Florian Birk¹, Felix Glang¹, Christoph Birkl², Klaus Scheffler^1,3, and Rahel Heule^1,3
1High Field Magnetic Resonance, Max Planck Institute for Biological Cybernetics, Tübingen, Germany, ²Department of Neuroradiology, Medical University of Innsbruck, Innsbruck, Austria, ³Department of Biomedical Magnetic Resonance, University of Tübingen, Tübingen, Germany

Asymmetries in the balanced SSFP frequency profile are known to reflect information about intravoxel tissue microenvironment with strong sensitivity to white matter fiber tract orientation. Phase-cycled bSSFP has demonstrated potential for multi-parametric quantification of relaxation times, static and transmit field inhomogeneity, or conductivity, but has not yet been investigated for diffusion quantification. Therefore, a neural network approach is suggested, which learns a model for voxelwise quantification of diffusion metrics from bSSFP profiles. Not only the feasibility for robust predictions of mean diffusivity (MD) and fractional anisotropy (FA) is shown, but also potential to estimate the principal diffusion eigenvector.

Philip Müller¹, Vladimir Golkov¹, Valentina Tomassini², and Daniel Cremers^1
1Computer Vision Group, Technical University of Munich, Munich, Germany, ²D’Annunzio University, Chieti–Pescara, Italy

Convolutional networks are successful, but have recently been outperformed by new neural networks that are equivariant under rotations and translations. These new networks do not struggle with learning each possible orientation of each image feature separately. So far, they have been proposed for 2D and 3D data. Here we generalize them to 6D diffusion MRI data, ensuring joint equivariance under 3D roto-translations in image space and the matching rotations in q-space, as dictated by the image formation. We validate our method on multiple-sclerosis lesion segmentation. Our proposed neural networks yield better results and require less training data.

Radhika Madhavan¹, Jaemin Shin², Nastaren Abad¹, Luca Marinelli¹, J Kevin DeMarco³, Robert Y Shih³, Vincent B Ho³, Suchandrima Banerjee⁴, and Thomas K Foo^1
1GE Global Research, Niskayuna, NY, United States, ²GE Healthcare, New York, NY, United States, ³Walter Reed National Military Medical Center and Uniformed Services University of the Health Sciences, Bethesda, MD, United States, ⁴GE Healthcare, Menlo Park, CA, United States

Structure-preserved denoising of MRI images is a critical step in medical image analysis. This is particularly critical in diffusion MRI where higher spatial and angular resolutions required to map tissue microstructure in low SNR (especially at higher b-values) situations, if longer acquisition times are not used. Denoising using deep convolutional neural networks (DCNN) can reduce noise without requiring extensive averaging, enabling shorter scan times and high image quality, especially in the resulting tensor-derived maps. Preliminary results using DCNN based denoising on multi-shell diffusion data demonstrates improved image quality and reduced noise, without compromising on structural integrity and tensor derived metrics.

Seema S Bhat¹, Pavan Poojar², Hanumantharaju M C³, and Sairam Geethanath^1,2
1Medical Imaging Research center, Dayananda sagar college of engineering, Bangalore, India, ²Columbia University in the City of New York, Newyork, NY, United States, ³Department of Electronic and Communications, BMS Institute of Technology and Management, Bangalore, India

Deep learning based denoising can improve signal-to-noise ratio in high b-value diffusion imaging. Denoising CNN (DnCNN) model is trained with clean and simulated noisy patches of b=2000 s/mm² DWI images from Openneuro database. We applied DnCNN to simulated noisy DWI and prospective DWI at b=2000 s/mm².Unsharp masking used during testing to emphasize medium contrast details. Peak signal-to-noise ratio(>32dB) for simulated noisy DWI and image entropy(>7.17) for prospective DWI obtained after denoising. This denoising can be leveraged to shorten acquisition time by reducing the number of signal averages or increase resolution in through plane by acquiring with smaller slice thickness.

