Zi Wang¹, Chen Qian¹, Di Guo², Hongwei Sun³, Rushuai Li⁴, and Xiaobo Qu^1
1Department of Electronic Science, Biomedical Intelligent Cloud R&D Center, Fujian Provincial Key Laboratory of Plasma and Magnetic Resonance, 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, ³United Imaging Research Institute of Intelligent Imaging, Beijing, China, ⁴Department of Nuclear Medicine, Nanjing First Hospital, Nanjing Medical University, Nanjing, China

Deep learning has shown astonishing performance in accelerated MRI. Most methods adopt the convolutional neural network and perform 2D convolution since many MR images or their corresponding k-space are in 2D. In this work, we try a different approach that explores the memory-friendly 1D convolution, making the deep network easier to be trained and generalized. Furthermore, a one-dimensional deep learning architecture (ODL) is proposed for MRI reconstruction. Results demonstrate that, the proposed ODL provides improved reconstructions than state-of-the-art methods and shows nice robustness to some mismatches between the training and test data.

Yirong Zhou¹, Chen Qian¹, Yi Guo², Zi Wang¹, Jian Wang¹, Biao Qu³, Di Guo², Yongfu You⁴, and Xiaobo Qu^1
1Biomedical Intelligent Cloud R&D Center, Department 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 Instrumental and Electrical Engineering, Xiamen University, Xiamen, China, ⁴Biomedical Intelligent Cloud R&D Center, School of Electronic Science and Engineering, China Mobile Group, Xiamen, China

Machine learning and artificial intelligence have shown remarkable performance in accelerated magnetic resonance imaging (MRI). Cloud computing technologies have great advantages in building an easily accessible platform to deploy advanced algorithms. In this work, we develop a high-performance medical intelligence cloud computing platform (XCloud-pFISTA) to reconstruct MRI images from undersampled k-space data. Two state-of-the-art approaches of the Projected Fast Iterative Soft-Thresholding Algorithm (pFISTA) family have been successfully implemented on the cloud. This work can be considered as a good example of cloud-based medical image reconstruction and may benefit the future development of integrated reconstruction and online diagnosis system.

Yu Zhang¹, Chunchao Xia¹, Xiaoyong Zhang², and Zhenlin Li^1
1Department of Radiology, West China Hospital of Sichuan University, Chengdu, China, ²Clinical Science, Philips Healthcare, Chengdu, China

 In this work, we aimed to use a deep learning based Compressed SENSE reconstruction algorithm, presented here as Artificial Intelligence Compressed-SENSE (AI-CS), to improve image quality of 3D magnetic resonance cholangiopancreatography (MRCP) in a breath hold (BH). The results demonstrated that AI-CS BH MRCP can enable improved image quality and show great visibility of small ductal structures than other previous methods.

Tomoki Miyasaka¹, Satoshi Funayama², Daiki Tamada², Hiroyuki Morisaka², Hiroshi Onishi², and Yasuhiko Terada^1
1Graduate School of Science and Technology, University of Tsukuba, Tsukuba, Japan, ²Department of Radiology, University of Yamanashi, Chuo, Japan

Deep learning has been attracting attention as a new tool for image reconstruction. However, there is a lack of appropriate automatic evaluation metrics for reconstruction performance of small structures such as lesions, which poses a high hurdle for clinical application. Here, we explored the relationship between radiomic features of tumors and various DL reconstruction conditions, and proposed a new method based on radiomics to evaluate the reconstruction performance of DL against lesions. Based on the analysis using the concordance correlation coefficients for ground truth images, we explored several texture features that are sensitive to differences in reconstruction methods and conditions.
 

Yi-Cheng Hsu¹, Patrick Liebig², and Ying-Hua Chu^1
1Siemens Healthineers Ltd., Shanghai, China, ²Siemens Healthcare GmbH, Erlangen, Germany

We proposed a method to approximate the magnitude of B₁^- and correct the bias field induced signal inhomogeneity at 7T. For 2D and 3D T2 weighted images, the signal intensity was more homogeneous and the signal void was recovered using the proposed method. This method can be implemented to correct bias field effect for other imaging methods using the estimated |B₁^-| and |B₁⁺|.

Yukio Kaneko¹, Atsuro Suzuki¹, Tomoki Amemiya¹, Chizue Ishihara¹, Yoshitaka Bito², and Toru Shirai^1
1Innovative Technology Laboratory, FUJIFILM Healthcare Corporation, Tokyo, Japan, ²Radiation Diagnostic Systems Division, FUJIFILM Healthcare Corporation, Tokyo, Japan

Recently, deep learning techniques for high-speed or high-quality imaging in MRI have been reported. However, deep learning techniques for the inhomogeneous spatial distribution of noise caused by parallel imaging have not been fully established. In this study, “Multi-Adaptive Convolutional Neural Network Reconstruction (MA-CNNR)” has been investigated. A noisy image was segmented into four regions by g-factor map, and different optimized CNNs were selected for each region. A denoised image was generated by combining the four denoised regions. The denoising effect was evaluated for 1.5T brain images, and it was confirmed that MA-CNNR can reduce the inhomogeneous noise in parallel imaging.

