Ben A Duffy¹, Srivathsa Pasumarthi Venkata¹, Long Wang¹, Sara Dupont¹, Lei Xiang¹, Greg Zaharchuk¹, and Tao Zhang^1
1Subtle Medical Inc., Menlo Park, CA, United States

We present an automated deep learning-based quality control system that generalizes to images of different orientations, images with and without contrast as well as those from different acquisition sites. Because the same model was able to classify images with different orientations, test-time augmentation substantially improved performance. Images that were moderately affected by artifacts were able to be identified with 95% accuracy. Furthermore, robustness to different data types (and potentially artifact types) was ensured by using an out-of-distribution detection procedure. This was able to discriminate spine MRI images from T1 brain images with an AUC of 0.98.

Zhixuan Liang¹, Yin Guo², and Chun Yuan^3
1Electrical Engineering, Zhejiang University, Hangzhou, China, ²Bioengineering, University of Washington, Seattle, WA, United States, ³Radiology, University of Washington, Seattle, WA, United States

This study aims to develop an automatic and robust algorithm for longitudinal registration of knee Magnetic Resonance Imaging (MRI) across the time span of eight years. We propose a technique that firstly achieves rigid registration based on femoral segmentation, and then makes evaluations based on tibial alignment. Both qualitative and quantitative results show solid performance of the proposed algorithm. This registration algorithm lays a solid foundation for serial analysis of knee images taken over a multi-year time frame. 

Kaia Ingerdatter Sørland¹, Mohammed R. S. Sunoqrot¹, Pål Erik Goa^2,3, Elise Sandsmark³, Sverre Langørgen³, Helena Bertilsson^4,5, Gigin Lin⁶, Tone F. Bathen^1,3, and Mattijs Elschot^1,3
1Department of Circulation and Medical Imaging, Norwegian University of Science and Technology, Trondheim, Norway, ²Department of Physics, Norwegian University of Science and Technology, Trondheim, Norway, ³Department of Radiology and Nuclear Medicine, St. Olavs University Hospital, Trondheim, Norway, ⁴Department of Cancer Research and Molecular Medicine, Norwegian University of Science and Technology, Trondheim, Norway, ⁵Department of Urology, St. Olavs University Hospital, Trondheim, Norway, ⁶Department of Medical Imaging and Intervention, Chang Gung Memorial Hospital at Linkou, Institute for Radiological Research, Chang Gung University, Taoyuan, Taiwan

Scanner-dependent variations induce non-standard signal intensities (SI) in T2-weighted (T2W) MR images. These variations make computer aided diagnosis of prostate cancer based on T2W images challenging, and SI normalization is necessary. Autoref is a normalization method for axial T2W prostate MRI based on two reference tissues of high and low intensity. The aim of this work was to evaluate Autoref’s performance on a large dataset, and to investigate its performance for various reference tissues. Femoral head and fat were proven to be stable reference tissues, significantly reducing inter-scan variation.

Ryan McNaughton¹, Hernan Jara^1,2, Chris Pieper², Laurie Douglass², Rebecca Fry³, Karl Kuban², and T. Michael O'Shea^3
1Boston University, Boston, MA, United States, ²Boston University Medical Center, Boston, MA, United States, ³University of North Carolina at Chapel Hill School of Medicine, Chapel Hill, MA, United States

Purpose: To describe an integrated, semi-automated image processing pipeline for multispectral qMRI, termed qVision. Methods: Dual-clustering and MS-qMRI python algorithms for the Tri-TSE pulse sequence are automatically calculated and harmonized across a dataset of neuroimaging data from adolescents born extremely preterm. Results: Automated processing is completed in 30 minutes per subject, resulting in high-resolution mappings of T1, T2, PD, and spatial entropy, as well as heavily R1-weighted images of white matter texture via Synthetic-MRI. Conclusion: qVision has been validated on a large-scale, multi-site, and multi-vendor dataset of neuroimaging data, capable of producing a broad spectrum of MS-qMRI outcomes.