Brian McCrindle^1,2, Nicholas Simard^1,2, Ethan Samson^2,3, Ethan Danielli ^2,3, Thomas E. Doyle^1,3,4, and Michael D. Noseworthy^1,2,3
1Electrical and Computer Engineering, McMaster University, Hamilton, ON, Canada, ²Imaging Research Center, St. Joseph's Healthcare, Hamilton, ON, Canada, ³School of Biomedical Engineering, McMaster University, Hamilton, ON, Canada, ⁴Vector Institute, Toronto, ON, Canada

Patients who experience a mild traumatic brain injury often suffer from microstructural white matter damage that even radiologists are unable to detect. By employing diffusion tensor imaging and a deep 2D-UNet ensemble network, we developed an image processing pipeline capable of detecting and segmenting damaged white matter regions. We show that ensemble networks are more reliable compared to any single model over the prediction threshold range under test-time-augmentation.

Hayato Nozaki^1,2, Yasuhiko Tachibana³, Yujiro Otsuka⁴, Wataru Uchida^1,2, Yuya Saito¹, Koji Kamagata¹, and Shigeki Aoki^1
1Department of Radiology, Graduate School of Medicine, Juntendo University, Tokyo, Japan, ²Graduate School of Human Health Sciences, Tokyo Metropolitan University, Tokyo, Japan, ³Department of Molecular Imaging and Theranostics National Institute of Radiological Sciences National Institutes for Quantum and Radiological Science and Technology, Chiba, Japan, ⁴Miliman, Tokyo, Japan

Deep learning-based noise reduction technique for DWI contains a risk of outputting values that are greatly deviating from what it should be because of the instability problem of deep learning. The neural network model was designed in this study to suppress this risk which can fix the generated value for each pixel within the range of values of neighboring pixels in the original image. The results of the volunteer study suggested that the proposed method has potential to provide effective denoising beside suppressing the instability risk.

Yuhao Yan^1,2 and Zheng Chang^1,2
1Medical Physics Graduate Program, Duke University, Durham, NC, United States, ²Department of Radiation Oncology, Duke University, Durham, NC, United States

This work focused on accelerating DTI using deep learning methods. Three neural networks including U-net, PD-net and Cascade-net were investigated on reconstructing DTI images, ADC maps and FA maps from Cartesian under-sampled k-space data. The results indicated that Cascade-net out-performed the other two networks, obtaining comparable image quality as compared with the reference reconstructed from full k-space data. In summary, neural networks can be used to accelerate DTI while maintaining image quality.

Ali Bilgin^1,2,3,4, Loi Do¹, Phillip A Martin², Ethan Lockhart⁴, Adam S Bernstein¹, Chidi Ugonna¹, Laurel Dieckhaus¹, Courtney Comrie¹, Elizabeth B Hutchinson¹, Nan-Kuei Chen¹, Gene E Alexander^5,6, Carol A Barnes^5,7,8, and Theodore P Trouard^1,3,5
1Biomedical Engineering, University of Arizona, Tucson, AZ, United States, ²Electrical and Computer Engineering, University of Arizona, Tucson, AZ, United States, ³Medical Imaging, University of Arizona, Tucson, AZ, United States, ⁴Program in Applied Mathematics, University of Arizona, Tucson, AZ, United States, ⁵Evelyn F. McKnight Brain Institute, University of Arizona, Tucson, AZ, United States, ⁶Departments of Psychology and Psychiatry, University of Arizona, Tucson, AZ, United States, ⁷Division of Neural System, Memory & Aging, University of Arizona, Tucson, AZ, United States, ⁸Departments of Psychology, Neurology and Neuroscience, University of Arizona, Tucson, AZ, United States

The aim of this work is to accelerate analysis of diffusion weighted MRI (dMRI) of the rat brain using deep learning. The proposed approach allows prediction of unacquired diffusion-weighted images (DWIs) from a small set of acquired DWIs. By combining the acquired and predicted DWIs, accurate and reliable diffusion tensor metrics can be obtained with up to ten-fold reduction in scan time.

