Zheyuan Yi^1,2,3, Yujiao Zhao^1,2, Linfang Xiao^1,2, Yilong Liu^1,2, Christopher Man^1,2, Jiahao Hu^1,2,3, Vick Lau^1,2, Alex Leong^1,2, Fei Chen³, and Ed X. Wu^1,2
1Laboratory of Biomedical Imaging and Signal Processing, The University of Hong Kong, Hong Kong SAR, China, ²Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong SAR, China, ³Department of Electrical and Electronic Engineering, Southern University of Science and Technology, Shenzhen, China

In traditional parallel imaging, calibration data need to be acquired, prolonging data acquisition time or/and sometimes increasing the susceptibility to motion. Low-rank parallel imaging has emerged as a calibrationless alternative that formulates reconstruction as a structured low-rank matrix completion problem while incurring a cumbersome iterative reconstruction process. This study achieves a fast and calibrationless low-rank reconstruction by estimating high-quality multi-channel spatial support directly  from undersampled data via deep learning. It offers a general and effective strategy to advance low-rank parallel imaging by making calibrationless reconstruction more efficient and robust in practice.

Yuan Lian¹, Xinyu Ye¹, Hai Luo², Ziyue Wu², and Hua Guo^1
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, School of Medicine, Tsinghua University, Beijing, China, ²Marvel Stone Healthcare Co., Ltd., Wuxi, China

Owing to hardware advancements, interest in low-field MRI system has increased recently. However, the imaging quality of low-field MRI is limited due to intrinsic low signal to noise ratio (SNR). Here we propose a deep-learning model to jointly denoise multi-contrast images using Noise2Noise training strategy. Our method can promote the SNRs of multi-contrast low-field images, and experiments show the effectiveness of the proposed strategy.

Min Feng¹, Qiqi Tong², Yingying Li¹, Bo Dong¹, JIanhui Zhong^1,3, and Hongjian He^1
1Center for Brain Imaging Science and Technology, College of Biomedical Engineering and Instrumental Science, Zhejiang University, HANGZHOU, China, ²Research Center for Healthcare Data Science, Zhejiang Lab, HANGZHOU, China, ³Department of Imaging Sciences, University of Rochester, Rochester, NY, United States

This study proposed a novel 3D residual network to learn end-to-end reconstruction from as few as eight DWIs to volumetric DKI parameters. The weighted loss function combining perceptual loss is utilized, which helps the network capture in-depth feature of DKI parameters. The results show that our method achieves superior performance over state-of-the-art methods for providing accurate DKI parameters as well as preserves rich textural details and improves the visual quality of reconstructions.

Linfang Xiao^1,2, Yilong Liu^1,2, Zheyuan Yi^1,2, Yujiao Zhao^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

Recently, deep learning methods have shown superior performance on image reconstruction and noise suppression by implicitly yet effectively learning prior information. However, end-to-end deep learning methods face the challenge of potential numerical instabilities and require complex application specific training. By taking advantage of the multichannel spatial encoding (as exploited by conventional parallel imaging reconstruction) and prior information (exploited by deep learning methods), we propose to embed a deep learning module into the iterative low-rank matrix completion based image reconstruction. Such strategy significantly suppresses the noise amplification and accelerates iteration convergence without image blurring.

Uten Yarach¹, Itthi Chatnuntawech², Congyu Liao³, Surat Teerapittayanon², Siddharth Srinivasan Iyer^4,5, Tae Hyung Kim^6,7, Jaejin Cho^6,7, Berkin Bilgic^6,7, Yuxin Hu⁸, Brian Hargreaves^3,8,9, and Kawin Setsompop^3,8
1Department of Radiologic Technology, Faculty of Associated Medical Science, Chiang Mai University, Chiang Mai, Thailand, ²National Nanotechnology Center (NANOTEC), National Science and Technology Development Agency (NSTDA), Pathum Thani, Thailand, ³Department of Radiology, Stanford University, Stanford, CA, United States, ⁴Department of Radiology, Stanford University, Stanford, Stanford, CA, United States, ⁵Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ⁶Athinoula A. Martinos Center for Biomedical Imaging, Charlestown, MA, United States, ⁷Department of Radiology, Harvard Medical School, Boston, MA, United States, ⁸Department of Electrical Engineering, Stanford University, Stanford, CA, United States, ⁹Department of Bioengineering, Stanford University, Stanford, CA, United States

