Veronica Ravano^1,2,3, Jean-François Démonet², Daniel Damian², Reto Meuli², Gian Franco Piredda^1,2,3, Till Huelnhagen^1,2,3, Bénédicte Maréchal^1,2,3, Jean-Philippe Thiran^2,3, Tobias Kober^1,2,3, and Jonas Richiardi^2
1Advanced Clinical Imaging Technology, Siemens Healthcare AG, Lausanne, Switzerland, ²Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ³LTS5, Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland

In radiology, the deployment of automated clinical decision support tools to new institutions is often hindered by inter-site data variability. In MRI, data heterogeneity often arises from differences in acquisition protocols. To overcome this issue, we propose a post-hoc harmonization technique based on generative adversarial networks (GAN). Seventy-seven patients suffering from dementia were scanned with two distinct T₁-weighted MP-RAGE protocols. We show that cross-protocol harmonization of brain images using a conditional GAN improves image similarity and reduces the variability of brain morphometry.

Kerstin Hammernik^1,2 and Thomas Küstner^3
1Lab for AI in Medicine, Technical University of Munich, Munich, Germany, ²Department of Computing, Imperial College London, London, United Kingdom, ³Medical Image And Data Analysis (MIDAS.lab), Department of Interventional and Diagnostic Radiology, University Hospital of Tuebingen, Tuebingen, Germany

Machine Learning (ML) methods have evolved tremendously during the last decade. A number of frameworks support the development of new ML methods. However, support for high-dimensional and complex-valued data processing is often limited. Therefore, we propose a Machine Enhanced Reconstruction Learning and Interpretation Networks (MERLIN) framework that seamlessly integrates with existing ML solutions such as Tensorflow/Keras and Pytorch and complements them by high-dimensional, complex-valued and MR-specific operators and layers. Furthermore, standard data processing pipelines in Python are provided in the context of MRI reconstruction.

Zahra Hosseini¹, Thorsten Feiweier², John Conklin³, Stephan Kannengiesser², Marcel Dominik Nickel², Min Lang³, Azadeh Tabari³, Augusto Lio Concalves Filho³, Wei-Ching Lo⁴, Maria Gabriela Figueiro Longo³, Michael Lev³, Pamela Schaefer³, Otto Rapalino³, Susie Huang³, Stephen Cauley⁵, and Bryan Clifford^4
1MR R&D Collaboration, Siemens Medical Solutions USA, Atlanta, GA, United States, ²Siemens Healthcare GmbH, Erlangen, Germany, ³Department of Radiology, Massachusetts General Hospital, Boston, MA, United States, ⁴MR R&D Collaboration, Siemens Medical Solutions USA, Boston, MA, United States, ⁵Department of Radiology, A. A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States

The integration of deep learning priors into regularized CG-SENSE reconstructions enables high quality MR images to be generated from noisy, undersampled data. The regularization parameter in these methods can be tuned to control the level of denoising, allowing a network to generalize to novel SNR conditions without retraining. However, manual tuning of the regularization parameter can be time consuming. This work presents a data-driven method for automatic regularization selection using commonly acquired noise calibration data. Results indicate the method generalizes across clinically relevant imaging scenarios and provides diagnostically equivalent image quality to that obtained by manual parameter tuning.

Peter Dawood^1,2, Martin Blaimer³, Felix Breuer³, Paul R. Burd⁴, István Homolya^5,6, Peter M. Jakob¹, and Johannes Oberberger^2
1Department of Physics, University of Würzburg, Würzburg, Germany, ²Department of Internal Medicine I, University Hospital Würzburg, Würzburg, Germany, ³Magnetic Resonance and X-ray Imaging Department, Fraunhofer IIS, Fraunhofer Institute for Integrated Circuits IIS, Division Development Center X-Ray Technology, Würzburg, Germany, ⁴Institute for Theoretical Physics and Astrophysics, University of Würzburg, Würzburg, Germany, ⁵Brain Imaging Centre, Research Centre for Natural Sciences, Budapest, Hungary, ⁶Institute of Nuclear Techniques, Budapest University of Technology and Economics, Budapest, Hungary

Recently, the Parallel Imaging method GRAPPA has been generalized by the deep-learning method RAKI, in which Convolutional Neural Networks are used for non-linear k-space interpolation. RAKI uses scan-specific training data, however, due to its increased parameter-space, its reconstruction quality may deteriorate given a limited training-data amount. We evaluate an approach that includes augmented training-data via an initial GRAPPA k-space reconstruction, and weights refinement by iterative training. Thereby, severe residual artefacts are suppressed in RAKI, while preserving its resilience against g-factor noise enhancement in GRAPPA for standard 2D imaging at medium accelerations, for strongly varying contrast between training- and interpolation-data, too. 

Tobias Senft¹, Reina Ayde¹, Najat Salameh¹, and Mathieu Sarracanie^1
1Department of Biomedical Engineering, University of Basel, Allschwil, Switzerland

Low magnetic field (LF) MRI is gaining popularity as a flexible and cost-effective complement to conventional MRI. However, LF-MRI suffers from a low signal-to-noise ratio per unit time which calls for signal averaging and hence prolonged acquisition times, challenging the clinical value of LF MRI.

In this study, we show that Deep Learning (DL) can reconstruct artifact-free heavily undersampled 2D LF MR images (34% sampling) with great success, both retrospectively and prospectively. Our results also highlight that a transfer learning approach combined with data augmentation improves the overall reconstruction performances, even when only small LF training datasets are available.

