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

We propose a U-net based architecture for multi-coil MRI reconstruction. The model is able to utilize the multi-coil nature of the data by processing images before they are combined, unlike U-nets which require coil combination before inputting the image. We achieved SSIM scores higher than the U-net with much fewer parameters, enabling lower memory usage.

Shohei Ouchi¹ and Satoshi ITO^1
1Utsunomiya University, Utsunomiya, Japan

CNN based Compressed Sensing reconstruction has attracted much attention. It is difficult to restore higher spatial frequency components even with CNN-CS. We proposed a new transformed image domain CNN-CS method based on equal-sized multi-resolution image decomposition (eFREBAS transform). The eFREBAS transform is multi resolution analysis method that divide the image into equal-sized sub-images. This CNN consisting of three U-Net CNNs, each with multi-channel input and output reconstructs MR phase varied images in frequency band-to-band through estimating artifact-free sub-images from under-sampled sub-images. Reconstruction experiments showed that eFREBAS-CNN could reconstruct sharp images that have strong phase variation.

Maarten Terpstra^1,2, Matteo Maspero^1,2, Jan Lagendijk¹, and Cornelis A.T. van den Berg^1,2
1Department of Radiotherapy, University Medical Center Utrecht, Utrecht, Netherlands, ²Computational Imaging Group for MR diagnostics & therapy, University Medical Center Utrecht, Utrecht, Netherlands

The \(\ell^2\) norm is the default loss function for complex image reconstruction. In this work, we investigate the behavior of the \(\ell^1\) and \(\ell^2\) loss functions for complex image reconstruction with non-complex-valued models. Simulations show that these norms assign a lower loss to reconstructions with lower magnitude, introducing an asymmetry in the loss function. To address this, we propose a new, symmetric loss function, and train deep learning models to show that the proposed loss function achieves better performance and faster convergence on complex image reconstruction tasks.

Madiha Arshad¹, Mahmood Qureshi¹, Omair Inam¹, and Hammad Omer^1
1Medical Image Processing Research Group (MIPRG), Department of Electrical and Computer Engineering, COMSATS University, Islamabad, Pakistan

In parallel imaging, the array coil with a large number of channels accelerates the data acquisition speed and provides high quality reconstructed images at the cost of an increase in computational complexity, image reconstruction time and memory requirement. Recently, Principal Component Analysis (PCA) has been used to compress multiple physical channels to a few virtual channels. However, the results showed a degradation in image quality and loss of sensitivity information. We introduce a novel channel compression technique combining PCA and U-Net before Compressed Sensing MRI reconstruction. The experimental results show less channel compression losses and retention of coil sensitivity information.

Fanwen Wang¹, Hui Zhang¹, Jiawei Han¹, Fei Dai¹, Yuxiang Dai¹, Weibo Chen², Chengyan Wang³, and He Wang^1,3
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, ²Philips Healthcare, Shanghai, China, ³Human Phenome Institute, Fudan University, Shanghai, China

This study proposed a novel MAGnitude Image to Complex (MAGIC) Network to reconstruct images using deep learning with limited number of training data. Collecting complex multi-coil data is inconvenient since it is beyond the routine examination. However, there are many magnitude images available in hospitals. By applying deformation between the magnitude image and complex image, MAGIC Net succeeded in synthesizing deformed data for training and enabled deep learning methods. Results show that with the same original data, MAGIC-Net outperforms the conventional CG-SENSE in PSNR for all undersampling trajectories with high resolution b = 0 and b = 1000 s/mm².

Moritz Blumenthal¹ and Martin Uecker^1,2,3,4
1Institute for Diagnostic and Interventional Radiology, University Medical Center Göttingen, Göttingen, Germany, ²DZHK (German Centre for Cardiovascular Research), Göttingen, Germany, ³Campus Institute Data Science (CIDAS), University of Göttingen, Göttingen, Germany, ⁴Cluster of Excellence “Multiscale Bioimaging: from Molecular Machines to Networks of Excitable Cells” (MBExC), University of Göttingen, Göttingen, Germany

Deep learning offers powerful tools for enhancing image quality and acquisition speed of MR images. Standard frameworks such as TensorFlow and PyTorch provide simple access to deep learning methods. However, they lack MRI specific operations and make reproducible research and code reuse more difficult due to fast changing APIs and complicated dependencies. By integrating deep learning into the MRI reconstruction toolbox BART, we have created a reliable framework combining state-of-the-art MRI reconstruction methods with neural networks. For demonstration, we reimplemented the Variational Network and MoDL. Both implementations achieve similar performance as implementations using TensorFlow.