Blip-up and Blip-down EPI (BUDA) is a rapid, distortion-free imaging method for diffusion-imaging and quantitative-imaging. Recently, we developed BUDA-circular-EPI (BUDA-cEPI) to shorten the readout-train and reduce T2* blurring for high-resolution applications. In this work, we further improve encoding efficiency of BUDA-cEPI by leveraging partial-Fourier in both phase-encode and readout directions, where complimentary conjugate k-space information from the blip-up and blip-down EPI-shots and S-LORAKS constraint are used to effectively fill-out missing k-space. While effective, S-LORAKS is computationally expensive. To enable clinical deployment, we also proposed a machine-learning reconstruction derived from RUN-UP (unrolled K-I network) that accelerates reconstruction by >300x.

Thai Akasaka¹, Koji Fujimoto², Yasutaka Fushimi³, Dinh Ha Duy Thuy¹, Atsushi Shima¹, Nobukatsu Sawamoto⁴, Yuji Nakamoto³, Tadashi Isa¹, and Tomohisa Okada^1
1Human Brain Research Center, Kyoto University, Kyoto, Japan, ²Department of Real World Data Research and Development, Kyoto University, Kyoto, Japan, ³Diagnostic Imaging and Nuclear Medicine, Kyoto University, Kyoto, Japan, ⁴Department of Human Health Sciences, Kyoto University, Kyoto, Japan

The effect of transmit field (B1⁺) inhomogeneity at 7T remains even after correction using a B1⁺-map. By a deep learning approach using pix2pix, a neural network was trained to generate 3T-like anatomical images from 7T MP2RAGE images (dataset 1: T1WI and T1-map after B1⁺ correction and dataset 2: Inversion time [INV]1, INV2 and B1⁺-map). When the HCP anatomical pipeline was applied and compared, low regressions of original 7T data to 3T data were largely improved by using generated images by  pix2pix, especially for dataset 2. 

Yuxiang Dai¹, Wei Tang¹, Ying-Hua Chu², Chengyan Wang³, and He Wang^1,3
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, ²Siemens Healthineers Ltd., Shanghai, China, ³Human Phenome Institute, Fudan University, Shanghai, China

Existing methods often fail to capture sufficient anatomical details which lead to unsatisfactory 7T MRI predictions, especially for 3D prediction. We proposed a 3D prediction model which introduces high frequency information learned from 7T images into generative adversarial network. Specifically, the prediction model can effectively produce 7T-like images with sharper edges, better contrast and higher SNR than 3T images.  

Jing Ren¹, Yonglan He¹, Chong Liu¹, Shifei Liu ¹, Jinxia Zhu², Marcel Dominik Nickel³, Zhengyu Jin¹, and Huadan Xue^1
1Peking Union Medical College Hospital, Peking Union Medical College and Chinese Academy of Medical Sciences, Beijing, China, ²MR collaboration, Siemens Healthineers Ltd., Beijing, China, ³MR Application Predevelopment, Siemens Healthcare GmbH, Erlangen, Germany

Novel deep learning (DL) reconstruction methods may accelerate female pelvis MRI protocols keeping high image quality. The value of a novel DL reconstruction of T2-weighted (T2_DLR) turbo spin-echo (TSE) sequences for female pelvis MRI in three orthogonal planes was evaluated. We evaluated examination times, image quality, and lesion conspicuity of benign uterine disease. The T2_DLR quantitative parameters remained similar or were significantly improved compared with that of standard T2 TSE (T2_S), allowing for a 62.7% reduction in acquisition times. Applying this novel T2_DLR sequence achieved better image quality and shorter acquisition time than T2_S.

