Peyman Shokrollahi¹, Juan M. Zambrano Chaves¹, Jonathan P.H. Lam¹, Avishkar Sharma¹, Debashish Pal², Naeim Bahrami², Akshay S. Chaudhari¹, and Andreas M. Loening^1
1Radiology, Stanford University, Stanford, CA, United States, ²GE Healthcare, Sunnyvale, CA, United States

The use of inappropriate radiology protocols introduces risk of missed and incomplete diagnoses, thus endangering patient health, potentially prolonging treatment, and increasing healthcare costs. A clinical decision support system based on machine learning and electronic medical records of patients undergoing MRI was developed to predict radiology titles and their probabilities for radiologist review. A cumulative F1-score of ~85% was obtained for the top three predicted titles. The proposed system can guide physicians toward selecting appropriate titles and alert radiologists of potentially inappropriate selections, thereby improving imaging utility and increasing diagnostic accuracy, which favors better patient outcomes.  

Omer Faruk Gulban^1,2, Renzo Huber³, Konrad Wagstyl⁴, Saskia Bolmann⁵, Dimo Ivanov³, Kendrick Kay⁶, and Rainer Goebel^2,3
1Cognitive Neuroscience Department, Maastricht University, Maastricht, Netherlands, ²Brain Innovation, Maastricht, Netherlands, ³Cognitive Neuroscience, Maasticht University, Maastricht, Netherlands, ⁴UCL, London, United Kingdom, ⁵Center for Advanced Imaging, The University of Queensland, Queensland, Australia, ⁶Center for Magnetic Resonance Research, University of Minnesota, Minneapolis, MN, United States

High resolution imaging poses analysis challenges in the daily lives of high-resolution imaging researchers (high res. MRI, histology). One challenge is flattening patches of the cortex to study the cortical topography. We demonstrate a method that works in volume without requiring triangular meshes to represent the cortical surfaces at any stage of the algorithm. This approach for surface mapping enables mesoscopic analysis of cortical architecture. Our method is specifically developed for -but not limited to- flattening patches of the cortex. We implement our patch flattening as a C++ program and make it freely available within LayNii v2.1.0.

Brian S Winston¹, Hoyt Patrick Taylor², Pew-Thian Yap², Frederick S Barrett¹, and James J Pekar^3,4
1Department of Psychiatry and Behavioral Sciences, Johns Hopkins University, Baltimore, MD, United States, ²Department of Radiology and BRIC, University of North Carolina, Chapel Hill, NC, United States, ³F.M. Kirby Research Center, Kennedy Krieger Institute, Baltimore, MD, United States, ⁴Department of Radiology, Johns Hopkins University, Baltimore, MD, United States

The connectome harmonic decomposition provides a principled approach to dimensionality reduction of human brain functional data using the eigenmodes of the graph Laplacian derived from the anatomical connectome. In the present report, test-retest data from the Human Connectome Project are used to assess the reliability of the structural harmonics, and of decompositions of resting-state functional MRI data using these harmonics. For both structure and function, reliability was higher within than between subjects, and decreased with harmonic rank.

Nian Wang^1,2 and Surendra Maharjan^1
1Radiology and Imaging Sciences, Indiana University, Indianapolis, IN, United States, ²Stark Neurosciences Research Institute, Indiana University, Indianapolis, IN, United States

Magnetic resonance imaging (MRI) offers a noninvasive method for characterizing brain neuroanatomy and provides insights into various disease processes. Diffusion MRI (dMRI) has been used for assessment of tissue microstructure due to its exquisite sensitivity to finer scales far above imaging resolution. The recently developed Allen Mouse Brain Common Coordinate Framework (CCFv3) has opened the opportunities to integrate multimodal and multiscale datasets of whole mouse brain in 3D, and thus, this approach can be used to compare quantitative MRI parameters with conventional histology features throughout the whole brain.

Nicola Katharine Dinsdale¹, Mark Jenkinson^2,3, and Ana IL Namburete^4
1WIN / Oxford Machine Learning in NeuroImaging Lab, University of Oxford, Oxford, United Kingdom, ²WIN, University of Oxford, Oxford, United Kingdom, ³Australian Institute for Machine Learning, University of Adelaide, Adelaide, Australia, ⁴Computer Science, Oxford Machine Learning in NeuroImaging Lab, University of Oxford, Oxford, United Kingdom

We propose an algorithm to simultaneously prune and train CNNs, leading to networks which have increased generalisability across imaging sites. Through segmentation of data from the ABIDE dataset, we show that through reducing the number of parameters in the network throughout training, we are able to reduce model overfitting, creating a model which is more robust to expect image variations across scanners. We also introduce a novel Targeted Dropout algorithm, which aids the process of model pruning. We demonstrate the approach on a UNet architecture, the basis of nearly all segmentation approaches across medical imaging.

