Bryony L. McGarry^1,2, Elizabeth Hunter¹, Robin A. Damion², Michael J. Knight², Philip L. Clatworthy³, George Harston⁴, Keith W. Muir⁵, Risto A. Kauppinen⁶, and John D. Kelleher^1
1PRECISE4Q Predictive Modelling in Stroke, Information Communications and Entertainment Institute, Technological University Dublin, Dublin, Ireland, ²School of Psychological Science, University of Bristol, Bristol, United Kingdom, ³Stroke Neurology, North Bristol NHS Trust, Bristol, United Kingdom, ⁴Acute Stroke Programme, Radcliffe Department of Medicine, University of Oxford, Oxford, United Kingdom, ⁵Institue of Neuroscience and Psychology, University of Glasgow, Glasgow, United Kingdom, ⁶Faculty of Engineering, University of Bristol, Bristol, United Kingdom

Unknown onset time is a common contraindication for anti-thrombolytic treatment of ischaemic stroke.T₂ relaxation-based signal changes within the lesion can identify patients within or beyond the 4.5-hour intravenous thrombolysis treatment-window. However, now that intra-arterial thrombolysis is recommended between 4.5 and 6 hours from symptom onset and mechanical thrombectomy is considered safe between 6 and 24 hours, there are three treatment-windows to consider. Here we show a cumulative ordinal regression model, incorporating the T₂ relaxation time, predicts the probabilities of a patient being within one of the three treatment-windows and is more accurate than signal intensity changes from T₂ weighted images. 

Mauro Costagli^1,2, Graziella Donatelli^3,4, Paolo Cecchi^3,4, Gabriele Siciliano^4,5, and Mirco Cosottini^3,4,5
1University of Genova, Genova, Italy, ²IRCCS Stella Maris, Pisa, Italy, ³IMAGO 7 Research Foundation, Pisa, Italy, ⁴Azienda Ospedaliero Universitaria Pisana, Pisa, Italy, ⁵University of Pisa, Pisa, Italy

The Primary Motor Cortex (M1) in Amyotrophic Lateral Sclerosis (ALS) patients and healthy controls (HC) has a different appearance in Quantitative Susceptibility Mapping (QSM) images. The purpose of this study was to identify a set of distribution indices of QSM values in M1 that enable to better classify ALS patients and HC. Taken individually, the mean value, standard deviation, skewness and kurtosis of the QSM value distributions enabled to obtain diagnostic accuracies 0.43 < A < 0.78. When the distribution indices were jointly used in support vector machine (SVM) classifiers, it was possible to achieve a diagnostic accuracy of 0.90.

Takashi Fujiwara¹, Haben Berhane^2,3, Michael Baran Scott³, Zachary King², Michal Schafer⁴, Brian Fonseca⁴, Joshua Robinson³, Cynthia Rigsby^2,3, Lorna Browne⁴, Michael Markl³, and Alex Barker^1,5
1Department of Radiology, Children's Hospital Colorado, University of Colorado Anschutz Medical Campus, Aurora, CO, United States, ²Lurie Children's Hospital of Chicago, Chicago, IL, United States, ³Northwestern University, Evanston, IL, United States, ⁴Children's Hospital Colorado, University of Colorado Anschutz Medical Campus, Aurora, CO, United States, ⁵Department of Bioengineering, University of Colorado Anschutz Medical Campus, Aurora, CO, United States

A convolutional neural network (CNN) is presented to quantify 4D flow MRI-based hemodynamics using automated segmentation of the proximal vasculature. The intent is to reduce time and user variability for cumbersome 4D flow MRI analyses; however, the pediatric setting is challenging given the complex arterial geometry often seen in congenital heart diseases. Multi-site and -vender datasets were used to train a CNN for 3D segmentation. Flow quantification was conducted with the automated segmentations to test if datasets from multiple institutions and vendors improves flow quantification. We found the multi-site approach improved flow measurements in the setting of complex disease.