Ivan I. Maximov^1,2, Dennis van der Meer², Ann-Marie de Lange², Tobias Kaufmann², Alexey Shadrin², Oleksandr Frei², Thomas Wolfers², and Lars T Westlye^2
1Western Norway University of Applied Sciences, Bergen, Norway, ²NORMENT, University of Oslo, Oslo, Norway

Diffusion MRI is a powerful approach to quantify brain architecture. However, diffusion scalar maps derived from raw data are sensitive to the data quality and processing choices. Many quality control algorithms exist that perform a robust check of raw diffusion data, there is a lack of QCs for inspecting the derived maps from different diffusion approaches. We present a novel QC algorithm for processed scalar maps using mean skeleton values (in the context of tract-based spatial statistics) and structural similarity metric based on the scalar maps. The algorithm builds on clustering of scalar diffusion metrics from 18609 UK Biobank individuals.

William J Matloff¹, Daniel J Matloff¹, Arthur W Toga¹, Taeuk Cheon², Jangwook Gwak², Yehree Kim², Hong Ju Park², and Hosung Kim^1
1Laboratory of Neuro Imaging, USC Mark and Mary Stevens Neuroimaging and Informatics Institute, Keck School of Medicine, University of Southern California, Los Angeles, CA, United States, ²Department of Otorhinolaryngology-Head and Neck Surgery, Asan Medical Center, Seoul, Korea, Republic of

For MRI-based studies of cochlea-related diseases, segmenting the cochlea and modiolus is often essential for quantifying disease-related morphological and intensity changes. Methods for automatic 3D segmentation of these structures on MRI scans, however, are not well established. We aimed to develop a semi-automated, multi-atlas based approach to segmenting the cochlea and modiolus, distinguishing between the cochlear basal and mid-apical turns. We found this approach to have good agreement with manual segmentation, demonstrating its potential in quantitative analyses. Moreover, the definition of the three subregions had good validity, as indicated by the high inter-rater agreement in segmentation.

Luigi Lorenzini¹, Silvia Ingala¹, Alle Meije Wink¹, Joost PA Kuijer¹, Viktor Wottschel¹, Carole Sudre^2,3,4,5, Sven Haller^6,7, José Luis Molinuevo^8,9,10,11, Juan Domingo Gispert^8,10,11,12, David M Cash¹³, David L Thomas¹⁴, Sjoerd B Vos^14,15, Ferran Prados Carrasco^16,17,18, Jan Petr¹⁹, Robin Wolz^20,21, Alessandro Palombit²⁰, Adam J Schwarz²², Gael Chételat²³, Pierre Payoux^24,25, Carol Di Perri²¹, Cyril Pernet²⁶, Frisoni Giovanni^27,28, Nick C Fox¹³, Craig Ritchie²⁹, Joanna Wardlaw^26,30, Adam Waldman^26,31, Frederik Barkhof^1,32, and Henk JMM Mutsaerts^1,33
1VUmc Amsterdam, Amsterdam, Netherlands, ²MRC unit for Lifelong Health and Ageing at UCL, London, London, United Kingdom, ³Department of Neurodegenerative Disease, Dementia Research Centre, London, United Kingdom, ⁴Centre for Medical Image Computing UCL, London, United Kingdom, ⁵School of Biomedical Engineering & Imaging Sciences, King’s College London, London, United Kingdom, ⁶CIRD Centre d’Imagerie Rive Droite, Geneva, Switzerland, ⁷Department of Surgical Sciences, Radiology, Uppsala University, Uppsala, Sweden, ⁸Barcelonaβeta Brain Research Center (BBRC), Pasqual Maragall Foundation, Barcelona, Spain, ⁹CIBER Fragilidad y Envejecimiento Saludable (CIBERFES), Madrid, Spain, ¹⁰IMIM (Hospital del Mar Medical Research Institute), Barcelona, Spain, ¹¹Universitat Pompeu Fabra, Barcelona, Spain, ¹²CIBER Bioingeniería, Biomateriales y Nanomedicina (CIBER-BBN), Madrid, Spain, ¹³Department of Neurodegenerative Disease, Dementia Research Centre, UCL, London, United Kingdom, ¹⁴Neuroradiological Academic Unit, UCL Queen Square Institute of Neurology, London, United Kingdom, ¹⁵Centre for Medical Image Computing, University College London, London, United Kingdom, ¹⁶Nuclear Magnetic Resonance Research Unit, Queen Square Multiple Sclerosis Centre, University College London Institute of Neurology, London, United Kingdom, ¹⁷Department of Medical Physics and Biomedical Engineering, Centre for Medical Image Computing, University College London, London, United Kingdom, ¹⁸e-Health Centre, Open University of Catalonia, Barcelona, Spain, ¹⁹Helmholtz‐Zentrum Dresden‐Rossendorf, Institute of Radiopharmaceutical Cancer Research, Dresden, Germany, ²⁰IXICO, London, United Kingdom, ²¹Imperial College London, London, United Kingdom, ²²Takeda Pharmaceuticals Ltd., Cambridge, MA, United States, ²³Université de Normandie, Unicaen, Inserm, U1237, PhIND "Physiopathology and Imaging of Neurological Disorders", institut Blood-and-Brain @ Caen-Normandie, Cyceron, Caen, France, ²⁴Department of Nuclear Medicine, Toulouse CHU, Purpan University Hospital, Toulouse, France, ²⁵Toulouse NeuroImaging Center, University of Toulouse, INSERM, UPS, Toulouse, France, ²⁶Centre for Clinical Brain Sciences, The University of Edinburgh, Edinburgh, Scotland, ²⁷Laboratory Alzheimer’s Neuroimaging & Epidemiology, IRCCS Istituto Centro San Giovanni di Dio Fatebenefratelli, Brescia, Italy, ²⁸University Hospitals and University of Geneva, Geneva, Switzerland, ²⁹Centre for Dementia Prevention, The University of Edinburgh, Edinburgh, Scotland, ³⁰UK Dementia Research Institute at Edinburgh, University of Edinburg, Edinburgh, Scotland, ³¹Department of Medicine, Imperial College London, London, United Kingdom, ³²Institute of Neurology and Healthcare Engineering, University College London, London, United Kingdom, ³³Ghent Institute for Functional and Metabolic Imaging (GIfMI), Ghent University, Ghent, Belgium