Zimeng Li^1,2, Sa Xiao², Cheng Wang², Chaohui Ye^1,2, and Xin Zhou^2
1School of Physics, Huazhong University of Science and Technology, Wuhan, China, ²Key Laboratory of Magnetic Resonance in Biological Systems, State Key Laboratory of Magnetic Resonance and Atomic and Molecular Physics, National Center for Magnetic Resonance in Wuhan, Wuhan Institute of Physics and Mathematics, Innovation Academy for Precision Measurement Science and Technology, Wuhan, China

Hyperpolarized gas MRI is a non-invasive and non-radiation imaging modality that can provide lung structure and function information. However, the problem of long imaging time limits its clinical application. Deep learning-based methods have shown great potential to accelerate MRI. In this work, we proposed a multi-task network to perform pulmonary hyperpolarized gas MRI reconstruction and lung region segmentation simultaneously, which allows sharing representations between two tasks. The results show that the proposed multi-task network has better reconstruction performance, stronger robustness and fewer model parameters than the comparison method.

Geonhui Son¹, Yohan Jun¹, Sewon Kim¹, Dosik Hwang¹, and Taejoon Eo*^1
1Electrical and Electronic Engineering, Yonsei university, Seoul, Korea, Republic of

Multi-contrast images acquired with magnetic resonance imaging (MRI) provide abundant diagnostic information. However, the applicability of multi-contrast MRI is often limited by slow acquisition speed and high scanning cost. To overcome this issue, we propose a contrast generation method for arbitrary missing contrast images. First, StyleGAN2-based multi-contrast generator is trained to generate paired multi-contrast images. Second, pSp-based encoder network is used to predict style vectors from input images. Consequently, the imputation for arbitrary missing contrast is achieved by the process of (1) embedding one or more kinds of contrast images and (2) forward-propagating the style vector to the multi-contrast generator.

Sena Azamat^1,2, Buse Buz-Yaluğ¹, Abdullah Baş¹, Alpay Ozcan³, Ayça Ersen Danyeli^4,5, Kubra Tan⁶, Ozge Can⁷, Necmettin Pamir^5,8, Alp Dinçer^5,9, Koray Ozduman^5,8, and Esin Ozturk-Isik^1
1Institute of Biomedical Engineering, Bogazici University, Istanbul, Turkey, ²Department of Radiology, Basaksehir Cam and Sakura City Hospital, Istanbul, Turkey, ³Electric and Electronic Engineering Department, Bogazici University, Istanbul, Turkey, ⁴Department of Medical Pathology, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁵Center for Neuroradiological Applications and Reseach, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁶Health Institutes of Turkey, Istanbul, Turkey, ⁷Department of Medical Engineering, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁸Department of Neurosurgery, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁹Department of Radiology, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey

Meningiomas are the most common primary intracranial tumors in adults. S100 protein expression (S100+) in meningiomas is a marker of neural crest origin. Eighty-four patients with preoperative MRI were included in this IRB approved study. The whole tumor volumes were segmented from FLAIR, followed by co-registration onto SWI. Supervised machine and deep learning methods were employed to categorize meningiomas into S100+ and S100- groups based on SWI histogram values. Ensemble bagged trees resulted in an accuracy of 85.7% (sensitivity=87.0 % and specificity=84.4 %), while a Resnet 50 architecture had 70.5% accuracy (sensitivity=80%, specificity=57.1%) for predicting S100 protein expression in meningiomas.

Banu Sacli-Bilmez¹, Abdullah Baş², Kübra Tan³, Ayça Erşen Danyeli^4,5, Özge Can⁶, M.Necmettin Pamir^5,7, Alp Dinçer^5,8, Koray Özduman^5,7, and Esin Ozturk-Isik^1
1Institute of Biomedical Engineering, Bogazici University, İstanbul, Turkey, ²Institute of Biomedical Engineering, Bogazici University, Istanbul, Turkey, ³Health Institutes of Turkey, İstanbul, Turkey, ⁴Department of Medical Pathology, Acıbadem Mehmet Ali Aydınlar University, İstanbul, Turkey, ⁵Center for Neuroradiological Applications and Reseach, Acıbadem Mehmet Ali Aydınlar University, İstanbul, Turkey, ⁶Department of Medical Engineering, Acıbadem Mehmet Ali Aydınlar University, İstanbul, Turkey, ⁷Department of Neurosurgery, Acıbadem Mehmet Ali Aydınlar University, İstanbul, Turkey, ⁸Department of Radiology, Acıbadem Mehmet Ali Aydınlar University, İstanbul, Turkey

Loss of neurofibromatosis 2 (NF2-L) is a well-known genetic alteration of meningiomas and causes meningiomas to evolve into more aggressive and infiltrating form. This study aims to investigate single-voxel proton magnetic resonance spectroscopy (¹H-MRS) correlations of NF2-L in meningiomas and to develop machine learning and deep learning models to identify NF2-L in meningiomas. NF2-L meningiomas had significantly higher Ins, Lac, and Ins+Glyc, and lower tNAA than tumors with no copy number loss. While a subspace discriminant model achieved a classification accuracy of 77.25%, a 1D-CNN model obtained a classifcation accuracy of 88.9% for identifying NF2-L meningiomas.

