development Center X-ray Technology EZRT, Fraunhofer Institute for Integrated Circuits IIS, Würzburg, Germany

The parallel imaging method GRAPPA has been generalized within the Machine Learning framework by introducing the deep-learning method RAKI, in which Convolutional Neural Networks are used for non-linear k-space interpolation. RAKI is a database-free approach that uses scan-specific calibration data. Here, we study the influence of the calibration data on the image quality of 2D imaging sequences. The results indicate that RAKI yields superior signal-to-noise ratio but introduces blurring and loss of detail for typical calibration data amounts at high accelerations. Furthermore, the contrast information in the calibration data must be similar to that of the accelerated scans. 

Yuze Li¹, Zhangxuan Hu², Haikun Qi³, Guangqi Li¹, Dongyue Si¹, Haiyan Ding¹, Hua Guo¹, and Huijun Chen^1
1Center for Biomedical Imaging Research, Medical School, Tsinghua University, Beijing, China, ²MR Research China, GE Healthcare, Beijing, China, ³King’s College London, London, United Kingdom

In this study, a deep learning-based MR reconstruction framework called DLNUFFT (Deep Learning-based Non-Uniform Fast Fourier Transform) was proposed, which can restore the under-sampled non-uniform k-space to fully sampled Cartesian k-space without NUFFT gridding. Novel network blocks with fully learnable parameters were built to replace the hand-crafted convolution kernel and the density compensation in NUFFT. Simulations and in-vivo results showed DLNUFFT can achieve higher performance than conventional NUFFT, compressed sensing and state-of-the-art deep learning methods in terms of PSNR and SSIM.

Linfang Xiao^1,2, Yilong Liu^1,2, Yujiao Zhao^1,2, Zheyuan Yi^1,2,3, Vick Lau^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, ³Department of Electrical and Electronic Engineering, Southern University of Science and Technology, Shenzhen, China

Fast spin echo (FSE) is the most commonly used multi-shot sequence in clinical MRI. In this study, we propose to acquire single-channel FSE data with incomplete number of shots (TRs), and reconstruct such periodically undersampled k-space data using a deep learning approach. The results demonstrate that the proposed method can effectively remove the aliasing artifacts and recover the high frequency information without noise amplification, enabling a FSE acceleration that can be readily implemented in practice.