Alistair Lamb¹, Anna Barnes², Stuart A Taylor², and Hui Zhang^3
1Department of Medical Phyics and Biomedical Engineering, University College London, London, United Kingdom, ²Centre for Medical Imaging, University College London, London, United Kingdom, ³Centre for Medical Image Computing, University College London, London, United Kingdom

Despite its potential as an imaging biomarker in assessing tumor response to therapy, use of apparent diffusion coefficient (ADC) as a quantitative endpoint is not routine in clinical practice. One factor that limits the usefulness of ADC is the presence of artifacts in the constituent diffusion-weighted imaging (DWI) data. In this study, we propose a supervised deep-learning approach to detect the presence of Nyquist ghosts in axial DWI slices of the abdomen, achieving a test accuracy of 81.5%. The detection and removal of these artifacts could help improve the reproducibility of quantitative ADC measurements.

Qiyuan Tian^1,2, Ziyu Li³, Qiuyun Fan^1,2, Chanon Ngamsombat¹, Yuxin Hu⁴, Congyu Liao^1,2, Fuyixue Wang^1,2, Kawin Setsompop^1,2, Jonathan R Polimeni^1,2, Berkin Bilgic^1,2, and Susie Y Huang^1,2
1Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Boston, MA, United States, ²Harvard Medical School, Boston, MA, United States, ³Department of Biomedical Engineering, Tsinghua University, Beijing, China, ⁴Department of Electrical Engineering, Stanford University, Stanford, CA, United States

High-resolution diffusion tensor imaging (DTI) is beneficial for probing tissue microstructure in fine neuroanatomical structures, but long scan times and limited signal-to-noise ratio pose significant barriers to acquiring DTI at sub-millimeter resolution. To address this challenge, we propose a deep learning-based super-resolution method entitled “SRDTI” to synthesize high-resolution diffusion-weighted images (DWIs) from low-resolution DWIs. SRDTI employs a deep convolutional neural network (CNN), residual learning and multi-contrast imaging, and generates high-quality results with rich textural details and microstructural information, which are more similar to high-resolution ground truth than those from trilinear and cubic spline interpolation.

Jonathan Weine¹, Robbert J. H. van Gorkum¹, Christian T. Stoeck¹, Valery Vishnevskiy¹, Thomas Joyce¹, and Sebastian Kozerke^1
1Institute for Biomedical Engineering, University and ETH Zurich, Zürich, Switzerland

Cardiac DTI provides invaluable information about the state of myocardial microstructure. Motion and systematic signal variations of the imaging process influence the tensor inference. Image registration prior to tensor fitting with an LSQ estimator is the common data processing approach. The feasibility of training a neural network with simulated data modelling tensors and slice misalignment due to free breathing for inference of diffusion tensors from free-breathing in vivo data is investigated. Evaluation on simulated test data demonstrates feasibility of the training process. Application to in vivo data shows promising results of the CNN especially at myocardial borders.

Emmanuelle Weber¹, Christoph Leuze¹, Daniel A. N. Barbosa¹, Gustavo Chau Loo Kung¹, Kalanit Grill-Spector¹, and Jennifer A. McNab^1
1Stanford, Stanford, CA, United States

We demonstrate the feasibility of machine learning direct prediction of tissue microstructure from raw diffusion MRI data. We test the approach by attempting to predict the main fiber orientation as this metric is well-understood and easily extracted using a diffusion tensor model. We present results on both simulated data and a matched dMRI-3D histology dataset.

Tobias Goodwin-Allcock¹, Ting Gong¹, Robert Gray², Parashkev Nachev², and Hui Zhang^1
1Centre for Medical Image Computing (CMIC), UCL, London, United Kingdom, ²High-Dimensional Neurology, University College London Queen Square Institute of Neurology, London, United Kingdom

We present Patch-CNN-DTI, a deep-learning method to estimate diffusion tensors (DT) accurately from only 6 diffusion-weighted images. Early voxel-wise deep-learning methods can only estimate scalar measures of DT. Later work shows DT can be estimated using image-wise methods based on convolutional neural networks (CNN), but they require large training cohort. Patch-CNN-DTI can estimate DT with only one training subject, by pooling information from local neighbourhood of a voxel similar to the CNN but at a much smaller scale to minimise training data requirements. Results show it outperforms conventional model fitting with twice the number of diffusion directions.

