Javier Bisbal^1,2,3, Julio Sotelo^1,3,4, Cristobal Arrieta^1,3, Pablo Irarrázabal^1,2,3, Marcelo Andia^1,3,5, Cristian Tejos^1,2,3, and Sergio Uribe^1,3,5
1Biomedical Imaging Center UC, Pontificia Universidad Catolica de Chile, Santiago, Chile, ²Electrical Engineering Department, School of Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile, ³ANID – Millennium Science Initiative Program – Millennium Nucleus for Cardiovascular Magnetic Resonance, Santiago, Chile, ⁴School of Biomedical Engineering, Universidad de Valparaíso, Valparaíso, Chile, ⁵Department of Radiology, School of Medicine, Pontificia Universidad Católica de Chile, Santiago, Chile

One major limitation on 4D flow MRI is the time-consuming and user-dependent post-processing. We developed an automated reinforced deep learning framework for plane planning in 4D flow data. This method sequentially updates plane parameters towards a target plane based on a continuous policy. A total of 83 4D flow MRI scans were considered, 41 for training, 14 for validation and 28 for test. Our method achieves good results in terms of angulation and distance error (9.21 ± 3.85 degrees and 3.72 ± 2.19 mm).

Kofi Deh¹, Nathaniel Kim¹, Guannan Zhang¹, Miloushev Vesselin¹, and Kayvan Keshari^1
1Memorial Sloan Kettering Cancer Center, New York, NY, United States

Hyperpolarized ¹³C spectroscopic images are acquired at a low spatial resolution, making it necessary to apply superresolution techniques to the metabolite maps prior to fusion with the proton anatomic image for visualization of metabolite biodistribution. In this work, we demonstrate great improvement in image quality for preclinical and clinical images when the metabolite map is upsampled by high spatial frequency transfer from magnetic resonance images of thermally polarized ¹³C phantoms to the invivo metabolic image.

David E J Waddington¹, Christopher Chiu¹, Nicholas Hindley^1,2, Neha Koonjoo², Tess Reynolds¹, Paul Liu¹, Bo Zhu², Chiara Paganelli³, Matthew S Rosen^2,4,5, and Paul J Keall^1
1ACRF Image X Institute, Faculty of Medicine and Health, The University of Sydney, Sydney, Australia, ²A. A. Martinos Center for Biomedical Imaging, Charlestown, MA, United States, ³Dipartimento di Elettronica, Informazione e Bioingegneria, Politecnico di Milano, Milan, Italy, ⁴Department of Physics, Harvard University, Cambridge, MA, United States, ⁵Harvard Medical School, Boston, MA, United States

Today’s MRI does not have the spatio-temporal resolution to image the anatomy of a patient in real-time. Therefore, novel solutions are required in MRI-guided radiotherapy to enable real-time adaptation of the treatment beam to optimally target the cancer and spare surrounding healthy tissue. Neural networks could solve this problem, however, there is a dearth of sufficiently large training data required to accurately model patient motion. Here, we use the YouTube-8M database to train the AUTOMAP network. We use a virtual dynamic lung tumour phantom to show that the generalized motion properties learned from YouTube lead to improved target tracking accuracy.

Patrick Salome^1,2,3,4, Francesco Sforazzini^1,2,3,4, Andreas Kudak^3,5,6, Matthias Dostal^3,5,6, Nina Bougatf^3,5,6, Jürgen Debus^3,4,5,7, Amir Abdollahi^1,3,4,5, and Maximilian Knoll^1,3,4,5
1CCU Translational Radiation Oncology, German Cancer Research Center (DKFZ), Heidelberg, Germany, ²Medical Faculty, Heidelberg University Hospital, Heidelberg, Germany, ³Heidelberg Ion-Beam Therapy Center (HIT), Heidelberg, Germany, ⁴German Cancer Consortium (DKTK) Core Center, Heidelberg, Germany, ⁵Radiation Oncology, Heidelberg University Hospital, Heidelberg, Germany, ⁶CCU Radiation Therapy, German Cancer Research Center (DKFZ), Heidelberg, Germany, ⁷National Center for Tumor Diseases (NCT), Heidelberg, Germany

