View Presentation Video

Danyal Bhutto^1,2, Bo Zhu², Jeremiah Zhe Liu^3,4, Neha Koonjoo^2,5, Bruce R Rosen^2,5, and Matthew S Rosen^2,5,6
1Biomedical Engineering, Boston University, Boston, MA, United States, ²Athinoula A. Martinos Center for Biomedical Imaging, Charlestown, MA, United States, ³Google, Mountain View, CA, United States, ⁴Biostatistics, Harvard University, Cambridge, MA, United States, ⁵Harvard Medical School, Boston, MA, United States, ⁶Physics, Harvard University, Boston, MA, United States

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

As deep learning approaches for image reconstruction become increasingly used in the radiological space, strategies to estimate reconstruction uncertainties become critically important to ensure images remain diagnostic. We estimate reconstruction uncertainty through calculation of the Local Lipschitz value, demonstrate a monotonic relationship between the Local Lipschitz and Mean Absolute Error, and show how a threshold can determine whether the deep learning technique was accurate or if an alternative technique should be employed. We also show how our technique can be used to identify out-of-distribution test images and outperforms baseline metrics, i.e. deep ensemble and Monte-Carlo dropout.    

View Presentation Video

Zehuan Zhang¹, Matej Genči¹, Hongxiang Fan¹, Wayne Luk¹, and Andreas Wetscherek^2
1Department of Computing, Imperial College London, London, United Kingdom, ²Joint Department of Physics, The Institute of Cancer Research, London, United Kingdom

Keywords: Machine Learning/Artificial Intelligence, Diffusion/other diffusion imaging techniques, Uncertainty estimation

We transformed the state-of-the art IVIM-NET for IVIM parameter fitting into a Bayesian Neural Network (BNN). BNNs can estimate uncertainty for quantitative MRI parameters, which is relevant for clinical decision making. We found that training on data with the highest SNR outperformed IVIM-BNNET models trained on matching SNR regarding parameter errors and uncertainties. A region with artificially increased noise could be identified from IVIM-BNNET's uncertainty output. Compared with traditional fitting, IVIM-BNNET achieved comparable accuracy, while being 21 times faster and providing less correlated parameter estimates. Monte-Carlo dropout rate 0.4 provided the best trade-off between low errors and low uncertainty.

Sirui Wu¹, WenXuan Chen¹, Yibing Chen², Zhongsen Li¹, and Xiaolei Song^1
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, Tsinghua University, Beijing 100084, China, Beijing, China, ²Xi’an Key Lab of Radiomics and Intelligent Perception, School of Information Sciences and Technology, Northwest University, Xi’an, Shaanxi 710069, China, Xi'an, China

Keywords: Machine Learning/Artificial Intelligence, CEST & MT, Super-resolution, Unrolled network, quantitation

CEST MRI suffers from the low resolution, leading to the disability to image small structures. We proposed a new optimization process adapting the CEST data form. Based on deep learning, our optimization flow is unfolded into an unrolled network for super resolution, termed as SC-Net. Besides, we introduced ROI-based normalization loss to guarantee quantitation accuracy. Results showed that SC-Net could reconstruct high-resolution image series and quantitative maps (PSNR = 44.27dB, CNR APTw = 41.79dB), when down sampling rate = 8. And the ROI-based normalization loss could calibrate errors. In conclusion, SC-Net has the potential to image small regions and acceleration. 

View Presentation Video

Meng Zhang¹, Zheng Ye¹, Bo Zhang², Miaoqi Zhang², and Zhenlin Li^1
1Department of Radiology, West China Hospital, Sichuan University, Chengdu, China, ²MR Research, GE Healthcare, Beijing, China

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

At present, the resolution of magnetic resonance image is still limited to the detection of pancreatic space occupying lesions. In this work, quantitative and qualitative analysis of pancreatic images of T1WI and DWI with built-in DL Recon was carried out. The results showed that the images obtained by the sequence with built-in DL Recon were better than the original images in terms of SNR, CNR and subjective scores. It is confirmed that deep learning reconstruction can improve image resolution and has potential in the detection of space occupying lesions in pancreas.

View Presentation Video

Congcong Liu¹, Zhuoxu Cui¹, Sen Jia¹, Zhilang Qiu², Xin Liu¹, Hairong Zheng¹, Dong Liang¹, and Haifeng Wang^1
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, ²Case Western Reserve University, Cleveland, OH, United States

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

The wave encoding model with virtual conjugate coil (Wave-VCC) extension can provide more powerful MRI-accelerated imaging performance. However, estimating the background phase covering the full frequency range is un-tractable by only acquiring the middle auto-calibration signals (ACS) line during reconstruction. Here, combining a neural network without training, a new method to generate more accurate background phase in Wave-VCC is proposed. Including ablation experiments and comparison, experiments were carried out to verify the feasibility and performance of the proposed methods, respectively.

