Junqi Wang^1,2, Hailong Li^1,2, Gang Qu³, Jonathan R Dillman^1,2,4, Nehal A Parikh^5,6, and Lili He^1,2,4
1Department of Radiology, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States, ²Imaging Research Center, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States, ³Department of Biomedical Engineering, Tulane University, New Orleans, LA, United States, ⁴Department of Radiology, University of Cincinnati College of Medicine, Cincinnati, OH, United States, ⁵Center for Prevention of Neurodevelopmental Disorders, Cincinnati Children's Hospital Medical Center, Cincinnati, OH, United States, ⁶Department of Pediatrics, University of Cincinnati College of Medicine, Cincinnati, OH, United States

The human brain is a highly interconnected network where local activation patterns are organized to cope with diverse environmental demands. We developed a hypergraph learning based convolutional neural network model to capture higher order relationships between brain regions and learn representative features for brain connectome classification. The model was applied to a large scale resting state fMRI cohort, containing hundreds of healthy developing adolescents, age 8 to 22. The proposed model is able to classify different age groups with a balanced accuracy of 86.8%.

Raghav Singhal¹, Mukund Sudarshan¹, Luke Ginocchio², Angela Tong², Hersh Chandarana², Daniel Sodickson², Rajesh Ranganath^3,4, and Sumit Chopra^1,2
1Courant Institute of Mathematical Sciences, New York University, New York, NY, United States, ²Center for Advanced Imaging Innovation and Research (CAI2R), Department of Radiology, New York University Grossman School of Medicine, New York, NY, United States, ³Department of Population Health, New York University, New York, NY, United States, ⁴Center for Data Science, New York University, New York, NY, United States

Despite its rich clinical information content, MR has seen limited adoption in population-level screenings, due to concerns about specificity combined with high scan duration and cost.  In order to begin to address such issues, and to accelerate the entire pipeline from data acquisition to diagnosis, we introduce ARMS, an algorithm that learns k-space undersampling patterns to maximize the accuracy of pathology detection. ARMS detects pathologies directly from undersampled k-space data, bypassing explicit image reconstruction. We use ARMS to detect clinically significant prostate cancer and knee abnormalities in 2D MR scans, achieving an acceleration of 12.5x without compromising accuracy. 

Cayden Murray¹, Olayinka Oladosu¹, and Yunyan Zhang ^2,3
1Neuroscience, University of Calgary, Calgary, AB, Canada, ²Radiology, University of Calgary, Calgary, AB, Canada, ³Clinical Neurosciences, University of Calgary, Calgary, AB, Canada

High Angular Resolution Diffusion Imaging (HARDI) is a promising method for the analysis of microstructural changes. However, HARDI acquisition is time-consuming and therefore impractical in clinical settings. We developed 2 neural networks for predicting non-acquired diffusion datasets based on diffusion MRI: Multi-layer Perceptron (MLP) and Convolutional Neural Network (CNN). Through systemic training and evaluation with healthy public data and local MS patient MRI, we found that both the MLP and CNN models could predict high b-value from low b-value data that allowed the assessment of Neurite Orientation Dispersion and Density Imaging (NODDI) outcomes. Neural networks can make NODDI clinically viable.

Junshen Xu¹, Molin Zhang¹, Lana Vasung^2,3,4, Esra Abaci Turk^2,3,4, Borjan Gagoski^3,4,5, Polina Golland^1,6, P. Ellen Grant^2,3,4,5, and Elfar Adalsteinsson^1,7
1Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ²Department of Pediatrics, Boston Children’s Hospital, Boston, MA, United States, ³Harvard Medical School, Boston, MA, United States, ⁴Fetal-Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Boston, MA, United States, ⁵Department of Radiology, Boston Children’s Hospital, Boston, MA, United States, ⁶Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, MA, United States, ⁷Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA, United States

Fetal motion is an important indicator of fetal health and nervous system development. Current assessments of fetal motion with MRI or ultrasound are qualitative and do not reflect the 3D motion of each body part . To study the detailed motion of fetuses, annotations of fetal pose are required, which would be time-consuming through manually-labelled data for each scan. In this work, we demonstrate an automated and efficient pipeline for fetal pose and motion estimation of fetal MRI using deep learning. The results of experiments show that the proposed pipeline outperforms other state-of-the-art fetal pose estimation methods.

Junshen Xu¹, Daniel Moyer², P. Ellen Grant^3,4,5,6, Polina Golland^1,2, Juan Eugenio Iglesias^2,4,7,8, and Elfar Adalsteinsson^1,9
1Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ²Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, MA, United States, ³Fetal-Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Boston, MA, United States, ⁴Harvard Medical School, Boston, MA, United States, ⁵Department of Pediatrics, Boston Children’s Hospital, Boston, MA, United States, ⁶Department of Radiology, Boston Children’s Hospital, Boston, MA, United States, ⁷Centre for Medical Image Computing, University College London, London, United Kingdom, ⁸Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Cambridge, MA, United States, ⁹Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA, United States

Volumetric reconstruction of fetal brains from MR slices is a challenging task, which is sensitive to the initialization of slice-to-volume transformations. Further complicating the task is the unpredictable fetal motion. In this abstract, we proposed a novel method for slice-to-volume registration using transformers, which models the stacks of MR slices as a sequence. With the attention mechanism, the proposed model predicts the transformation of one slice using information from other slices. Results show that the proposed method achieves not only lower registration error but also better generalizability compared with other state-of-the-art methods for slice-to-volume registration of fetal MRI.

