Ying Liao¹, Santiago Coelho¹, Jelle Veraart¹, Els Fieremans¹, and Dmitry S. Novikov^1
1Radiology, NYU School of Medicine, New York, NY, United States

Maximum likelihood estimation is challenging in multicompartmental models due to the degeneracy of the optimization landscape. As a result, machine learning (ML) methods are often applied for parameter estimation, interpolating the mapping of measurements to model parameters. Such mapping can essentially depend on the training set (prior), decreasing the sensitivity to the measurements, and yielding artificially “clean” maps. Here we quantify the effect of the training set on the Standard Model of diffusion in white matter as function of signal-to-noise ratio, in simulations and in vivo.

Santiago Coelho¹, Steven Baete¹, Gregory Lemberskiy¹, Benjamin Ades-aron¹, Jelle Veraart¹, Dmitry S. Novikov¹, and Els Fieremans^1
1Radiology, NYU School of Medicine, New York, NY, United States

Robust parameter estimation of the Standard Model (SM) for diffusion in white matter has been elusive due to intrinsic model degeneracies and insufficient measurements. To design optimal scanner-specific protocols, we couple multidimensional protocol optimization for estimation of microstructural tissue properties in 15 minute acquisitions on clinical scanners with varying gradient performance. We show reproducible scan-rescan results and assess inter-scanner variability ranging from 1 to 8% depending on parameters. Results suggest that combining denoising, protocol optimization, and robust parameter estimation may enable quantitative microstructure mapping in clinical settings.

Tianjia Zhu^1,2, Minhui Ouyang¹, Nikou Lei³, David Wolk⁴, Paul Yushkevich⁵, and Hao Huang^1,5
1Department of Radiology, Children's Hospital of Philadelphia, Philadelphia, PA, United States, ²Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, United States, ³Department of Physics, University of Washington Seattle, Seattle, WA, United States, ⁴Department of Neurology, University of Pennsylvania, Philadelphia, PA, United States, ⁵Department of Radiology, University of Pennsylvania, Philadelphia, PA, United States

Diffusion MRI (dMRI) has ushered in a new era in which conventional brain cortical histological measures such as soma and neurite densities may be assessed noninvasively through advanced dMRI. However, analytical dMRI microstructural models are restricted by the model assumptions and lack of validation from quantitative histology data. Individual dMRI parameters characterize only limited microstructural information. By leveraging a variety of dMRI-based parameters delineating cortical microstructure from multiple aspects, we established a machine learning based method accurately estimating cortical soma and neurite densities in the cortex, paving the way for data-driven noninvasive virtual histology for potential applications to Alzheimer’s diseases.  

Andrada Ianus¹, Francisca F. Fernandes¹, Joana Carvalho¹, Cristina Chavarrias¹, Marco Palombo², and Noam Shemesh^1
1Champalimaud Centre for the Unknown, Lisbon, Portugal, ²University College London, London, United Kingdom

Measuring micro-architectural features involving cell body morphologies is emerging as a frontier of diffusion MRI. This work aimed to map the apparent soma size and density and neurite density via the SANDI methodology in the mouse brain in-vivo at 9.4T. Our results show consistent parameters values between N=3 animals in both gray and white matter ROIs. Moreover, we have also compared the effects of using different number of shells and maximum b-values in the SANDI analysis. Our work augurs well for future investigations in animal models of plasticity and disease.

Christopher S Parker¹, Thomas Veale², Martina Bocchetta², Catherine F Slattery², Nick Fox², Jonathan M Schott², Dave M Cash^1,2, and Hui Zhang^1
1Centre for Medical Image Computing, Department of Computer Science, UCL, London, United Kingdom, ²Dementia Research Centre, Department of Neurodegenerative Disease, UCL Queen Square, Institute of Neurology, UCL, London, United Kingdom

Region-of-interest (ROI) metrics are typically computed as a mean over ROI voxels. However, for some NODDI metrics, this approach produces biased estimates in the presence of cerebrospinal fluid partial volume. We address this by introducing a tissue-weighted alternative. We compare the proposed mean to its conventional counterpart for periventricular and non-periventricular ROIs in healthy subjects and patients with young onset Alzheimer’s disease (YOAD). Results show the conventional mean overestimates orientation dispersion index and inflates inter-subject variation, particularly for periventricular ROIs and the YOAD cohort. This technique may improve detection of true regional effects in future group studies of neurodegenerative diseases.