View Presentation Video

Denis Prokopenko¹, Daniel Rueckert^2,3, and Joseph V. Hajnal^1
1Biomedical Engineering Department, School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom, ²Department of Informatics, Technical University of Munich, Munich, Germany, ³Department of Computing, Imperial College London, London, United Kingdom

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

Dynamic free-breathing fetal heart MRI requires high spatial and temporal resolution, which could be reconstructed by kt-SENSE from undersampled data guided by priors of the same anatomy. Doubled acquisition time and uncontrolled fetal motion between the 2 acquisitions affects the data quality for reconstruction. We explored an alternative deep learning approach using a 3D U-Net based model with time-averaged skip connection and data consistency. Assessment of the model set a baseline for prior preconstruction and underlines important pitfalls that will drive further improvements to achieve optimal reconstruction quality.

View Presentation Video

Ashmita Deb^1,2, Shu-Fu Shih^1,2, Zhaohuan Zhang^1,2, and Holden H Wu^1,2
1Department of Radiological Sciences, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, United States, ²Department of Bioengineering, University of California Los Angeles, Los Angeles, CA, United States

Keywords: Machine Learning/Artificial Intelligence, Data Analysis, Image Enhancement

Deep learning (DL) based image enhancement of undersampled 3D dual-echo steady-state knee MRI can achieve faster computation times compared to compressed sensing reconstruction. However, it is hard to interpret how DL models work. This introduces the risk of DL-enhanced images containing inaccuracies without the user’s knowledge and thus confounding diagnosis. This work aimed to calculate pixel-wise uncertainty maps for DL-enhanced images by incorporating Monte Carlo Dropout into a 2D UNET to estimate epistemic uncertainty. Analysis showed that the DL-enhanced images achieved good image quality and the spatial uncertainty maps reflected errors, compared to reference images.

View Presentation Video

Swetali Nimje^1,2, Ludovic de Rochefort¹, and Thierry Artières^2
1Aix Marseille University, CNRS, CRMBM, Marseille, France, ²Aix Marseille University, Ecole Centrale de Marseille, CNRS, LIS, Marseille, France

Keywords: Machine Learning/Artificial Intelligence, Quantitative Susceptibility mapping

MRI data is inherently complex-valued, the vast majority of deep learning frameworks do not yet support complex-valued data. Most reconstruction networks separate real and imaginary components into two separate real-valued channels, which may not be the most efficient way to represent complex numbers. Phase is essential for many MRI applications, including phase contrast velocity mapping and Quantitative Susceptibility Mapping (QSM) etc. We propose a new crRAKI, a scan-specific complex-valued residual convolutional neural network for 2D/3D MRI data for accelerating phase mapping and QSM. A comparison is made with GRAPPA and rRAKI for the accelerated reconstruction of MRI images.

View Presentation Video

Ramesh Paudyal¹, Akash Deelip Shah², Amaresha Shridhar Konar¹, Jaemin Shin³, Eve LoCastro¹, Nisha Bagchi⁴, Maggie Fung³, Suchandrima Banerjee⁵, Nancy Lee⁶, and Amita Shukla-Dave^1,2
1Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, United States, ²Radiology, Memorial Sloan Kettering Cancer Center, New York, NY, United States, ³GE Health Care, New York, NY, United States, ⁴Kimmel Medical College, Thomas Jefferson University, Philadelphia, PA, United States, ⁵GE Health Care, Menlo Park, CA, United States, ⁶Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, United States

Keywords: Machine Learning/Artificial Intelligence, Tumor

The head and neck (HN) region have complex anatomical structures that affect the image quality of diffusion-weighted MRI. Therefore deep learning (DL)-based Reconstruction (Recon) for DW-MRI could be a promising method that can help improve image sharpness and signal-to-noise ratio (SNR) without increasing signal averaging. The present study aimed to evaluate the performance of qualitative and quantitative multiple b-value DW-MRI powered by DL-based Recon for tumors in the HN region. The DL-based recon method improved the DW image quality and SNR compared to those without DL recon.

View Presentation Video

Nora Vogt¹, Karyna Isaieva¹, Jean-Sébastien Louis¹, Christian Weihsbach², Mattias Paul Heinrich², Freddy Odille^1,3, Pierre-André Vuissoz^1,3, Jacques Felblinger^1,3, and Julien Oster^1,3
1IADI, Université de Lorraine, INSERM U1254, Nancy, France, ²Institut für Medizinische Informatik, Universität zu Lübeck, Lübeck, Germany, ³CIC-IT, INSERM 1433, Université de Lorraine and CHRU Nancy, Nancy, France

Keywords: Machine Learning/Artificial Intelligence, Cardiovascular

Cardiac CINE MRI plays an essential role in cardiac diagnosis, but not all patients are eligible for 3D imaging, which is associated with long acquisition times. We propose a deep-learning super-resolution model to generate 3D CINE from multi-planar 2D real-time MRI using external signals for cardiac and respiratory motion estimation. The proposed neural field model is trained on a single subject and performs non-rigid motion compensation and implicit representation learning in an end-to-end manner. A preliminary study with three healthy volunteers demonstrates promising reconstruction performance and computation times compared to traditional registration-based approaches.