Lijun An^1,2,3, Pansheng Chen^1,2,3, Jianzhong Chen^1,2,3, Christopher Chen⁴, Juan Helen Zhou², and B.T. Thomas Yeo^1,2,3,5,6
1Department of Electrical and Computer Engineering, National University of Singapore, Singapore, Singapore, ²Centre for Sleep and Cognition (CSC) & Centre for Translational Magnetic Resonance Research (TMR), National University of Singapore, Singapore, Singapore, ³N.1 Institute for Health & Institute for Digital Medicine (WisDM), National University of Singapore, Singapore, Singapore, ⁴Department of Pharmacology, National University of Singapore, Singapore, Singapore, ⁵NUS Graduate School for Integrative Sciences and Engineering, National University of Singapore, Singapore, Singapore, ⁶Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA, United States

We propose a flexible application-specific harmonization framework utilizing downstream application performance to regularize the harmonization procedure. Our approach can be integrated with various deep learning models. Here, we apply our approach to the recently proposed conditional variational autoencoder (cVAE) harmonization model. Three datasets (ADNI, N=1735; AIBL, N=495; MACC, N=557) collected from three different continents were used for evaluation. Our results suggest our approach (AppcVAE) compares favorably with ComBat (named for “combating batch effects when combining batches”) and cVAE for improving downstream application performance.

Shao Chieh Lin^1,2, Chun-Jung Juan^2,3,4, Ya-Hui Li², Ming-Ting Tsai², Chang-Hsien Liu², Hsu-Hsia Peng⁵, Teng-Yi Huang⁶, Yi-Jui Liu⁷, and Chia-Ching Chang^2,8
1Ph.D. program in Electrical and Communication Engineering, Feng Chia University, Taichung, Taiwan, ²Department of Medical Imaging, China Medical University Hsinchu Hospital, Hsinchu, Taiwan, ³Department of Radiology, School of Medicine, College of Medicine, China Medical University, Taichung, Taiwan, ⁴Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan, ⁵Department of Biomedical Engineering and Environmental Sciences, National Tsing Hua University, Hsinchu, Taiwan, ⁶Department of Electrical Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan, ⁷Department of Automatic Control Engineering, Feng Chia University, Taichung, Taiwan, ⁸Department of Management Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan

Inter-observer agreement is commonly used to evaluate the consistency of clinical diagnosis for two or more doctors. However, it is seldom to use to evaluate the consistency of clinical diagnosis for two or more deep learning models. In this study, four deep learning models for segmentation of stroke lesion were trained using GTs defined by two neuroradiologists with two ADC thresholds. We found the addition of an ADC threshold (0.6 × 10^-3 mm²/s) helps eliminate inter-observer variation and achieve best segmentation performance. The inter-observer in two deep learning models shows the more consistent degree compared with inter-observer in two neuroradiologists.

Haonan Xiao¹, Yat Lam Wong¹, Wen Li¹, Chenyang Liu¹, Shaohua Zhi¹, Weiwei Liu², Weihu Wang², Yibao Zhang², Hao Wu², Ho-Fun Victor Lee³, Lai-Yin Andy Cheung⁴, Hing-Chiu Charles Chang⁵, Tian Li¹, and Jing Cai^1
1Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China, ²Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education/Beijing), Department of Radiation Oncology, Beijing Cancer Hospital & Institute, Peking University Cancer Hospital & Institute, Beijing, China, ³Department of Clinical Oncology, The University of Hong Kong, Hong Kong, China, ⁴Department of Clinical Oncology, Queen Mary Hospital, Hong Kong, China, ⁵Department of Radiology, The University of Hong Kong, Hong Kong, China