Priyanka Tupe Waghmare¹, Archith Rajan², Shweta Prasad³, Jitender Saini⁴, Pramod Kumar Pal⁵, and Madhura Ingalhalikar^6
1E &TC, Symbiosis Institute of Technology, Pune, India, ²Symbiosis Centre for Medical Image Analysis, Symbiosis Centre for Medical Image Analysis, Pune, India, ³Department of Clinical Neurosciences and Neurology, National Institute of Mental Health & Neurosciences, Bangalore, India, ⁴Department of Neuroimaging & Interventional Radiology, National Institute of Mental Health & Neurosciences, Bangalore, India, ⁵Department of Neurology, National Institute of Mental Health & Neurosciences, Bangalore, India, ⁶Symbiosis Center for Medical Image Analysis and Symbiosis Institute of Technology, Pune, India

Parkinson’s disease (PD), multiple system atrophy (MSA), and progressive supra-nuclear palsy (PSP) are neurodegenerative disorders which have parkinsonism as a core clinical feature. In the early stages PD and atypical parkinsonian syndrome (APS) (MSA and PSP) may often be indistinguishable and differential diagnosis is therefore crucial. Our work employs radiomics based features extracted from standard T1 weighted MRI images that are used in a machine learning framework to differentiate PD from APS. Results demonstrate a superior test accuracy of 92% that support our underlying hypothesis that radiomics on T1-weighted images can provide a discriminatory feature space between PD and APS.

Shuai Shen^1,2,3,4, Xiong Yang⁵, Jin Fang⁶, Guihua Jiang⁶, Shuheng Zhang⁵, Yanqun Teng⁵, Xiaomin Ren⁵, Lele Zhao⁵, Jiayu Zhu⁵, Qiang He⁵, Hairong Zheng^1,3,4, Xin Liu^1,3,4, and Na Zhang^1,3,4
1Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhen, China, ²College of Software, Xinjiang University, Urumqi,, China, ³Key Laboratory for Magnetic Resonance and Multimodality Imaging of Guangdong Province, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhen, China, ⁴CAS key laboratory of health informatics, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhen, China, ⁵Shanghai United Imaging Healthcare Co., Ltd., shanghai, China, ⁶Department of Radiology, Guangdong Second Provincial General Hospital, guangdong, China

A total of 124 patients were included in this study. We used U-net neural network architecture to segment the arterial vessel wall on original acquired MR vessel wall images and the corresponding images reconstructed from undersampled K-space data. The Dice coefficients based on the original K-space data, the K-space data with a sampling rate of 7.7%, and K-space data with a sampling rate of 1.9% were 88.66%, 88.19%, and 87.66%, respectively. The effectiveness of arterial vessel wall segmentation on undersampled images using U-net network was verified. The result demonstrated the potential to improve the acceleration performance of MR imaging.

Wallace Souza Loos^1,2, Roberto Souza^2,3, Linda Andersen^1,2, R. Marc Lebel^2,4, and Richard Frayne^1,2
1Radiology and Clinical Neuroscience, Hotchkiss Brain Institute, University of Calgary, Calgary, AB, Canada, ²Seaman Family MR Research Centre, Foothills Medical Centre, Calgary, AB, Canada, ³Electrical and Computer Engineering, Hotchkiss Brain Institute, University of Calgary, Calgary, AB, Canada, ⁴General Electric Healthcare, Calgary, AB, Canada

Dynamic contrast-enhanced magnetic resonance (DCE-MR) imaging is an important clinical tool for investigating the cerebrovascular microcirculatory systems including the blood-brain barrier through estimation of perfusion and permeability maps. However, a vascular function is needed to generate these maps. The estimation is usually performed manually, potentially leading to error, and is also time-consuming. In this work, we designed a deep learning model that leverages the temporal and spatial information from a time series of DCE-MR images to estimate the vascular function automatically. Our model was able to generalize well for unseen data and achieved good overall performance.

Sung-han Lin¹, Chih-Chien Tsai¹, Yi-Chun Chen^2,3, and Jiun-Jie Wang^1
1Department of Medical Imaging and Radiological Sciences, Chang-Gung University, TaoYuan, Taiwan, ²Department of Neurology, Chang Gung Memorial Hospital Linkou Medical Center, TaoYuan, Taiwan, ³College of Medicine, Chang Gung University, TaoYuan, Taiwan

Improvement of diagnosis in patients with Mild Cognitive Impairment (MCI) would help in the early diagnosis of Alzheimer's Disease (AD), since MCI was known as the transition state from normal aging to AD. We developed a novel feature extraction algorithm, which was based on the disease clinical-pathological understanding and combined with the spatial information between disease affected pattern and its surrounding regions. This novel feature set demonstrated improved diagnostic performance, compared to the conventional feature set, especially for the MCI patients. In addition, the identified disease affected pattern corresponded with the postmortem pathology of amyloid deposition in AD patients.  