Odélia Jacqueline Chitrit¹, Qingjia Bao¹, Silvia Chuartzman², Noga Silkha², Tali Kimchi², and Lucio Frydman^1
1Chemical and biological physics, Weizmann institute of Science, Rehovot, Israel, ²Neurobiology, Weizmann institute of Science, Rehovot, Israel

Single-shot fMRI executed at ultrahigh fields can reveal valuable insight about brain function, if it successfully overcomes field inhomogeneity problems. Spatiotemporal Encoding provides a route for achieving this, enabling studies on the olfactory bulbs of mice at 15.2T. Images collected with a 125 µm in-plane resolution yielded remarkably large and well-defined responses to olfactory cues, particularly in males. These were unambiguously linked to olfaction via single-nostril experiments.  The experiments highlighted specific activation regions in the external plexiform region and in glomeruli in the lateral part of the bulb, when stimulated by aversive or appetitive odors, respectively. 

Alexandru V Avram^1,2, Kadharbatcha S Saleem^1,2, Michal E Komlosh^1,2, and Peter J Basser^1
1Eunice Kennedy Shriver National Institute of Child Health and Human Development, National Institutes of Health, Bethesda, MD, United States, ²Center for Neuroscience and Regenerative Medicine, The Henry Jackson Foundation, Bethesda, MD, United States

We investigate the sensitivity of high-resolution MAP-MRI to distinguish cortical architectonic features by directly comparing laminar patterns in whole-brain volumes of MAP-derived parameters to the corresponding histological sections in the same macaque monkey brain. MAP parameters, in particular PA, NG, RTAP, DEC, and fODF maps, show cortical area-specific lamination patterns with excellent conspicuity in the mid-cortical layers, matching, and sometimes complementing, the contrast observed in the immuno- or histochemically stained sections. Delineating cortical areas with distinct cytoarchitectonic features non-invasively based on diffusion propagators could transform our ability to study brain networks and diagnose subtle microstructural changes in many clinical applications. 

Dengrong Jiang¹, Raymond C. Koehler², Xiuyun Liu², Ewa Kulikowicz², Jennifer K. Lee², Hanzhang Lu^1,3,4, and Peiying Liu^1
1Department of Radiology, Johns Hopkins University School of Medicine, Baltimore, MD, United States, ²Department of Anesthesiology and Critical Care Medicine, Johns Hopkins University School of Medicine, Baltimore, MD, United States, ³Department of Biomedical Engineering, Johns Hopkins University School of Medicine, Baltimore, MD, United States, ⁴F.M. Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Research Institute, Baltimore, MD, United States

The neonatal brain relies primarily on oxygen metabolism to meet its enormous energy demands. Cerebral venous oxygenation (Y_v) is an important parameter of the brain’s oxygen utilization, and has been demonstrated to be a potential biomarker in various neonatal diseases. We previously developed two non-invasive MRI techniques, TRUPC and accelerated-TRUPC, to measure Y_v in the neonatal brain. In this work, we validated the accuracy of these two techniques by comparing their Y_v measurements with gold-standard blood gas oximetry on a piglet model, and demonstrated that TRUPC and accelerated-TRUPC can provide accurate quantifications of Y_v.

Kilian Hett¹, Colin D. McKnight², Jarrod J. Eisma¹, Jason Elenberger¹, Ciaran M. Considine¹, Daniel O. Claassen¹, and Manus J. Donahue^1,3
1Neurology, Vanderbilt University Medical Center, Nashville, TN, United States, ²Radiology and Radiological Sciences, Vanderbilt University Medical Center, Nashville, TN, United States, ³Psychiatry and Behavioral Sciences, Vanderbilt University Medical Center, Nashville, TN, United States

Despite prior assumptions that the brain is devoid of lymphatic vessels, emerging evidence indicates that lymphatic vessels are present in the region alongside the superior sagittal sinus, the parasagittal dural (PSD) space, and could play a critical role in waste clearance or glymphatic physiology. However, limited methods are available for interrogating this space non-invasively in vivo. Here, in 62 healthy participants we apply novel deep learning algorithms in sequence with anatomical and phase contrast cerebrospinal fluid (CSF) flow assessments to test fundamental hypotheses regarding how PSD volume evolves across the lifespan and relates to measures of CSF flux and volume.

Joseph Kus ¹, Shakthi Kumaran Ramasamy¹, Brett Ploussard ¹, Joyce Cara ², Stacey Bennis³, Collen Fitzgerald ³, Steve Shea ¹, Ari D Goldberg ¹, Judy James ⁴, and Anugayathri Jawahar^1
1Radiology, Loyola University Medical Center, Maywood, IL, United States, ²Biostatistics, Stritch School of Medicine, Maywood, IL, United States, ³Physical Medicine and Rehabilitation, Loyola University Medical Center, Maywood, IL, United States, ⁴Radiology and Nuclear Medicine, Loyola University Medical Center, Maywood, IL, United States

Pudendal neuralgia is a common cause of chronic pelvic pain, especially in females. This is caused by pudendal nerve entrapment and can be a severely disabling neuropathic pain syndrome. It is currently a clinical diagnosis and most of the time a diagnosis of exclusion without definitive imaging criteria. We retrospectively compared DESS with the routine T2, T1 and DWI described in literature for the evaluation of pudendal nerves in patients with unresolved pelvic pain. Our study showed that DESS is effective and better than T2 and DWI combined sequences for diagnosing pudendal neuropathy.