We have developed and validated a deep learning-based real-time high-quality (HQ) multi-parametric (Mp) 4D-MRI technique. A dual-supervised downsampling-invariant deformable registration (D3R) model was trained on retrospectively downsampled 4D-MRI with 100 radial spokes in the k-space. The deformations obtained from the downsampled 4D-MRI were applied to 3D-MRI to reconstruct HQ Mp 4D-MRI. The D3R model provides accurate and stable registration performance at up to 500 times downsampling, and the HQ Mp 4D-MRI shows significantly improved quality with sub-voxel level motion accuracy. This technique provides HQ Mp 4D-MRI within 500 ms and holds great potential in online tumor tracking in MR-guided radiotherapy.

Christopher Man^1,2, Zheyuan Yi^1,2, Vick Lau^1,2, Jiahao Hu^1,2, Yujiao Zhao^1,2, Linfang Xiao^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 SAR, China, ²Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong SAR, China

RAKI is recently proposed as a deep learning version of GRAPPA, which trains on auto-calibration signal (ACS) to estimate the missing k-space data. However, RAKI requires a larger amount of ACS for training and reconstruction due to its multiple convolutions which resulting in lower effective acceleration. In this study, we propose to incorporate the virtual conjugate coil and enhanced non-linearity into the RAKI framework to improve the noise resilience and artifact removal at high effective acceleration. The results demonstrate that such strategy is effective and robust at high effective acceleration and in presence of pathological anomaly.

Vick Lau^1,2, Christopher Man^1,2, Yilong Liu^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

Supervised deep learning (DL) methods for MRI reconstruction is promising due to their improved reconstruction quality compared with traditional approaches. However, all current DL methods do not utilise anatomical features, a potentially useful prior, for regularising the network. This preliminary work presents a 3D CNN-based training framework that attempts to incorporate learning of anatomy prior to enhance model’s generalisation and its stability to perturbation. Preliminary results on single-channel HCP, unseen pathological HCP and IXI volumetric data (effective R=16) suggest its potential capability for achieving high acceleration while being robust against unseen anomalous data and data acquired from different MRI systems.

Rameesha Khawaja¹, Amna Ammar¹, Madiha Arshad¹, Faisal Najeeb¹, and Hammad Omer^1
1MIPRG Research Group, ECE Department, Comsats University, Islamabad, Pakistan

Reconstruction of cine Cardiac MRI (CMRI) is an active research area with room for improvement in motion detection (particularly irregular cardiac motion) and modeling in order to significantly enhance the quality of reconstructed images. Moreover, the reduction of scan time and image reconstruction time of cine CMRI is also a key aspect of today’s clinical requirement. We propose a dual domain cascade of neural networks intercalated with multi-coil data consistency layers for the reconstruction of cardiac MR images from Variable Density under-sampled data. The results show successful reconstruction results of our proposed method when compared with conventional compressed sensing reconstruction.

Jihun Kwon¹, Takashige Yoshida², Masami Yoneyama¹, Johannes M Peeters³, and Marc Van Cauteren^3
1Philips Japan, Tokyo, Japan, ²Tokyo Metropolitan Police Hospital, Tokyo, Japan, ³Philips Healthcare, Best, Netherlands

Time-of-flight MR angiography (TOF-MRA) is a non-contrast-enhanced imaging technique widely used to visualize intracranial vasculature. In this study, we investigated the use of ultra-thin slice (up to 0.4 mm) to improve the delineation of the cerebral arteries in TOF-MRA. To reduce the noise while preserving the image quality, Compressed SENSE AI (CS-AI) reconstruction was used. Our results showed that the improved noise reduction by CS-AI enabled better visualization of vessels, especially on the thinner slices compared to conventional Compressed-SENSE. The usefulness of CS-AI was also demonstrated in clinical cases with moyamoya disease and suspected aneurysm patients.