The neuroimaging community strives to obtain large data cohorts, usually through association within consortia spanning different sites and countries. This results in increased variability of acquisition parameters and scan quality, which can affect image processing and statistical analyses. We propose a semi-automatic data management pipeline to process raw data, assess quality and compute image-derived phenotypes from multi-modal MRI scans, as developed for the multi-centre European Prevention of Alzheimer Dementia longitudinal cohort study (EPAD LCS).

Venkata Veerendranadh Chebrolu¹, Michael Wullenweber², Andreas Schaefer², Johann Sukkau², and Peter Kollasch^1
1Siemens Medical Solutions USA, Inc., Rochester, MN, United States, ²Siemens Healthcare GmbH, Erlangen, Germany

Automated identification of proton spectral characteristics has potential utility in accurate spectral fat saturation, improving dynamic shim routines, and optimizing bandwidth of radiofrequency pulses used in multi-slice or multi-band excitation. In this work, we present an algorithm for automated identification of fat and water proton spectral characteristics and evaluate its performance in 30 proton spectra from breast (number of subjects: n=20), ankle (n=11), and knee (n=9) anatomical regions.

Samah Khawaled¹ and Moti Freiman^2
1Applied Mathematics, Technion, Haifa, Israel, ²Biomedical Engineering, Technion, Haifa, Israel

Unsupervised deep neural networks (DNN) are successfully employed to predict deformation-fields in neuroimaging studies. Bayesian DNN models enable safer utilization of DNN methods in neuroimaging studies, improve generalization and enable assessment of uncertainty in the predictions. We propose a non-parametric Bayesian approach to estimate the uncertainty in DNN-based algorithms for brain MRI deformable registration. We demonstrated the added-value of our Bayesian registration framework on the brain MRI (LPBA40) dataset compared to state-of-the-art VoxelMorph DNN. Further, we quantified the uncertainty of the registration and assessed its correlation with the out-of-distribution data.

Fangxu Xing¹, Riwei Jin², Imani Gilbert³, Xiaofeng Liu¹, Georges El Fakhri¹, Jamie Perry³, Bradley Sutton², and Jonghye Woo^1
1Department of Radiology, Massachusetts General Hospital, Boston, MA, United States, ²Department of Bioengineering, University of Illinois at Urbana-Champaign, Champaign, IL, United States, ³Department of Communication Sciences and Disorders, East Carolina University, Greenville, NC, United States

Visual assessment of articulatory motion of the vocal tract in speech has been drastically advanced by recent developments in high-speed real-time MRI. However, the variation of speaking rate among different subjects prevents quantitative analysis of data from a larger study group. We present a pipeline of methods that aligns audio and image data in the time domain and produces temporally matched image volumes for various subjects. Comparison of the cross-correlation score before and after time alignment showed an increased similarity between source and target image sequences, enabling production of preprocessed multi-subject data for the subsequent statistical atlas construction studies.