MR-Class is a deep learning-based MR image classification tool that facilitates and speeds up the initialization of big data MR-based research studies by providing fast, robust, and quality-assured MR image classifications. It was observed in this study that corrupt and misleading DICOM metadata could lead to a misclassification of about 10%. Therefore,  in a field where independent datasets are frequently needed for study validations, MR-Class can eliminate the cumbrousness of data cohorts curation and sorting. This can greatly impact researchers interested in big data multiparametric MRI studies and thus contribute to the faster deployment of clinical artificial intelligence applications.

Andjela Dimitrijevic^1,2, Vincent Noblet³, and Benjamin De Leener^1,2,4
1NeuroPoly Lab, Institute of Biomedical Engineering, Polytechnique Montréal, Montréal, QC, Canada, ²Research Center, Ste-Justine Hospital University Centre, Montréal, QC, Canada, ³ICube-UMR 7357, Université de Strasbourg, CNRS, Strasbourg, France, ⁴Computer Engineering and Software Engineering, Polytechnique Montréal, Montréal, QC, Canada

Deep learning techniques have a potential in allowing fast deformable registration tasks. Studies around registration often focus on adult populations, even if there is a need for pediatric research where less data and studies are being produced. In this study, we compared three methods for intra-subject registration on publicly available Calgary Preschool dataset. Using the DeepReg framework, pre-registering with a rigid and affine transformation (proposed RigidAffineReg method) showed the least negative JD values and the highest Dice score (0.924±0.045). By achieving faster alignments, this tool for pediatric MRI scans could help proliferate larger scale population research in brain developmental studies.

David Lohr¹, Theresa Reiter^1,2, and Laura Schreiber^1
1Chair of Cellular and Molecular Imaging, Comprehensive Heart Failure Center (CHFC), Würzburg, Germany, ²Department of Internal Medicine I, University Hospital Würzburg, Würzburg, Germany

Diffusion weighted imaging has become a key imaging modality for the assessment of brain connectivity as well as the structural integrity, but the method is limited by low SNR and long scan times. In this study we demonstrate that AI models trained on a moderate number of publicly available 7T datasets (n=12) are able to enhance image resolutions in diffusion MRI up to 25 times. The applied NoGAN model performs well for smaller resolution enhancements (4- and 9-fold), but generates "hyper" realistic images for higher enhancements (16- and 25-fold). Models trained using perceptual loss seem to avoid this limitation.  

Pooya Ashtari^1,2, Berardino Barile^1,2, Dominique Sappey-Marinier², and Sabine Van Huffel^1
1Department of Electrical Engineering (ESAT), KU Leuven, Leuven, Belgium, ²CREATIS (CNRS UMR5220 & INSERM U1294), Université Claude-Bernard Lyon 1, Lyon, France

Automated segmentation of new multiple sclerosis (MS) lesions in MRI data is crucial for monitoring and quantifying MS progression. Manual delineation of such lesions is laborious and time-consuming since experts need to deal with 3D images and numerous small lesions. We propose a 3D encoder-decoder architecture with pre-activation blocks to segment new MS lesions in longitudinal FLAIR images. We also applied intensive data augmentation and deep supervision to mitigate the limited data and the class imbalance problem. The proposed model, called Pre-U-Net, achieved a Dice score of 0.62 and a sensitivity of 0.58 on the public challenge MSSEG-2 dataset.