Elisabetta Pagani¹, Paolo Preziosa^1,2, Raffaello Bonacchi^1,2, Laura Cacciaguerra^1,2,3, Massimo Filippi^1,2,3,4,5, and Maria A. Rocca^1,2,3
1Neuroimaging Research Unit, Division of Neuroscience, IRCCS San Raffaele Scientific Institute, Milan, Italy, ²Neurology Unit, IRCCS San Raffaele Scientific Institute, Milan, Italy, ³Vita-Salute San Raffaele Unversity, Milan, Italy, ⁴Neurorehabilitation Unit, IRCCS San Raffaele Scientific Institute, Milan, Italy, ⁵Neurophysiology Service, IRCCS San Raffaele Scientific Institute, Milan, Italy

In multiple sclerosis (MS), cortical damage is a relevant predictor of clinical disability. We applied Neurite orientation dispersion and density imaging (NODDI) to characterize microstructure of normal-appearing cortex (NA-cortex) and cortical lesions (CLs) and their relations with disease clinical phenotypes and disability. We found that a significant neurite loss occurs in MS NA-cortex, being more severe with longer disease duration, more severe disability and progressive MS. CLs show a further reduction of neurite density, together with an increased extracellular space, possibly due to inflammation and gliosis, and a reduced dispersion suggestive of increased tissue coherence and simplification of neurite complexity.

Abrar Faiyaz¹, Marvin M Doyley^1,2,3, Giovanni Schifitto^2,4, Jianhui Zhong^2,3,5, and Md Nasir Uddin^4
1Department of Electrical and Computer Engineering, University of Rochester, Rochester, NY, United States, ²Department of Imaging Sciences, University of Rochester, Rochester, NY, United States, ³Department of Biomedical Engineering, University of Rochester, Rochester, NY, United States, ⁴Department of Neurology, University of Rochester, Rochester, NY, United States, ⁵Department of Physics and Astronomy, University of Rochester, Rochester, NY, United States

The study explores the possibility to create neurite orientation dispersion and density imaging (NODDI) parameter maps with single-shell diffusion MRI data, using isotropic volume fraction (f_ISO) as prior. Proposed dictionary based deep learning prior NODDI (DLpN) framework leverages f_ISO estimation with a diffusion imaging scalar and T2 weighted non-diffusion signal. f_ISO estimation was validated in simulation and in-vivo. DLpN derived NDI and ODI maps for single-shell protocols are comparable with original multi-shell NODDI (error <5%) and may allow NODDI evaluation of retrospective brain studies on single-shell diffusion MRI data by multi-shell scanning of two subjects for DictNet f_ISO training.

Kilian Stumpf¹, Anna-Katinka Bracher², Thomas Hüfken¹, Britta Huch², Meinrad Beer², Henning Neubauer², and Volker Rasche^1
1Department of Internal Medicine II, Ulm University Medical Center, Ulm, Germany, ²Department of Diagnostic and Interventional Radiology, Ulm University Medical Center, Ulm, Germany

Intravoxel incoherent motion imaging allows the quantification of diffusion and pseudo-perfusion in tissues and can be used as an alternative to conventional T1-weighted scans that require the use of contrast agents. Its benefits for diagnosing juvenile idiopathic arthritis(JIA) in the knee has previously be demonstrated, but usually requires multi-b-value DWI scans with long scan times. In this work we investigate the influence of different b-value combinations on the calculated diffusion and perfusion fraction values and explore the possibility for IVIM MRI of JIA with fewer b-values and shorter scan times.