Stefan Winzeck^1,2, Ben Glocker¹, Virginia F. J. Newcombe², David K. Menon², and Marta M. Correia^3
1BioMedIA, Department of Computing, Imperial College London, London, United Kingdom, ²Division of Anaesthesia, Department of Medicine, University of Cambridge, Cambridge, United Kingdom, ³MRC Cognition and Brain Sciences Unit, University of Cambridge, Cambridge, United Kingdom

The comparison of TractSeg and JHU atlas-based white matter parcellation on DWI showed the impact of different acquisition schemes on both region volumes and variation in diffusion metrics. TractSeg defined fibre bundles more accurately, but volumes varied for different DWI parameters. The atlas-based segmentation of fibre tracts was more robust to acquisition differences, but showed higher variation in diffusion metrics suggesting a less precise differentiation of white and grey matter.

Emily Koons¹, Eric G Stinson², Patrick Liebig³, Peter Speier³, Krystal Kirby¹, Kirk M Welker¹, and Andrew J Fagan^1
1Radiology, Mayo Clinic, Rochester, MN, United States, ²Siemens Medical Solutions USA, Inc., Rochester, MN, United States, ³Siemens Healthcare GmbH, Erlangen, Germany

A compressed sensing technique was adapted for use in acquiring 3D susceptibility-weighted image data on a clinical 7T MRI system.  The image quality was highly dependent on the choice of regularization parameter used in the reconstruction for all accelerated acquisitions, with some residual artifacts manifest in images thought due to the acquisition of anisotropic voxels.  Nevertheless, CNR and vein detectability was significantly enhanced in images acquired using the CS-SWI technique compared to the clinical SWI protocol, for CS acceleration factors up to 7.2 (corresponding to a 48% scan time reduction compared to the clinical protocol). 

Hendrik Mattern^1
1Biomedical Magnetic Resonance, Otto-von-Guericke-University Magdeburg, Magdeburg, Germany

An easy-to-use and Openly available sMall vEsseL sEgmenTa-Tion pipelinE (OMELETTE) was developed and compared to a Frangi-based benchmark segmentation. Segmentations performance was evaluated for 20 high resolution Time-of-Flight angiograms. Qualitative and quantitative comparison showed OMELETTE's superior segmentation performance.

Warda T. Syeda¹, Vanessa Brait², Alex Oman², Charlotte Ermine ², Jess Nithianantharajah ², Lachlan Thompson², Leigh A. Johnston^3,4, David K. Wright⁵, and Amy Brodtmann^2
1Melbourne Neuropsychiatry Centre, The University of Melbourne, Parkville, Australia, ²The Florey Institute of Neuroscience and Mental Health, The University of Melbourne, Parkville, Australia, ³Department of Biomedical Engineering, The University of Melbourne, Melbourne, Australia, ⁴Melbourne Brain Centre Imaging Unit, The University of Melbourne, Parkville, Australia, ⁵Department of Neuroscience, Central Clinical School, Monash University, Melbourne, Australia

A multicomponent registration technique has been proposed to perform image normalization in the presence of largely deformed ventricles in the mouse brain. By employing multicomponent similarity metrics, the proposed technique combines information from multiple filtered and contrast-enhanced copies of the original images in the optimization process. We apply the proposed method to a mouse brain dataset with enlarged ventricles after focal ischaemic stroke, and compare the performance with single-component registration.

Reza Kalantar¹, Susan Lalondrelle², Jessica M Winfield^1,3, Christina Messiou^1,3, Dow-Mu Koh^1,3, and Matthew D Blackledge^1
1Division of Radiotherapy and Imaging, The Institute of Cancer Research, London, United Kingdom, ²Gynaecological Unit, The Royal Marsden Hospital, London, United Kingdom, ³Department of Radiology, The Royal Marsden Hospital, London, United Kingdom

Radiotherapy (RT) is a cornerstone treatment for cervical cancer. Accurate delineation of organs-at-risk (OARs) is critical to prevent radiation toxicity to healthy organs surrounding the tumour. However, OARs delineation is currently performed manually by clinicians which is labour- and time-intensive. Deep-learning-based algorithms have shown potential in automating this task. This study developed and evaluated a novel framework to incorporate the superior soft-tissue contrast of MRI as well as multi-wavelet image decompositions for improved OARs segmentation on CT images. The proposed framework appears to be a promising addition to the cervical cancer treatment workflow.