David C Alsop^1,2, Manuel Taso^1,2, and Arnaud Guidon^3
1Radiology, Beth Israel Deaconess Medical Center, Boston, MA, United States, ²Radiology, Harvard Medical School, Boston, MA, United States, ³Global MR applications and workflow, GE Healthcare, Boston, MA, United States

STEAM diffusion offers longer diffusion times, reduced gradient demands and other advantages for diffusion imaging but has suffered from long acquisition times. We propose a non-selective preparation of a conventional gradient echo echoplanar sequence that enables the rapid acquisition of many slices. Initial application to the brain readily enabled up to 32 slice acquisition within a single TR.

Arun Venkataraman¹, Benjamin Risk², Deqian Qiu^3,4, Jianhui Zhong^1,5, and Zhengwu Zhang^6
1Physics and Astronomy, University of Rochester, Rochester, NY, United States, ²Biostatistics and Bioinformatics, Emory University, Atlanta, GA, United States, ³Radiology and Imaging Sciences, Emory University School of Medicine, Atlanta, GA, United States, ⁴Biomedical Engineering, Emory University, Atlanta, GA, United States, ⁵Imaging Sciences, University of Rochester Medical Center, Rochester, NY, United States, ⁶Biostatistics and Computational Biology, University of Rochester Medical Center, Rochester, NY, United States

In this study, we sought to better understand the effect of multiband and phase acceleration on diffusion image reconstruction. As part of an on-going study on SMS in dMRI, 10 test-retest scans from 5 young, healthy subjects were scanned. We found that g-factor, reflecting noise amplification, increased with higher acceleration factors, the g-factor was also higher in frontal regions compared to occipital regions. We also found that the DTI fitting performed worse where we saw increased g-factor. Finally, we found that streamlines were stable as a function of acceleration in the occipital areas, but not frontal areas.

Li Guo^1,2,3, Lyu Jian^2,3, Yingjie Mei⁴, Mingyong Gao¹, Yanqiu Feng^2,3, and Xinyuan Zhang^2,3,5
1Department of MRI, The First People’s Hospital of Foshan (Affiliated Foshan Hospital of Sun Yat-sen University), Foshan, China, ²School of Biomedical Engineering, Southern Medical University, Guangzhou, China, ³Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, Guangzhou, China, ⁴Philips Healthcare, Guangzhou, China, ⁵Guangdong-Hong Kong-Macao Greater Bay Area Center for Brain Science and Brain-Inspired Intelligence, Guangzhou, China

Accurate tensor estimation for DKI is usually challenged by noise. The noncentral Chi distribution noise would introduce bias in the estimated DKI tensors. Although several noise-corrected models are statistically unbiased, the DKI tensors generated by these estimators have large variances. In addition, severe noise easily causes the estimated kurtosis values outside a physically acceptable range. The goal of this work is to propose a unified framework that integrates multiple prior information including nonlocal structural self-similarity (NSS), local spatial smoothness (LSS), physical relevance (PR) of DKI model, and noise characteristic of magnitude diffusion images for improved DKI tensor estimation.

Julian Rauch^1,2, Dominik Ludwig^1,2, Frederik B. Laun³, and Tristan A. Kuder^1
1Division of Medical Physics in Radiology, German Cancer Research Center (DKFZ), Heidelberg, Germany, ²Faculty of Physics and Astronomy, Heidelberg University, Heidelberg, Germany, ³Institute of Radiology, University Hospital Erlangen, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Erlangen, Germany

Apparent exchange rate (AXR) mapping can be used to non-invasively investigate water exchange between the intra- and extracellular compartment, which might yield insight into cell membrane permeability. To obtain the AXR properly with the filter exchange imaging (FEXI) approach, signal stability is crucial. We show that electrocardiogram (ECG) triggering can lead to a slight improvement of the signal stability for AXR measurements in the human brain. However, pulsation-induced fluctuations are not the main source of signal variations. Thus, it remains questionable if investing additional measurement time in triggering is justified.