Alberto De Luca^1,2, Geert-Jan Biessels¹, Ofer Pasternak³, and Chantal MW Tax^4,5
1Department of Neurology, University Medical Center Utrecht, Utrecht, Netherlands, ²PROVIDI Lab, Image Sciences Institute, University Medical Center Utrecht, Utrecht, Netherlands, ³Brigham and Women's Hospital, Harvard Medical School, Boston, MA, United States, ⁴Cardiff University Brain Research, Imaging Centre (CUBRIC), University of Cardiff, Cardiff, United Kingdom, ⁵Image Sciences Institute, University Medical Center Utrecht, Utrecht, Netherlands

Diffusion MRI scans sensitive to intra-voxel incoherent motion (IVIM) are increasingly being applied to derive properties of capillary blood pseudo-diffusion in-vivo. A robust quantification of pseudo-diffusion measures requires the acquisition of diffusion MRI with multiple weightings and directions, which is time-consuming. Here, we investigate whether the isotropic averaging of spherical tensor diffusion encoding (STE) can improve IVIM estimates and shorten the acquisition time. Our simulations and in-vivo data show that STE can substantially shorten the acquisition time required to robustly quantify pseudo-diffusion while also removing a previously unreported bias in voxels containing complex white matter fiber configurations. 

Yen-Peng Liao^1,2, Shin-ichi Urayama^1,2, Tadashi Isa^1,2,3, and Hidenao Fukuyama^2,4
1Department of Neuroscience, Kyoto University Graduate School of Medicine, Kyoto, Japan, ²Human Brain Research Center, Kyoto University Graduate School of Medicine, Kyoto, Japan, ³Institute for the Advanced Study of Human Biology (ASHBi), Kyoto University, Kyoto, Japan, ⁴Yasu City Hospital, Yasu, Japan

To improve the uncertainties caused by the inappropriate diffusion model fitting in the IVIM-MRI, an optimal model mapping method has been proposed and demonstrated its usefulness. Here, we investigated the effect of the SNR in the method. Six different SNR data sets, derived by changing the numbers of average (NA) of six sets of DWI volumes, were analyzed. Then, the resultant maps reached a steady-state when the NA is four or higher. This study revealed that further investigations on b-value distribution optimization and denoising are necessary to obtain such high SNR data within practical scan time.

Lucas M da Costa¹, Bruno Hebling Vieira¹, Renata Ferranti Leoni¹, and Andre Monteiro Paschoal^1,2
1InBrain Lab - University of Sao Paulo, Ribeirao Preto, Brazil, ²LIM44, Instituto e Departamento de Radiologia, Faculdade de Medicina, Universidade de Sao Paulo, Sao Paulo, Brazil

The analysis of the Intravoxel Incoherent Motion (IVIM) signal can be performed with the Non-Negative Least Squares (NNLS) method. It allows detecting an indeterminate number of peaks associated with the multiple exponentials that compose the IVIM signal. Thus, it becomes essential to identify and classify these peaks. We performed a new method to analyze the NNLS spectrum applying the Gaussian Mixture. We obtained pseudo-diffusion maps with better contrast to visualize gliomas. Gaussian Mixture seems to replace the traditional peak search and classification algorithms, opening up possibilities to explore the NNLS spectrum.

Susanne Rauh¹, Oliver Maier², Oliver Gurney-Champion³, Melissa Hooijmans³, Rudolf Stollberger^2,4, Aart Nederveen³, and Gustav Strijkers^1
1Department of Biomedical Engineering and Physics, Amsterdam UMC, location AMC, Amsterdam, Netherlands, ²Institute of Medical Engineering, Graz University of Technology, Graz, Austria, ³Department of Radiology and Nuclear Medicine, Amsterdam UMC, location AMC, Amsterdam, Netherlands, ⁴BioTechMed-Graz, Graz, Austria

Model-based reconstruction can overcome many short-comings of IVIM and DTI by accelerating the acquisition, suppressing systematic errors from Rician noise and improving SNR. In this study we assess the feasibility of model-based reconstruction to obtain IVIM and combined IVIM-DTI parameter maps in the liver and kidneys in healthy volunteers. Results were compared to a conventional IVIM and IVIM-DTI least-squares fitting. Model-based reconstructions produced maps with less artifacts in liver and more details in the kidneys as compared to those from conventional fitting. Mean parameter values were similar for both methods.