Sjoerd Ypma¹, Ivo Maatman¹, Matthan Caan², Dimitris Karkalousos², Marnix Maas¹, and Tom Scheenen^1
1Radboud UMC, Radiology and Nuclear Medicine, University of Nijmegen, Nijmegen, Netherlands, ²Amsterdam UMC, Biomedical Engineering and Physics, University of Amsterdam, Amsterdam, Netherlands

Undersampled k-space data reconstruction results in aliasing artifacts. Compressed sensing theory enables image reconstruction by using a priori knowledge in the form of regularization. Increasingly, Machine Learning methods are used to learn the regularization from data itself, but these methods can result in unstable reconstructions.

We propose a translation equivariant  single-layer neural network for reconstruction of radially measured k-space data. By exploiting translation symmetry, it can learn from randomly simulated data while still being applicable to in-vivo measurements. We tested robustness to small perturbations and reliability of the reconstruction of unexpected objects.

Robin Antony Birkeland Bugge¹, Jon Andre Ottesen², Elies Fuster³, Atle Bjørnerud², and Kyrre Eeg Emblem^1
1Department of Diagnostic Physics, Oslo University Hospital, Oslo, Norway, ²Department of Computational Radiology and Artificial Intelligence, Oslo University Hospital, Oslo, Norway, ³Biomedical Data Science Laboratory, Instituto Universitario de Tecnologías de la Información y Comunicaciones, Universitat Politècnica de València, Valencia, Spain

Problem summary: Brain MR elastography is associated with extended acquisition times, which is alleviated by reduced coverage or resolution. The aim of this project is to utilize deep learning to accelerate MRE. Methods: We employ an MRE for fully sampled acquisition. Undersampled data is simulated by masking phase-encoding steps. A cascaded reconstruction network is used to reconstruct the phase image from undersampled k-space. Results: There are subtle differences between the reconstructed and fully sampled phase images. We observe a non-significant difference for stiffness values in our preliminary results. Conclusions: The method shows promise for accelerating MR elastography data.

Aymen Ayaz¹, Kirsten Lukassen¹, Cristian Lorenz², Juergen Weese², and Marcel Breeuwer^1,3
1Biomedical Engineering Department, Eindhoven University of Technology, Eindhoven, Netherlands, ²Philips Research Laboratories, Hamburg, Germany, ³MR R&D – Clinical Science, Philips Healthcare, Best, Netherlands

We propose to simulate a large set of anatomically variable voxel-aligned and artifact-free brain MRI data at different resolutions to be used for training deep-learning based Super Resolution (SR) networks. To the best of our knowledge, no such efforts have been made in past regarding use of simulated data to train a SR network. We trained a SR network using such simulated data and tested the performance on real-world MRI data. The trained network could significantly sharpen low-resolution input MR images and clearly outperformed classic image interpolation methods. 

Paul Weiser¹, Stanislav Motyka², Wolfgang Bogner², and Georg Langs^1
1Computational Imaging Research Lab - Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Vienna, Austria, ²High Field MR Center - Department of Biomedical Imaging and Image‐Guided Therapy, Medical University of Vienna, Vienna, Austria

In this work a Graph Neural Network for the reconstruction of undersampled k-space data is proposed. The results were evaluated and compared against a state-of-the-art method. Overall, the results suggest that this Deep Learning-based approach for reconstruction is promising.

Jisu Yun¹, Seul Lee¹, Muyul Park¹, and Dong-Hyun Kim^1
1Department of Electrical and Electronic Engineering, Yonsei University, Seoul, Korea, Republic of

The mGRE(multi-echo gradient echo) sequence has previously been used for MWF(Myelin water fraction) imaging. Such mGRE observes signal decay by obtaining multiple echo signals, but scan time increases as more echoes are obtained. To solve this trade-off, we developed a deep learning model based on a LSTM model that can reduce scan time by predicting the later echoes using only the early echoes. Looking at the in vivo results and various performance test, our network has lower RMSE and higher PSNR than NLLS(Non-linear least squares), a conventional fitting algorithm.

R. Marc Lebel¹, Suryanarayanan Kaushik², Trevor Kolupar², Nate White³, Weidong Luo³, and Suchandrima Banerjee^4
1GE Healthcare, Calgary, AB, Canada, ²GE Healthcare, Waukesha, WI, United States, ³Cortechs, San Diego, CA, United States, ⁴GE Healthcare, Menlo Park, CA, United States

This work evaluates brain volumetry measurements obtained from T1w images with fast (PI) and rapid (PI+CS) protocols (< 3 mins) and reconstructed with AIR Recon DL 3D, a deep learning-based reconstruction to reduce noise and ringing. Images were segmented using FDA-cleared software NeuroQuantÒ for quantitative volumetric analysis. Segmentation was successful in all cases and key brain volumes are unchanged between fast and rapid protocols, and further between conventional and DL reconstructions. This work demonstrates that highly accelerated acquisitions and advanced reconstruction methods are suitable for segmentation and volumetric studies and can improve repeatability of measurements.

Marius Arvinte¹, Ajil Jalal¹, Giannis Daras², Eric Price², Alex Dimakis¹, and Jonathan I Tamir^1,3,4
1Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX, United States, ²Computer Science, 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, ⁴Diagnostic Medicine, Dell Medical School, The University of Texas at Austin, Austin, TX, United States

This work deals with the problem of few-shot adaptation in data-driven MRI reconstruction, where models must efficiently adapt to new distributions. We introduce score-based models for MRI reconstruction and an algorithm for adjusting inference parameters (step size, noise level and stopping point), investigate the impact of these parameters on reconstruction performance, and demonstrate average gains of at least 2 dB in PSNR across a range of acceleration values, all while using a pretrained model that was trained for brain MRI and fine-tuned using only a single fully-sampled 2D knee scan from the fastMRI database.