Yahang Li^1,2, Loreen Ruhm^3,4, Anke Henning^3,4, and Fan Lam^1,2,5
1Department of Bioengineering, University of Illinois Urbana-Champaign, Urbana, IL, United States, ²Beckman Institute for Advanced Science and Technology, Urbana, IL, United States, ³Advanced Imaging Research Center, University of Texas Southwestern Medical Center (UTSW), Dallas, TX, United States, ⁴Max Planck Institute for Biological Cybernetics, Tübingen, Germany, ⁵Cancer Center at Illinois, Urbana, IL, United States

We proposed here a novel method for computationally efficient reconstruction from noisy MRSI data. The proposed method is characterized by (a) a strategy that jointly learns a nonlinear low-dimensional representation of high-dimensional spectroscopic signals and a projector to recover the low-dimensional embeddings from noisy FIDs; and (b) a formulation that integrates forward encoding model, a spectral constraint from the learned representation and a complementary spatial constraint. The learned projector allows for the derivation of a highly efficient algorithm combining projected gradient descent and ADMM. The proposed method has been evaluated using simulation and in vivo data, demonstrating impressive SNR-enhancing performance. 

Yan Zhang¹ and Jun Shen^1
1National Institute of Mental Health, Bethesda, MD, United States

Multi echo techniques such as JPRESS consist of both short and long echoes and provide more diversified information for spectral fitting than techniques based on a single echo. However, fitting multi echo data is more challenging because signals attenuate with increasing echo time due to T2 relaxation, and the macromolecule background also varies across the echoes. We present a novel neural network architecture that directly maps the time domain JPRESS input onto metabolite concentrations. The testing results show the model can successfully predict in vivo metabolite concentrations from multi-echo JPRESS data after being trained with quantum mechanics simulated spectral data.

Burhaneddin Yaman^1,2, Seyed Amir Hossein Hosseini^1,2, and Mehmet Akcakaya^1,2
1Electrical and Computer Engineering, University of Minnesota, Minneapolis, MN, United States, ²Center for Magnetic Resonance Research, Minneapolis, MN, United States

While self-supervised learning enables training of deep learning reconstruction without fully-sampled data, it still requires a database. Moreover, performance of pretrained models may degrade when applied to out-of-distribution data. We propose a zero-shot subject-specific self-supervised learning via data undersampling (ZS-SSDU) method, where acquired data from a single scan is split into at least three disjoint sets, which are respectively used only in physics-guided neural network, to define training loss, and to establish an early stopping strategy to avoid overfitting. Results on knee and brain MRI show that ZS-SSDU achieves improved artifact-free reconstruction, while tackling generalization issues of trained database models.

Qiyuan Tian¹, Ziyu Li², Wei-Ching Lo³, Berkin Bilgic¹, Jonathan R. Polimeni¹, and Susie Y. Huang^1
1Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Harvard Medical School, Charlestown, MA, United States, ²Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom, ³Siemens Medical Solutions, Charlestown, MA, United States

The requirement for high-SNR reference data for training reduces the practical feasibility of supervised deep learning-based denoising. This study improves the accessibility of deep learning-based denoising for MRI using transfer learning that only requires high-SNR data of a single subject for fine-tuning a pre-trained convolutional neural network (CNN), or self-supervised learning that can train a CNN using only the noisy image volume itself. The effectiveness is demonstrated by denoising highly accelerated (R=3×3) Wave-CAIPI T1w MPRAGE images. Systematic and quantitative evaluation shows that deep learning without or with very limited high-SNR data can achieve high-quality image denoising and brain morphometry.

Matt Hemsley¹, Liam S.P Lawrence¹, Rachel W Chan², Brige Chuge^3,4, and Angus Z Lau^1,2
1Medical Biophysics, University of Toronto, Toronto, ON, Canada, ²Physical Sciences, Sunnybrook Research Insitute, Toronto, ON, Canada, ³Department of Radiation Oncology, Sunnybrook Health Sciences Centre, Toronto, ON, Canada, ⁴Department of Physics, Ryerson University, Toronto, ON, Canada

Convolutional Neural Networks behave unpredictively when test images differ from the training images, for example when different sequences or acquisition parameters are used. We trained models to generate synthetic CT images, and tested the models on both in-distribution and out-of-distribution input sequences to determine the magnitude of performance loss. Additionally, we evaluated if uncertainty estimates made using dropout-based variational inference could detect spatial regions of failure. Networks tested on out of distribution images failed to generate accurate synthetic CT images. Uncertainty estimates identified spatial regions of failure and increased with the difference between the training and testing sets.