Lyu Jian^1,2, Xinyuan Zhang^1,2,3, Yingjie Mei⁴, and Li Guo^1,2,5
1School of Biomedical Engineering, Southern Medical University, Guangzhou, China, ²Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, Guangzhou, China, ³Guangdong-Hong Kong-Macao Greater Bay Area Center for Brain Science and Brain-Inspired Intelligence, Guangzhou, China, ⁴Philips Healthcare, Guangzhou, China, ⁵Department of MRI, The First People’s Hospital of Foshan (Affiliated Foshan Hospital of Sun Yat-sen University), Foshan, China

Non-Gaussian intravoxel incoherent motion (NG-IVIM) has been proposed to simultaneously quantify the perfusion and non-Gaussian diffusion properties in tissues. However, accurate parameter estimation for NG-IVIM is usually challenged by noise. The noncentral χ-distribution noise would introduce bias in the estimated NG-IVIM parameters. In addition, severe noise easily causes the estimated parameter values have large variance. To improve the accuracy and precision of parameter estimation for NG-IVIM, we propose to use an unbiased vector non-local means (UVNLM) filter to denoise and correct the noise bias before NG-IVIM model fitting.

Ratthaporn Boonsuth¹, Marco Battiston¹, Francesco Grussu^1,2, Marios Yiannakas¹, Torben Schneider³, Rebecca Samson¹, Ferran Prados^1,4,5, and Claudia A. M. Gandini Wheeler-Kingshott^1,6,7
1NMR research Unit, Queen Square MS Centre, Department of Neuroinflammation, UCL Queen Square Institute of Neurology, London, United Kingdom, ²Radiomics Group, Vall d’Hebron Institute of Oncology, Vall d’Hebron Barcelona Hospital Campus, Barcelona, Spain, ³Philips Healthcare, Guildford, Surrey, United Kingdom, ⁴Centre for Medical Image Computing, Medical Physics and Biomedical Engineering, University College London, London, United Kingdom, ⁵E-Heath Centre, Universitat Oberta de Catalunya, Barcelona, Spain, ⁶Department of Brain & Behavioural Sciences, University of Pavia, Pavia, Italy, ⁷Department of Brain Connectivity Centre Research Department, IRCCS Mondino Foundation, Pavia, Italy

Major MRI scanner upgrades can potentially introduce systematic changes in quantitative MRI metrics and affect reliability and reproducibility of quantitative parameters. Advanced diffusion weighted (DW) imaging methods are a powerful diagnostic and research tool in the brain and in spinal cord, but it is unclear to what extent DW-derived metrics can be affected by upgrades. Here we assess the effect of scanner upgrades on state-of-the-art spinal cord DW techniques. We found that diffusion matrices such as FA, MD and MK remain stable following a major upgrade when measured in the whole spinal cord, white matter, and grey matter areas.

Jiyoon Yoo¹, Leevi Kerkelä¹, Patrick W. Hales¹, Kiran K. Seunarine¹, Iulius Dragonu², Enrico Kaden¹, and Christopher A. Clark^1
1UCL Great Ormond Street Institute of Child Health, London, United Kingdom, ²Siemens Healthcare Ltd, Frimley, United Kingdom

In this work, our protocol leverages the most recent advances in the b-tensor diffusion encoding, and also uses a zoomed imaging technique, to specifically scan the hippocampus. Our sequence is able to acquire high-resolution images which enable direct delineation of hippocampal structures and allows estimation of advanced diffusion metrics, as well as parameters that can be measured from a standard multi-shell diffusion data set. Therefore, our acquisition can be used as an all-in-one sequence for microstructural imaging of the hippocampus. We also defined a subsampled acquisition protocol that is feasible for a clinical setting.