Arjun D Desai^1,2, Andrew M Schmidt², Elka B Rubin², Christopher M Sandino¹, Marianne S Black², Valentina Mazzoli², Kathryn J Stevens², Robert Boutin², Christopher Ré³, Garry E Gold^2,4, Brian A Hargreaves^1,2,4, and Akshay S Chaudhari^2,5
1Electrical Engineering, Stanford University, Stanford, CA, United States, ²Radiology, Stanford University, Stanford, CA, United States, ³Computer Science, Stanford University, Stanford, CA, United States, ⁴Bioengineering, Stanford University, Stanford, CA, United States, ⁵Biomedical Data Science, Stanford University, Stanford, CA, United States

While deep-learning-based MRI reconstruction and image analysis methods have shown promise, few have been translated to clinical practice. This may be a result of (1) paucity of end-to-end datasets that enable comprehensive evaluation from reconstruction to analysis and (2) discordance between conventional validation metrics and clinically useful endpoints. Here, we present the Stanford Knee MRI with Multi-Task Evaluation (SKM-TEA), a dataset of 155 clinical quantitative 3D knee MRI scans with k-space data, DICOM images, and dense tissue segmentation and pathology annotations to facilitate clinically relevant, comprehensive benchmarking of the MRI workflow. Dataset, code, and trained baselines are available at https://github.com/StanfordMIMI/skm-tea. 

Frank Zijlstra^1,2 and Peter T. While^1,2
1Department of Radiology and Nuclear Medicine, St. Olav's University Hospital, Trondheim, Norway, ²Department of Circulation and Medical Imaging, NTNU - Norwegian University of Science and Technology, Trondheim, Norway

This study demonstrates the potential of using deep learning for generating raw data to augment raw datasets of limited size for use in deep-learning-based MR image reconstruction. Using an adversarial autoencoder architecture, variability in 10 T1-weighted raw datasets from the fastMRI database was learned and recombined into new, random raw data. We trained a deep-learning-based Compressed Sensing reconstruction using both conventional approaches with real raw data and compared the results to training with generated raw data. The reconstruction from generated raw data showed improved reconstruction results and better generalization to scans with slightly different echo times.

Wei Peng¹, Li Feng², Guoying Zhao¹, and Fang Liu^3
1Computer Science and Engineering, University of Oulu, Oulu, Finland, ²Diagnostic, Molecular and Interventional Radiology, Icahn School of Medicine at Mount Sinai, New York, NY, United States, ³Radiology, Harvard Medical School, Boston, MA, United States

This work proposes a novel optimization framework to learn k-space sampling trajectories using deep learning by considering it an Ordinary Differential Equation (ODE) problem that can be solved using neural ODE. Particularly, the sampling of k-space data is framed as a dynamic system, in which neural ODE is formulated to approximate the system with additional constraints on MRI physics. Experiments were conducted on different in-vivo datasets (e.g., brain and knee images) acquired with different sequences. Initial results have shown that our proposed method can generate better image quality in accelerated MRI than conventional undersampling schemes in Cartesian and non-Cartesian acquisitions.

Andreas Kofler¹, Tobias Schaeffter^1,2,3, and Christoph Kolbitsch^1,3
1Physikalisch-Technische Bundesanstalt, Braunschweig and Berlin, Germany, ²Department of Biomedical Engineering, Technische Universität Berlin, Berlin, Germany, ³School of Imaging Sciences and Biomedical Engineering, King's College London, London, United Kingdom

Iterative neural networks (INNs) currently define the state-of-the-art for image reconstruction methods. With these methods, the obtained regularizers are not only optimally adapted to the employed physical model but also tailored to the reconstruction method the network implicitly defines. However, comparing the performance of different INNs-based methods is often challenging because of the black-box character of neural networks. In this work we construct an example which highlights the importance of keeping the number of trainable parameters approximately fixed when comparing INNs-based methods. If this aspect is not taken into consideration, wrong conclusions could be drawn.
 

Hongjun An¹, Sooyeon Ji¹, Sehong Oh², Jiye Kim¹, Dongmyung Shin¹, and Jongho Lee^1
1Department of Electrical and Computer Engineering, Seoul National University, Seoul, Korea, Republic of, ²Division of Biomedical Engineering, Hankuk University of Foreign Studies, Gyeonggi-do, Korea, Republic of

A new pulse sequence design method that requires no prior knowledge of MR physics or training dataset is proposed. This method utilizes neural architecture search, generating an optimal sequence for given properties (e.g., T₂*, B₁⁺) and target objectives. It successfully discovered FLAIR-like and spin-echo-like sequences. Furthermore, a less-intuitive sequence using three RF pulses was created when targeted for spin-echo. In an in-vivo experiment, this new sequence had almost identical contrasts and SNR with spin-echo (our SNR: 297.1±70.1; spin-echo SNR: 295.8±78.2) while utilizing 90% of RF energy of the spin-echo, suggesting a potential of the automated sequence design.

Sophie Schauman¹, Siddharth Srinivasan Iyer^1,2, Mahmut Yurt¹, Xiaozhi Cao¹, Congyu Liao¹, Zheng Zhong¹, Guanhua Wang³, Greg Zaharchuk¹, Shreyas Vasanawala¹, and Kawin Setsompop^1
1Department of Radiology, Stanford University, Stanford, CA, United States, ²Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ³Deptartment of Biomedical Engineering, University of Michigan, Ann Arbor, MI, United States

MRI is a notoriously slow imaging method, but in recent years many technical developments have improved acquisition speed. However, many of these new fast sequences have not made their way to clinical practice. Barriers to uptake include long reconstruction times, diminished image quality, and non-standard contrast in the resulting images. Our objective is to translate a ~1 minute MRF sequence into clinical practice, providing five high quality images with common clinical contrasts at 1 mm isotropic-resolution, as well as quantitative T1, T2, and proton density maps, all within a 5 minute reconstruction pipeline that we have deployed clinically.