Jyoti Mangal^1,2, Sila Ayse Dokumaci^1,2, David Leitao^1,2, Raphael Tomi-Tricot^1,2,3, Amer Ajanovic^1,2, Stephen Ogier^1,2, Tom Wilkinson^1,2, Sharon Giles^1,2, Pip Bridgen^1,2, Claudia Prieto¹, Joseph V Hajnal^1,2, Shaihan Malik^1,2, and David W Carmichael^1,2
1Biomedical Engineering Department, School of Biomedical Engineering and Imaging Sciences, King's College London, London, United Kingdom, ²London Collaborative Ultra high field System (LoCUS), King’s College London, London, United Kingdom, ³MR Research Collaborations, Siemens Healthcare Limited, Frimley, United Kingdom

Quantitative MRI at 7T has advantages of enhanced resolution and uniform tissue contrast however protocol optimisation can be challenging as B1(+) field variability is increased. We maximized the precision of T1 estimation for a dual flip angle SPGR acquisition using Cramer-Rao Lower Bound framework and validated the results using a Monte Carlo simulation and both phantom and in-vivo experiments. For a TR of 20ms, simulations suggested optimal flip angles of 4° and 26° similar to in-vivo results of 4° and 28°. By accounting for other parameters, the CRLB framework can be flexibly extended to optimize more complex mapping protocols.

Santiago Coelho¹, Filip Szczepankiewicz², Jelle Veraart¹, Sohae Chung¹, Yvonne W. Lui¹, Dmitry S. Novikov¹, and Els Fieremans^1
1Center for Advanced Imaging Innovation and Research (CAI2R), Department of Radiology, New York University, School of Medicine, New York, NY, United States, ²Department of Diagnostic Radiology, Clinical Sciences, Lund University, Lund, Sweden

Joint modeling of diffusion and relaxometry in white matter is attractive due to its potential to provide unique insights into tissue integrity. We designed a diffusion MRI protocol containing varying b-values, b-tensor shapes, and echo times, employed machine learning parameter estimation, and studied the reproducibility of the Standard Model (SM) parameters. We quantified scan-rescan reproducibility in six healthy volunteers for compartmental water fractions, diffusivities, and $$$T_2$$$ relaxation times. We also found good agreement between SM parameters obtained from diffusion-only data at fixed echo time, and those from joint diffusion-relaxometry acquisitions, providing a consistency check for the SM assumptions.

Evgenios N. Kornaropoulos^1,2, Stefan Winzeck^2,3, Linda Knutsson^4,5,6, Marta M. Correia⁷, Pia C. Sundgren^1,8,9, and Markus Nilsson^1
1Diagnostic Radiology, Lund University, Lund, Sweden, ²Division of Anaesthesia, University of Cambridge, Cambridge, United Kingdom, ³Computing, Imperial College London, London, United Kingdom, ⁴Lund University, Lund, Sweden, ⁵Johns Hopkins University, Baltimore, MD, United States, ⁶Kennedy Krieger Institute, Baltimore, MD, United States, ⁷MRC Cognition and Brain Sciences Unit, University of Cambridge, Cambridge, United Kingdom, ⁸Skane University Hospital, Lund, Sweden, ⁹BioImaging Center, Lund University, Lund, Sweden

We examined whether more advanced dMRI increase the sensitivity to white matter pathology in systemic lupus erythematosus. Specifically, we examined whether more advanced acquisitions (DKI versus DTI, 7T versus 3T) and more complex processing (e.g. denoising, Gibbs-ringing removal) provide increased sensitivity. Tract-based spatial statistical analysis was applied, and the fraction of significant voxels in the white matter skeleton was used as the sensitivity metric. Results showed that the choice of the acquisition protocol had a greater impact on sensitivity than the choice of the post-processing strategy. 3T-DTI was more sensitive than 3T-DKI and 7T-DTI, regardless of the applied post-processing pipeline.

Jo Schlemper¹, Neel Dey², Seyed Sadegh Mohseni Salehi¹, Kevin Sheth³, W. Taylor Kimberly⁴, Lorraine Cullen⁵, and Michal Sofka^1
1Hyperfine, Guilford, CT, United States, ²New York University, New York City, NY, United States, ³Yale University, New Haven, CT, United States, ⁴Massachusetts General Hospital, Boston, MA, United States, ⁵Gaylord Specialty Healthcare, Wallingford, CT, United States

We present an unsupervised deep image registration framework for MR image reconstruction. Specifically, even and odd echo images from fast spin echo-based sequence are nonlinearly registered using a convolutional network that estimates the deformation field. The registered echo images are then  averaged for noise reduction. The proposed framework was evaluated across four imaging contrasts (T1w, T2w, FLAIR, and DWI) from a low-field MR scanner and was found to outperform nonlinear registration from advanced normalization tools, yielding sharper image quality and preserving important pathology features.