View Presentation Video

Guan Qiu Hong^1,2,3, William Morley², Matthew Wan², Yuan Tao Wei^1,2,3, Yang Su², and Hai-Ling Margaret Cheng^1,2,3
1Institute of Biomedical Engineering, University of Toronto, Toronto, ON, Canada, ²The Edward S. Rogers Sr. Department of Electrical and Computer Engineering, University of Toronto, Toronto, ON, Canada, ³Translational Biology & Engineering Program, Ted Rogers Centre for Heart Research, Toronto, ON, Canada

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

Deep learning architectures such as Convolutional Neural Networks (CNNs) are widely used alongside fixed k-space undersampling for reconstructing highly undersampled MRI k-space data. However, CNN-based architectures may perform sub-optimally due to their limited ability to capture long range dependencies, and fixed undersampling patterns may not be amenable to optimal reconstruction. We propose dual-domain (image and k-space), transformer-based reconstruction architectures paired with learning-based undersampling to reconstruct undersampled k-space data. Experimental results demonstrate improved reconstruction quality compared to CNN-based (UNet) architecture across 5x to 100x acceleration factors; dual domain outperformed single domain reconstructions.

View Presentation Video

Dilek Mirgun Yalcinkaya^1,2, Hazar Benan Unal¹, Subha Raman^3,4, Abolfazl Hashemi², Rohan Dharmakumar^3,4, and Behzad Sharif^1,3,4
1Laboratory for Translational Imaging of Microcirculation, Indiana University (IU) School of Medicine, Indianapolis, IN, United States, ²Elmore Family School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN, United States, ³Weldon School of Biomedical Engineering, Purdue University, West Lafayette, IN, United States, ⁴Krannert Cardiovascular Research Center, IU School of Medicine/IU Health Cardiovascular Institute, Indianapolis, IN, United States

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction, Plug-and-play, denoiser

In this work, we demonstrated that a deep-learning based denoiser trained on a limited dataset of first-pass myocardial perfusion MRI studies enables low-latency image reconstruction using the plug-and-play iterative reconstruction framework. Our proof-of-concept results suggest that the data-adapted denoiser resulted in superior performance versus a generic denoiser especially if there are constraints on the number of iterations (total computation time) which is the case in certain clinical settings specifically interventional MRI. Our findings also imply that radial sampling may be a more desirable data acquisition strategy for PnP-based image reconstruction in first-pass MP MRI studies.

View Presentation Video

Quan Dou¹, Zhixing Wang¹, Xue Feng¹, and Craig Meyer^1
1Biomedical Engineering, University of Virginia, Charlottesville, VA, United States

Keywords: Machine Learning/Artificial Intelligence, Machine Learning/Artificial Intelligence

Signal-to-noise ratio is crucial for MR image analysis. In this work, we proposed a non-blind complex-valued convolutional neural network for MRI denoising. Compared to existing denoising methods, the proposed method shows better performance on handling noise at different levels and on spatially variant noise.

View Presentation Video

Xinyang Lv¹, Zheng Ye¹, Miaoqi Zhang², Bo Zhang², and Zhenlin Li^1
1radiology department, West China Hospital，Sichuan University, Chengdu, China, ²MR Research, GE Healthcare, Beijing, China

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

Magnetic resonance imaging (MRI) is a useful tool to diagnose lumbar endplate inflammation. It is thus important to improve diagnostic accuracy by improving image quality. In this study, we compared signal-to-noise ratio (SNR), contrast noise ratio (CNR) and subjective scores between original images and deep learning reconstruction (DL Recon) images in 31 patients diagnosed with lumbar endplate inflammation. It was observed that the deep learning reconstructed images outperformed conventional images in terms of both subjective scores and objective values.

View Presentation Video

Jonas Kleineisel¹, Bernhard Petritsch¹, Thorsten A. Bley¹, Herbert Köstler¹, and Tobias Wech^1
1Department of Diagnostic and Interventional Radiology, University Hospital of Würzburg, Würzburg, Germany

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

Magnetic resonance cholangiopancreatography suffers from long examination times and artifacts originating from residual motion. To shorten the acquisition protocol, we acquired data at 12-fold undersampling and investigated a 3D Variational Network (VN) architecture for reconstruction. We compared a self-supervised training scheme and a supervised network trained on synthetic data. We find that the self-supervised method is only able to provide competitive reconstructions if the network is initialized with pre-trained weights, and even then does not offer superior performance over the supervised approach. For the presented exemplary data, the supervised VN showed comparable image quality as a reference Compressed Sensing model.