Nhan Duc Nguyen¹, Joon Sik Park¹, Seung Hong Choi², Roh-Eul Yoo², and Jaeseok Park^1,3
1Department of Intelligent Precision Healthcare Convergence, Sungkyunkwan University, Suwon, Korea, Republic of, ²Department of Radiology, Seoul National University Hospital, Seoul, Korea, Republic of, ³Department of Biomedical Engineering, Sungkyunkwan University, Suwon, Korea, Republic of

To take characterization of various vascular contrast dynamics into account, in this work we propose a novel, vascular heterogeneity model based deep learning reconstruction from highly undersampled data for high-definition whole brain DCE MRI. To this end, we introduce a new, vascular contrast dynamics (VCD) weighted deep attention neural network (VACAN) consisting of: 1) a vascular adaptive attention 3D U-Net, 2) a multilayered non-negative matrix factorization (NMF) layer, and 3) a data consistency layer. Experimental studies are performed using highly undersampled patient data to validate the effectiveness of the proposed VACAN against conventional 3D U-Net.

Congcong Liu^1,2, Zhuoxu Cui¹, Zhilang Qiu¹, Haoxiang Li^1,2, Yifan Guo¹, Chentao Cao^1,2, Xin Liu^1,2, Hairong Zheng^1,2, Dong Liang^1,2,3, and Haifeng Wang^1,2
1Shenzhen Institutes of Advanced Technology,Chinese Academy of Sciences, Shenzhen, China, ²Shenzhen College of Advanced Technology, University of Chinese Academy of Sciences, Shenzhen, China, ³Research Centre for Medical AI, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China

CSM or kernel needs to be estimated in conventional parallel image, which is very time-consuming and the estimation process may be inaccurate. Here, wave coding model based on virtual conjugate coil using deep generative modes (WV-DGM) is proposed for the virtual conjugate coil (VCC) extended model based on wave coding.  WV-DGM combined with deep DGM without training can realize advantages of wave encoding and introduce extra phase to reduce the ill-condition of the model via VCC. The results in vivo demonstrated that WV-DGM can achieve better quality compared with conventional SENSE while 10x times used to estimate CSM is reduced.

Junhao Zhang^1,2, Zheyuan Yi^1,2, Yujiao Zhao^1,2, Linfang Xiao^1,2, Jiahao Hu^1,2, Christopher Man^1,2, Yujiao Zhao³, 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, ³The University of HongKong, HongKong, China

Conventional ESPIRiT reconstruction requires accurate estimation of ESPIRiT maps from autocalibration samples or signals but acquiring such autocalibration signals takes time and may not be straightforward in some situations. This study aims to deploy deep learning to directly estimate ESPIRiT maps from uniformly undersampled multi-channel 2D MR data that contain no autocalibration signals.  Results show that the estimated ESPIRiT maps could be reliably obtained and they could be used for ESPIRiT and SENSE reconstruction with high acceleration.

Muhammad Shafique^1,2, Sohaib Ayyaz Qazi^1,3, Irfan Ullah¹, and Hammad Omer^1
1Electrical and Computer Engineering, COMSATS University Islamabad, Islamabad, Pakistan, ²Electrical Engineering, University of the Poonch Rawalakot, Rawalakot, Pakistan, ³Service of Radiology and Faculty of Medicine, University of Geneva University Switzerland, Geneva, Swaziland

The Low rank and Sparse (L+S) matrix decomposition model has been proposed in literature to reconstruct the undersampled dynamic MRI data. The limitations of L+S method include an effective separation of the low-rank and sparse components from the acquired dynamic MRI data; also the algorithm is computationally expensive. In this paper, Compressed Singular Value Decomposition (cSVD) is employed in L+S method. The results show that the proposed method provides effective separation of the L and S components as well as considerably reduces the computation time.

Xinyu Ye^1,2, Zepeng Wang^2,3, and Fan Lam^2,3
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, School of Medicine, Tsinghua University, Beijing, China, ²Department of Bioengineering, University of Illinois at Urbana-Champaign, Urbana, IL, United States, Urbana, IL, United States, ³Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, Urbana, IL, United States

Removing nuisance signals is an essential step for MRSI. A union-of-subspaces model that uses spatiospectral priors has achieved excellent water/lipid removal performance for ¹H-MRSI, but may not be sufficient when the initial water/lipids are too strong and/or when field inhomogeneity is severe. We propose a low-rank plus sparse method for improved nuisance removal. The sparsity term is used to capture residual nuisance failed to be captured by the union-of-subspace model and the low-rank term with learned subspaces protects metabolite signals. Results from in vivo ¹H-MRSI data show that the proposed method led to improved nuisance signal removal.