Shira Nemirovsky-Rotman¹, Elad Rotman¹, Onur Afacan², Sila Kurugol², Simon Warfield², and Moti Freiman^1
1Biomedical Engineering, Technion, Haifa, Israel, ²Boston's Children's Hospital, Harvard Medical School, Boston, MA, United States

Quantitative Diffusion-Weighted MRI with the Intra-Voxel Incoherent Motion (IVIM) model shows potential to produce quantitative biomarkers for multiple clinical applications.   Recently, deep-learning (DL) models were proposed for the estimation of the IVIM model parameters from DW-MRI data. While the DL models produce more accurate parameter estimates compared to classical methods, their capability to generalize the IVIM model to different acquisition protocols is very limited. For that end, we introduce a physically motivated DL model, by incorporating the acquisition protocol into the network architecture. Our approach provides a DL-based method for IVIM parameter estimates that is agnostic to the acquisition protocol.

Serge Vasylechko^1,2, Simon K. Warfield^1,2, Onur Afacan^1,2, and Sila Kurugol^1,2
1Computational Radiology Laboratory, Boston Children's Hospital, Boston, MA, United States, ²Harvard Medical School, Boston, MA, United States

The goal of this study was to assess the robustness and repeatability of intravoxel incoherent motion model (IVIM) parameter estimation for the diffusion weighted MRI in the abdominal organs under the constraints of noisy diffusion signal using a novel neural network training method. The method is based on the principle of a physics guided self-supervised neural network that does not require supervision for training. Such approach is beneficial in conditions where the reference methods are not available, or are not robust enough to provide good supervision. This work is targeting evaluations towards accelerated IVIM DWI scanning which exhibit low SNR.

Sam Sharifzadeh Javidi¹ and Hamidreza Saligheh Rad^1,2
1Department of Medical Physics and Biomedical Engineering, Tehran University of Medical Sciences, Tehran, Iran (Islamic Republic of), ²Quantitative MR Imaging and Spectroscopy Group, Research Center for Molecular and Cellular Imaging, Tehran, Iran (Islamic Republic of)

The IVIM model is capable of extracting functional and structural information simultaneously without the injection of contrast agents. The main limitation of this technique is the inaccuracy of the output of this model in low SNR regimes. In this study, we proposed the use of twelve diffusion imaging orientations and three b-values instead of three orthogonal DW imaging and several b-values. Simulation and in-vivo results showed that the proposed method outperforms the conventional IVIM reconstruction method. Improved quality and reproducibility can make this method more practical and attractive in clinical settings.

Ante Zhu¹, Luca Marinelli¹, and Thomas K.F. Foo^1
1GE Global Research, Niskayuna, NY, United States

Oscillating gradient spin echo (OGSE) diffusion imaging achieves a short diffusion time and has been performed to assess tissue characteristics at short length scale, including the correlation length of the brain tissues which may provide information on axonal beading and swelling in disease. Time-dependent diffusivity has been measured from OGSE acquisition at a fixed TE in a high-performance head-only high-gradient 3.0T MRI scanner (MAGNUS). In this study, we characterized time-dependent diffusivity at varying frequencies (0~80 Hz) and TEs (72~158 ms) using multi-TE OGSE acquisition in healthy volunteers. Preliminary results showed increased mean/parallel diffusivities in corpus callosum at shorter TE. 