Zhaoyuan Gong¹, Nikkita Khattar², Matthew Kiely¹, Curtis Triebswetter¹, Maryam H. Alsameen¹, and Mustapha Bouhrara^1
1National Institute on Aging, Baltimore, MD, United States, ²Yale University, New Haven, CT, United States

The Myelin water fraction (MWF) measure provides a direct assessment of myelin content. The widely utilized method is the multicomponent analysis of T2 relaxation time and MWF is determined by the fraction of the fast-relaxing component. However, using either conventional or advanced methods, such as the BMC-mcDESPOT, requires prolonged acquisition and computation times, hampering their integration in clinical investigations. In this proof-of-concept work, we propose artificial neural network models to derive MWF maps from under sampled mcDESPOT data through two distinct approaches. This work opens the way to further developments for practical and rapid MWF imaging.

Daniel Güllmar¹, Wei-Chan Hsu^1,2, Stefan Ropele³, and Jürgen R. Reichenbach^1,2
1Institute of Diagnostic and Interventional Radiology, Medical Physics Group, Jena University Hospital, Jena, Germany, ²Michael Stifel Center Jena for Data-Driven and Simulation Science, Jena, Germany, ³Division of General Neurology, Medical University Graz, Graz, Austria

Synthetic medical images can be generated with a StyleGAN and are indistinguishable from real data even by experts. However, the projection of real data via latent space onto a synthetic image shows clear deviations from the original (at least on the second image). This plays a major role especially when using GANs to perform tasks such as image correction (e.g. noise reduction), image interpolation or image interpretation by analyzing the latent space. Based on the results shown, it is highly recommended to perform an analysis of the projection accuracy before applying any of these applications.

Michael Stritt¹, Matthias Günther^1,2,3, Johannes Gregori^1,4, Daniel Mensing¹, Henk-Jan Mutsaerts⁵, and Klaus Eickel^1,3
1mediri GmbH, Heidelberg, Germany, ²Universität Bremen, Bremen, Germany, ³Fraunhofer MEVIS, Bremen, Germany, ⁴Darmstadt University of Applied Sciences, Darmstadt, Germany, ⁵Amsterdam University Medical Center, Amsterdam, Netherlands

The presented neural network with 3D-FiLM-cGAN architecture synthesizes cerebral blood flow maps from T1-weighted input images. Acquisition- and subject-specific metadata such as sex, arterial spin labeling (ASL) method and readout techniques were fed into the neural network as auxiliary input. The multi-vendor database including different ASL sequence types was created from ADNI data which were preprocessed in ExploreASL and transformed to MNI standard space. A subset of data from a single vendor (GE) were used for supervised training exemplarily and compared to CBF from acquired ASL data.

Chompunuch Sarasaen^1,2,3, Soumick Chatterjee^1,3,4,5, Georg Rose^2,3, Andreas Nürnberger^4,5,6, and Oliver Speck^1,3,6,7,8
1Biomedical Magnetic Resonance, Otto von Guericke University, Magdeburg, Germany, ²Institute for Medical Engineering, Otto von Guericke University, Magdeburg, Germany, ³Research Campus STIMULATE, Otto von Guericke University, Magdeburg, Germany, ⁴Data and Knowledge Engineering Group, Otto von Guericke University, Magdeburg, Germany, ⁵Faculty of Computer Science, Otto von Guericke University, Magdeburg, Germany, ⁶German Center for Neurodegenerative Disease, Magdeburg, Germany, ⁷Center for Behavioral Brain Sciences, Magdeburg, Germany, ⁸Leibniz Institute for Neurobiology, Magdeburg, Germany

Cartesian sampling techniques are available to speed up the measurement of dynamic MRI, such as k-t GRAPPA. However, radial samplings, such as iGRASP, are more robust to motion and can be applied for abdominal dynamic MRI. In this work, k-t GRAPPA inspired iGRASP has been created (so-called k-t GRASP)–which acquires the subsequent time points by starting the initial spoke of the time point with an angle with the last spoke of the previous time point, and extended by super-resolution reconstruction of dynamic abdominal MRI using DDoS-UNet. The method was evaluated in 3D dynamic data of four subjects with retrospective undersampling. 

Vishwanatha Mitnala Rao¹, Zihan Wan², David Ma¹, Ye Tian¹, and Jia Guo^3
1Department of Biomedical Engineering, Columbia University, New York, NY, United States, ²Department of Applied Mathematics, Columbia University, New York, NY, United States, ³Department of Psychiatry, Columbia University, New York, NY, United States

Despite achieving compelling performance, many deep learning automated brain tissue segmentation solutions struggle to generalize to new datasets due to properties inherent to MRI scans. We propose TABS, a new transformer-based deep learning architecture that achieves state-of-the-art-performance, generalization, and consistency. We tested TABS on three datasets of differing field strands and acquisition parameters. TABS outperformed RAUnet on our performance testing and remained consistent across test-retest repeated scans from a separate dataset. Moreover, TABS achieved impressive generality performance and even improved in performance across datasets. We believe TABS represents a generalized and accurate brain tissue segmentation alternative.