Yajing Zhang¹, Xiangyu Xiong¹, and Chuanqi Sun^1
1MR Clinical Science, Philips Healthcare, Suzhou, China

Image synthesis methods based on deep learning has recently achieved success in reducing the dosage of gadolinium-based contrast agents (GBCAs). However, these methods cannot focus on the region of interest to synthesize realistic images. To address this issue, a mask guided attention generative adversarial network (MGA-GAN) was proposed to synthesize contrast enhanced T1-weight images from the multi-channel inputs. Qualitive and quantitative results indicate that the proposed MGA-GAN can improve the synthesized images with higher quality for details of brainstem glioma, compared with state-of-the-art methods.

Seb Harrevelt¹, Daan Reesink², Astrid Lier, van³, Richard Meijer³, Josien Pluim⁴, and Alexander Raaijmakers^1
1TU Eindhoven, Utrecht, Netherlands, ²Meander Medisch Centrum, Utrecht, Netherlands, ³UMC Utrecht, Utrecht, Netherlands, ⁴TU Eindhoven, Eindhoven, Netherlands

Prostate imaging at ultra-high fields is heavily affected by B1 field induced inhomogeneities. This not only results in unattractive images but it also might affect clinical diagnosis . To remedy this we developed a deep learning model that retrospectively corrects for the bias field. We applied this model to a clinical data set and demonstrated its performance in a qualitative manner. The results indicate that the model is able to drastically reduce the inhomogeneities in a variety of cases while the tissue contrast is generally maintained and the underlying anatomy has been successfully recovered.

Fatima Sattar¹, Sadia Ahsan¹, Fariha Aamir¹, Ibtisam Aslam^1,2, Iram Shahzadi^3,4, and Hammad Omer^1
1Medical Image Processing Research Group (MIPRG), Department of Electrical and Computer Engineering, COMSATS University, Islamabad, Pakistan, ²Service of Radiology, Geneva University Hospitals and Faculty of Medicine, Hospital University of Geneva and University of Geneva, Geneva, Switzerland, ³OncoRay – National Center for Radiation Research in Oncology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universität Dresden, Helmholtz-Zentrum Dresden –Rossendorf, Dresden, Germany, Dresden, Germany, ⁴German Cancer Research Center (DKFZ), Heidelberg, Germany, Dresden, Germany

Echo-planar imaging suffers from Nyquist ghost (i.e., N/2 ghost) artifacts because of poor system gradients and delays. Many conventional methods have been used in literature to remove N/2 artifacts in Diffusion Weighted Imaging (DWI) but often produce erroneous results. This paper presents a deep learning approach to eliminate the phase error of k-space for removing the Nyquist ghost artifacts in DWI. Experimental results show successful removal of the ghost artifacts with improved SNR and reconstruction quality with the proposed method.

Tobias Senft¹, Reina Ayde¹, Marco Fiorito¹, Najat Salameh¹, and Mathieu Sarracanie^1
1Department of Biomedical Engineering, University of Basel, Allschwil, Switzerland

Low magnetic field (LF) MRI is currently gaining momentum as a complementary, more flexible and cost-effective approach to MRI diagnosis. However, the impaired signal-to-noise ratio challenges its relevance for clinical applications. Recently, denoising of low SNR images using deep learning techniques has shown promising results for MRI applications. In this study, we assess the denoising performance of residual U-net architecture on different SNR levels of LF MRI data (0.1 T). The model performance has been evaluated on both simulated and acquired LF MRI datasets.

Cristian Crisosto^1,2, Andreas Voskrebenzev^1,2, Marcel Gutberlet^1,2, Filip Klimeš^1,2, Till Kaireit^1,2, Gesa Pöhler^1,2, Tawfik Alsady^1,2, Lea Behrendt^1,2, Robin Müller^1,2, Maximilian Zubke^1,2, Frank Wacker^1,2, and Jens Vogel-Claussen^1,2
1Institute of Diagnostic and Interventional Radiology, Medical School Hannover, Hannover, Germany, ²Biomedical Research in Endstage and Obstructive Lung Disease Hannover (BREATH), Member of the German Centre for Lung Research (DZL), Hannover, Germany, Hannover, Germany