Sudhir Kumar Pathak¹, Vishwesh Nath², B V Rathish Kumar³, and Walter Schneider^1
1Psychology, University of Pittsburgh, Pittsburgh, PA, United States, ²Nvidia, Bethesda, MD, United States, ³Mathematics, Indian Institute of Technology Kanpur, Kanpur, India

Diffusion MRI based microstructural imaging is used to estimate voxel wise intra/extra-cellular volume fraction and compartment diffusivities. Previously proposed techniques like NODDI, SMT models are used to estimate intra/extra-axonal volume fraction, axial and radial diffusivity for each voxel. TEdDI is a framework to estimate intra/extra-axonal T₂ in human brain.  The proposed TEdDI model requires multiple sets of diffusion dataset varies with echo time. In this study we are proposing SMT based formulation to estimate intra/extra-axonalT₂. We also used this formulation to design numerical algorithm that approximate T₂ and other parameters for a typical multi-shell diffusion acquisition.

Jens T Rosenberg¹, Samuel Colles Grant^1,2, and Daniel Topgaard^3
1National High Magnetic Field Laboratory, Florida State University, Tallahassee, FL, United States, ²Chemical and Biomedical Engineering, FAMU-FSU College of Engineering, Tallahassee, FL, United States, ³Physical Chemistry, Lund University, Lund, Sweden

In vivo diffusion MRI is by default performed using single-shot EPI with TE>50 ms  and associated signal loss from transverse relaxation. The individual benefits of the current trends of increasing B₀ to boost SNR and employing more advanced signal preparation schemes to improve the specificity for selected microstructural properties eventually may be cancelled by increases in relaxation rates at high B₀ and TE with advanced encoding. Here, we make initial attempts to translate state-of-the-art diffusion-relaxation correlation methods to 21.1T to identify hurdles that need to be overcome to fulfill the promises of both high SNR and readily interpretable microstructural information.

Maryam Afzali¹, Lars Mueller¹, Ken Sakaie², Siyuan Hu³, Yong Chen⁴, Mark Griswold⁴, Derek K Jones¹, and Dan Ma^3
1Cardiff University Brain Research Imaging Centre (CUBRIC), School of Psychology, Cardiff University, Cardiff, United Kingdom, ²Imaging Institute, Cleveland Clinic, Cleveland, OH, United States, ³Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, ⁴Radiology, Case Western Reserve University, Cleveland, OH, United States

Diffusion magnetic resonance imaging provides information about microstructural features of a sample. Relaxometry on the other hand is sensitive to the biochemical environment of the underlying microstructure. Recent studies show that using conventional Stejskal-Tanner experiment, separating compartmental features is challenging and therefore b-tensor encoding and diffusion-relaxometry were proposed to mitigate this challenge. Magnetic resonance fingerprinting (MRF) can quantify multiple tissue parameters in one scan. Here, we implement multi-dimensional MR Fingerprinting scan with linear and spherical tensor encoding (LTE, and STE) and show the feasibility of estimating $$$T_1,\;\rm{and}\;T_2$$$ relaxation times and diffusivity on NIST, microstructure phantom and in vivo brain scans.

Jin Gao^1,2, Weiguo Li^2,3, Richard Magin³, and Danilo Erricolo^1,3
1Department of Electrical and Computer Engineering, University of Illinois at Chicago, Chicago, IL, United States, ²Research Resources Center, University of Illinois at Chicago, Chicago, IL, United States, ³Department of Bioengineering, University of Illinois at Chicago, Chicago, IL, United States

Multiple components analysis of nuclear magnetic resonance (MR) relaxation data using L₂-norm regularized non-negative least squares (NNLS) method has been widely used in myelin imaging for neurological diseases. When this analysis is applied to diffusion-weighted MR imaging to investigate water diffusion properties of biological tissues, noise corruption becomes a major problem which affects the accuracy and robustness of results. In this study, a L₁-norm regularized method was developed to process diffusion-weighted MRI data from spinal cords of amyotrophic lateral sclerosis affected mice.