Marius Arvinte¹, Sriram Vishwanath¹, Ahmed H Tewfik¹, and Jonathan I Tamir^1,2,3
1Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX, United States, ²Diagnostic Medicine, Dell Medical School, The University of Texas at Austin, Austin, TX, United States, ³Oden Institute for Computational Engineering and Sciences, The University of Texas at Austin, Austin, TX, United States

Accurate reconstruction using parallel imaging relies on estimating a set of sensitivity maps from a fully-sampled calibration region, which can lead to reconstruction artifacts in poor signal-to-noise ratio conditions. We introduce Deep J-Sense as a deep learning approach for jointly estimating the image and the sensitivity maps in the frequency-domain. We formulate an alternating minimization problem that uses convolutional neural networks for regularization and train the unrolled model end-to-end. We compare reconstruction performance with model-based deep learning methods that only optimize the image and show that our approach is superior.

Guanxiong Luo¹, Moritz Blumenthal¹, and Martin Uecker^1,2
1Institute for Diagnostic and Interventional Radiology, University Medical Center Göttingen, Germany, Göttingen, Germany, ²Campus Institute Data Science (CIDAS), University of Göttingen, Germany, Göttingen, Germany

The application of deep learning has is a new paradigm for MR image reconstruction. Here, we demonstrate how to incorporate trained neural networks into pipelines using reconstruction operators already provided by the BART toolbox. As a proof of concept, we demonstrate how to incorporate a deep image prior trained via TensorFlow into reconstruction within BART's framework.

Soumick Chatterjee^1,2,3, Chompunuch Sarasaen^1,4, Alessandro Sciarra^1,5, Mario Breitkopf¹, Steffen Oeltze-Jafra^5,6,7, Andreas Nürnberger^2,3,7, and Oliver Speck^1,6,7,8
1Department of Biomedical Magnetic Resonance, Otto von Guericke University, Magdeburg, Germany, ²Data and Knowledge Engineering Group, Otto von Guericke University, Magdeburg, Germany, ³Faculty of Computer Science, Otto von Guericke University, Magdeburg, Germany, ⁴Institute of Medical Engineering, Otto von Guericke University, Magdeburg, Germany, ⁵MedDigit, Department of Neurology, Medical Faculty, University Hopspital, Magdeburg, Germany, ⁶German Centre for Neurodegenerative Diseases, Magdeburg, Germany, ⁷Center for Behavioral Brain Sciences, Magdeburg, Germany, ⁸Leibniz Institute for Neurobiology, Magdeburg, Germany

One of the common problems in MRI is the slow acquisition speed, which can be solved using undersampling. But this might result in image artefacts. Several deep learning based techniques have been proposed to mitigate this problem. Most of these methods work only in the image space. Fine anatomical structures obscured by artefacts in the image can be challenging to reconstruct for a model working in the image space, but not in k-space. In this research, a novel complex-valued ResNet has been proposed to work directly in the k-space to reconstruct undersampled MRI. The preliminary experiments have shown promising results.

Nalini M Singh^1,2, Juan Eugenio Iglesias^1,3,4, Elfar Adalsteinsson^2,5,6, Adrian V Dalca^1,4, and Polina Golland^1,5,6
1Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, MA, United States, ²Harvard-MIT Division of Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, United States, ³Centre for Medical Image Computing, University College London, London, United Kingdom, ⁴A. A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Charlestown, MA, United States, ⁵Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ⁶Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA, United States

We propose a deep learning approach for reconstructing undersampled k-space data corrupted by motion. Our algorithm achieves high-quality reconstructions by employing a novel neural network architecture that captures the correlation structure jointly present in the frequency and image spaces. This method provides higher quality reconstructions than techniques employing solely frequency space or solely image space operations. We further characterize the motion severities for which the proposed method is successful. This analysis represents the first step toward fast image reconstruction in the presence of substantial motion, such as in pediatric or fetal imaging.