Accurate fully automated lung segmentation is needed to facilitate Fourier-Decomposition employment-based techniques in clinical routine among different centers. However, the lung parenchyma segmentation remains challenging for convolutional neural networks (CNN) when consolidations are present. To improve training balanced augmentation (BA) and artificially-generated consolidations (AGC) were introduced. The proposed CNN was compared to conventional CNNs without BA and AGC using Sørensen-Dice coefficient (SDC) and Hausdorff coefficient (HD). The SDC / HD of the proposed model is significantly higher (p of 0.0001 and p of 0.0146 / p of 0.0009 and p of 0.0152) when compared to CNNs without BA and AGC. 

Yujian Diao^1,2,3, Catarina Tristão Pereira^2,4, Ting Yin², and Ileana Ozana Jelescu^2,5
1Laboratory of Functional and Metabolic Imaging, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland, ²CIBM Center for Biomedical Imaging, Lausanne, Switzerland, ³Animal Imaging and Technology, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland, ⁴Faculdade de Ciências da Universidade de Lisboa, Lisbon, Portugal, ⁵Department of Radiology, Lausanne University Hospital, Lausanne, Switzerland

Impaired brain glucose consumption is a possible trigger of Alzheimer’s disease (AD). Previous work revealed affected brain structure and function by insulin resistance in terms of functional connectivity and white matter microstructure in a rat model of AD. Here, functional and structural metrics were further used to classify Alzheimer’s from control rats using logistic regression. Our study highlights the MRI-derived biomarkers that best discriminate Alzheimer’s vs control rats early in the course of the disease, with potential translation to human AD.

Georgia Doumou^1,2, Hongxiang Lin^2,3, Sara Lorio¹, Lenka Vaculčiaková⁴, Kerrin J. Pine⁴, Nikolaus Weiskopf^4,5, Jonathan O'Muircheartaigh^6,7,8, Daniel C. Alexander², and David W. Carmichael ^1
1Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom, ²Centre for Medical Image Computing, Department of Computer Science, University College London, London, United Kingdom, ³Research Center for Healthcare Data Science, Zhejiang Lab, Hangzhou, China, ⁴Department of Neurophysics, Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig, Germany, ⁵Felix Bloch Institute for Solid State Physics, Faculty of Physics and Earth Sciences, Leipzig University, Leipzig, Germany, ⁶Department of Forensic & Neurodevelopmental Sciences, King's College London, London, United Kingdom, ⁷Centre for the Developing Brain, Department of Perinatal Imaging and Health, King's College London, London, United Kingdom, ⁸MRC Centre for Neurodevelopmental Disorders, King's College London, London, United Kingdom

High-resolution quantitative MRI using ultra-high field scanners (7T) could advance a range of research and clinical applications if limitations, both practical (e.g. long acquisitions) and technical (e.g. B1 non-uniformity), can be avoided. This could be achieved by using 7T information to enhance conventional field strength images. To test this approach, paired 3T-7T R1 maps were used to train a U-Net variant to enhance 3T R1 maps. Leave-one-out cross-validation, quantitative evaluation, as well as external validation on an external clinical dataset, demonstrated promising enhancement with visual and quantitative metrics more similar to 7T R1 maps than the 3T equivalents.

Aymen Ayaz¹, Ronald de Jong¹, Samaneh Abbasi-Sureshjani¹, Sina Amirrajab¹, 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 a method to synthesize 3D consistent brain MRI utilizing multiple 2D conditional SPatially-Adaptive (DE)normalization (SPADE) GANs to preserve the spatial information of patient-specific brain anatomy and a 3D CNN based network to improve the image consistency in all directions in a cost-efficient manner. Three individual 2D SPADE-GAN networks are trained across the axial, sagittal and coronal slice directions and their outputs are thereafter used further to train a CNN based 3D image restoration network to combine them into one 3D volume, using real MRI as a reference. The resulting predicted-synthesized 3D brain MRI is evaluated quantitatively and qualitatively.