View Presentation Video

Yuqi Tan¹, Zheng Ye¹, Miaoqi Zhang², Bo Zhang², and Zhenlin Li^1
1West China Hospital of Sichuan University, Chengdu, China, ²GE Healthcare, MR Research, Beijing, China

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

Improving image resolution by denoising is an important research goal in MRI reconstruction. An emerging technique, deep learning reconstruction (DLR), has shown great potential in MRI denoising. In this study, we included 37 patients with pathologically diagnosed hepatic malignancy, and compared the image quality of DLR and original reconstruction regarding dual echo T1 weighted sequence, diffusion weighted imaging (DWI) and fat-suppressed T1 weighted gadolinium-enhancement. It was shown that DLR significantly improved the image quality by reducing background noise, thus making hepatic malignancy more conspicuous. 

View Presentation Video

Shohei Ouchi¹, Itona Fukatsu¹, Kazuki Yamato¹, and Satoshi Ito^1
1Utsunomiya University, Utsunomiya, Japan

Keywords: Machine Learning/Artificial Intelligence, Machine Learning/Artificial Intelligence

Complex-valued CNN based image reconstruction methods have been proposed to correspond to MR images with a spatial phase variation. However, using those CNN may lead to over-fitting because CNN layers for complex numbers are requires large number of parameters than real-valued CNN. We previously proposed a reconstruction method for complex-valued image using a real-valued DnCNN by introducing a symmetrical k-space under-sampling. In this study, we introduced this method to U-Net and ADMM-CSNet. Reconstruction experiments showed that a real-valued CNN has the possibility to have the same or better performance as a complex-valued CNN without perform complex calculations.

Wenxuan Chen¹, Sirui Wu¹, and Xiaolei Song^1
1Center for Biomedical Imaging Research, Department of Biomedical Engineering, Tsinghua University, Beijing, China

Keywords: Machine Learning/Artificial Intelligence, Machine Learning/Artificial Intelligence, Super-resolution

CEST is an important source of new contrast in MRI. However, it is time-consuming to obtain high-resolution CEST images. We proposed a deep learning based Two-Stage Super-Resolution (TSSR) model for CEST images. Compared with conventional SISR or MFSR models, TSSR model can effectively utilize the correlations among slices and those among saturation offsets for CEST images. We acquired brain CEST images on 14 volunteers using a 3T clinical MR scanner. Initial results suggested that the proposed TSSR model outperformed other methods for all the evaluation indicators. Our work showed the potential in reconstruction of high-resolution CEST images from low-resolution ones.

View Presentation Video

Gulfam Ahmed Saju¹, Zhiqiang Li², Reza Abiri³, Tianming Liu⁴, and Yuchou Chang^1
1Computer and Information Science, University of Massachusetts Dartmouth, North Dartmouth, MA, United States, ²Neuroradiology, Barrow Neurological Institute, Phoenix, AZ, United States, ³Electrical, Computer and Biomedical Engineering, University of Rhode Island, Kingston, RI, United States, ⁴Computer Science, University of Georgia, Athens, GA, United States

Keywords: Machine Learning/Artificial Intelligence, Parallel Imaging

Training data for MRI reconstruction are difficult to be acquired in clinical practice. In addition, machine learning or deep learning-based MRI reconstruction suffers the distribution shift problem between training data and testing data. Generalization error always exists, so reconstructed images are unstable. We proposed a joint estimation of coil sensitivity and image using the prior of an untrained neural network (UNN). Coil sensitivity map improvement gradually enhances the UNN prior and the image to be reconstructed in an iterative optimization process. The method outperforms other MRI reconstruction methods by suppressing noise and aliasing artifacts.

View Presentation Video

Junyu Wang¹ and Michael Salerno^1
1Cardiovascular Medicine, Stanford University, Stanford, CA, United States

Keywords: Machine Learning/Artificial Intelligence, Image Reconstruction

Keywords: artificial intelligence, image reconstruction, perfusion

 

Cardiac magnetic resonance first-pass contrast-enhanced myocardial perfusion imaging is valuable for evaluating coronary artery disease¹. 2D Cartesian perfusion imaging using compressed sensing (CS)-based reconstructions such as L1-SENSE² enables fast and high-resolution imaging, but whole-heart coverage cannot be achieved without simultaneous multi-slice (SMS) acquisitions and the CS-based iterative reconstruction takes ~30 minutes per slice. To address this, we have developed a deep learning-based motion-corrected rapid image reconstruction for high-resolution Cartesian perfusion imaging at 3 Tesla, for both 2D and SMS MB=2 acquisitions, which provides fast and high-quality motion-corrected reconstruction and makes rapid online reconstruction feasible.