Juhyung Park¹, Dongmyung Shin¹, Hyeong-Geol Shin¹, Jiye Kim¹, and Jongho Lee^1
1Seoul National University, Seoul, Korea, Republic of

            We designed a slab-encoding RF pulse set, referred to as DeepSlider, using deep reinforcement learning to improve the robustness and performance. The condition number of the RF encoding matrix, which determines the sensitivity of the design to noise, was optimized via deep reinforcement learning and gradient descent. Additionally, the design was extended from reconstructing the real signal to the complex signal, allowing us to utilize the complex signal instead of the real part, expanding the application of the pulse to phase-based imaging.

Renkuan Zhai¹, Xiaoqian Huang², Yawei Zhao¹, Meiling Ji¹, Xuyang Lv², Mengyao Qian¹, Shu Liao², and Guobin Li^1
1United Imaging Healthcare, Shanghai, China, ²United Imaging Intelligence, Shanghai, China

The advantages of Convolutional Neural Networks (CNN) for MRI acceleration have been widely reported, but one remaining problem is that the significantly complex network makes itself less explainable than conventional model-based methods. In this work, a novel deep learning assisted MRI acceleration method is introduced to address the uncertainty of CNN by integrating its output as another constraint into the framework of Compressed Sensing (CS).

Simon Weinmüller¹, Hoai Nam Dang¹, Alexander Loktyushin^2,3, Felix Glang², Arnd Doerfler¹, Andreas Maier⁴, Bernhard Schölkopf³, Klaus Scheffler^2,5, and Moritz Zaiss^1,2
1Neuroradiology, University Clinic Erlangen, Friedrich-Alexander Universität Erlangen-Nürnberg (FAU), Erlangen, Germany, ²Max-Planck Institute for Biological Cybernetics, Magnetic Resonance Center, Tübingen, Germany, ³Max-Planck Institute for Intelligent Systems, Empirical Inference, Tübingen, Germany, ⁴Pattern Recognition Lab Friedrich-Alexander-University Erlangen-Nürnberg, Erlangen, Germany, ⁵Department of Biomedical Magnetic Resonance, Eberhard Karls University Tübingen, Tübingen, Germany

MRzero is a fully differentiable Bloch-equation-based MRI sequence invention framework. Instead of using time-consuming average-isochromat-based Bloch simulations, analytic signal equations are used as alternative forward differentiable MR scan simulation method. Neural network reconstruction is used for efficient auto-encoding. The joint optimization of sequence and NN parameters for B1 and T1 mapping can be performed 2 to 3 orders of magnitude faster then in previous MRzero approaches. The optimized sequence is tested by measurements in vivo at 3T and compared to a standard inversion recovery. High quality B1 and T1 maps are provided with less total acquisition time and energy deposition.

Hongjun An¹, Dongmyung Shin¹, Hyeong-Geol Shin¹, Woojin Jung¹, and Jongho Lee^1
1Department of Electrical and computer Engineering, Seoul National University, Seoul, Korea, Republic of

In MRI, scan parameters are carefully selected for desired results. We propose a new optimization method, AutoSON, that determines optimal scan parameters for flexibly-selected objectives on given tissue properties such as distribution and noises. AutoSON optimizes not only scan parameters but also a neural estimator, which extracts the desired information from MR signals (e.g., quantification mapping). The method successfully optimized the flip angles in DESPOT1 for T₁ mapping and classification of white matter and gray matter. The last example does not have a simple analytic equation, and therefore, demonstrates a potential utility of the method.

Vihang Agarwal¹, Yue Cao¹, and James Balter^1
1Radiation Oncology, University of Michigan, Ann Arbor, MI, United States

Accelerating MRI acquisition by under-sampling measurements in k-space and learning an image reconstruction model with high image quality is necessary to expand its clinical utilization. In this work, we explore joint optimization of under-sampling patterns and image reconstruction neural networks for aggressive sub-sampling of images that require very long acquisition times (e.g., FLAIR T2 weighted images). We propose Attention Residual Non-Local Networks (ARNL-Net) trained with an uncertainty based L1 loss function for producing high quality images. Initial experiments demonstrate the practicability of this method, with reconstructions demonstrating superior fidelity to fully sampled images as compared to random under-sampling schemes.