Suguru Yokosawa¹, Toru Shirai¹, Hisako Nagao¹, and Hisaaki Ochi^1
1Healthcare Business Unit, Hitachi, Ltd, Tokyo, Japan

Generally, automated scan plane planning methods require the recognition of different landmarks depending on the examination parts. Therefore, algorithms must be tailored to each examination part. In this study, we have proposed a modular algorithm developing method for automatic scan plane planning. In this method, an algorithm is composed of several common processes, and by simply changing the combination of those processes, it can be tailored to different examination parts.

Yuki Matsumoto¹, Masafumi Harada¹, Yuki Kanazawa¹, Nagomi Fukuda², Syun Kitano², Yo Taniguchi³, Masaharu Ono³, and Yoshitaka Bito^3
1Graduate School of Biomedical Sciences, Tokushima University, Tokushima-city, Japan, ²School of Health Sciences, Tokushima University, Tokushima-city, Japan, ³Healthcare Business Unit, Hitachi, Ltd., Tokyo, Japan

In this study, we attempted to characterize brain tumors by combining quantitative parameter mapping and deep-learning-based semantic segmentation. T1, concentration of contrast media (CM), and pHe maps were calculated after the image dataset was obtained. The contrast-enhanced area was then automatically detected using a deep-learning-based semantic segmentation algorithm. The segmented mask was set as the region of interest on these calculated maps. The statistical significance of differences in brain tumors was evaluated to determine whether changes in the mean T1, CM, and pHe were malignancy-dependent.

Neha A Reddy^1,2, Andrew D Vigotsky^1,3, Rachael C Stickland², and Molly G Bright^1,2
1Department of Biomedical Engineering, Northwestern University, Evanston, IL, United States, ²Department of Physical Therapy and Human Movement Sciences, Feinberg School of Medicine, Northwestern University, Chicago, IL, United States, ³Department of Statistics, Northwestern University, Evanston, IL, United States

RETROICOR (Retrospective Image Correction) is a commonly used technique for physiological noise correction, but it may not be optimal in its standard implementation for use in studies of the human spinal cord involving atypical breathing patterns. We test the inclusion of an absolute chest position regressor derived from respiratory belt data, in addition to standard RETROICOR regressors, in modeling fMRI scans of the cervical spinal cord with and without breath-hold challenges. The chest position regressor may be a simple approach towards explaining additional noise variance in spinal cord fMRI data during breath-hold tasks.

Diana M. Marin-Castrillon¹, Arnaud Boucher¹, Siyu Lin¹, Chloe Bernard², Marie-Catherine Morgant^1,2, Alexandre Cochet^1,3, Alain Lalande ^1,3, Benoit Presles ¹, and Olivier Bouchot ^1,2
1ImViA Laboratory, University of Burgundy, Dijon, France, ²Department of Cardio-Vascular and Thoracic Surgery, University Hospital of Dijon, Dijon, France, ³Department of Medical Imaging, University Hospital of Dijon, Dijon, France

Analysis of aorta hemodynamics is useful in the evaluation of aortic diseases. 4D PC-MRI provides information of flow velocity in the aorta and automatic segmentation is one of the biggest challenges. We propose a fully 3D automatic segmentation of the aorta in systole using a multi-atlas approach. Evaluation on 16 patients provided an average performance of 29.55±24.33 mm and 0.859 ±0.024 for Hausdorff distance and Dice score respectively. With the proposed method, the automatic segmentation of the thoracic aorta that can be obtained from 4D PC-MRI is close enough to the manual one to be used in future studies.

Nicolas Pannetier¹, Kaleb Fischer¹, Justin Elhert¹, Ambrus Simon¹, Gunnar Schaefer¹, and Michael Perry^1
1Flywheel Exchange, Inc, Minneapolis, MN, United States

De-identification of medical data is complex and is a barrier for aggregating heterogeneous data repositories. We have developed a flexible open-source Python package for de-identification of medical images and related data. It is fully featured, supports DICOM and 8 other file types and 9 different field transformations. This package is used in production at Flywheel, Inc. We believe this package can contribute to enforce best standards in the protection of patient identity and privacy while relieving researchers from the burdensome task of developing their own custom tooling.