Nate Tran^1,2, Tracy Luks¹, Devika Nair¹, Angela Jakary¹, Yan Li¹, Janine Lupo¹, Javier Villanueva-Meyer¹, Nicholas Butowski³, Jennifer Clarke³, and Susan Chang^3
1Department of Radiology & Biomedical Imaging, University of California, San Francisco, SAN FRANCISCO, CA, United States, ²UCSF/UC Berkeley Graduate Program in Bioengineering, SAN FRANCISCO, CA, United States, ³Department of Neurological Surgery, University of California, San Francisco, SAN FRANCISCO, CA, United States

This study evaluates the predictive power of multi-parametric MRI at pre-therapy and mid-RT time points in predicting progression-free and overall survival of patients with glioblastoma (GBM). We trained and tested random forest models using metabolic, perfusion, and diffusion images at both preRT and midRT scans, and found that not confining these metrics to the anatomical lesion boundaries improved outcome prediction. The CEL volume mid-RT and type of treatment were among the most important features in predicting PFS, while the T2L volume and metabolic metrics at pre-RT were more relevant for OS prediction.

Mario Serrano-Sosa¹, Jared Van Snellenberg², and Chuan Huang^2,3
1Biomedical Engineering, Stony Brook University, Stony Brook, NY, United States, ²Psychiatry, Renaissance School of Medicine at Stony Brook University, Stony Brook, NY, United States, ³Radiology, Renaissance School of Medicine at Stony Brook University, Stony Brook, NY, United States

Although deep learning algorithms are a novel method for neuroimage analysis, at times they are used as a “black box” for classification task. Therefore, it is crucial to develop methods to comprehend the abstract features used for prediction. We have developed an interpretable deep learning algorithm to predict working memory scores from fMRI data; wherein prediction performance was compared to Kernel Ridge Regression, a traditional machine learning approach. Across all metrics of evaluation, our method outperformed KRR. Moreover, our method was able to create averaged saliency maps highlighting regions most predictive of working memory scores.

Daphne Hong¹ and Yunyan Zhang^1
1University of Calgary, Calgary, AB, Canada

Deep learning is becoming increasingly important in medical imaging analysis, but the ability to interpret deep learning models still lag behind. Here, based on a promising method, gradient-weighted class activation mapping (Grad-CAM), we developed new approaches to interpret arbitrary layers of a convolutional neural network (CNN). Further, using two common CNN models trained to classify brain MRI scans into 3 types, we demonstrated the promise of our new strategy. Characterizing features at low and high levels of a CNN may provide new biomarkers and new insight into disease mechanisms, deserving further validation. 

Sanjay Purushotham¹, Ashwathy Samivel Sureshkumar¹, Li Jiang², Shiyu Tang², Steven Roys², Chandler Sours Rhodes^2,3, Rao P. Gullapalli², and Jiachen Zhuo^2
1Department of Information System, University of Maryland, Baltimore County, Baltimore, MD, United States, ²Diagnostic Radiology and Nuclear Medicine, University of Maryland School of Medicine, Baltimore, MD, United States, ³National Intrepid Center of Excellence, Walter Reed National Military Medical Center, Bethesda, MD, United States

Mild traumatic brain injury (mTBI) patients account for more than 70% of all TBI, with a subset experiencing persistent post concussive symptoms more than 3 months post injury. Patients clinical presentation and conventional imaging findings acutely post injury often lack the ability to predict chronic outcome. In this study, we present a novel method for latent feature extraction from volumetric MRI. We show that with machine learning methods, these acute latent volumetric MRI features are able to improve our symptom prediction in mTBI patients at 18 months post injury.

René Schranzer^1,2, Steffen Bollmann³, Simon Hametner², Christian Menard¹, Siegfried Trattnig⁴, Fritz Leutmezer², Paulus Stefan Rommer², Thomas Berger², Assunta Dal-Bianco², and Günther Grabner^1,2,4
1Department of Medical Engineering, Carinthia University of Applied Sciences, Klagenfurt, Austria, ²Department of Neurology, Medical University of Vienna, Vienna, Austria, ³School of Information Technology and Electrical Engineering, The University of Queensland, Brisbane, Australia, ⁴Department of Biomedical Imaging and Image-guided Therapy, High Field Magnetic Resonance Centre, Vienna, Austria

In multiple sclerosis (MS) the presence of paramagnetic iron rim lesions has been shown to be indicative for progression with a more severe disease course. Our goal was to develop a pipeline based on neural networks to automatically detect, segment and classify lesions as either non-iron or iron loaded using multi-contrast 7T MRI data. A patch-based approach with two modified u-net architectures was used for segmentation and classification. Automatic, high quality lesion segmentation and their classification based on the presence or absence of iron-rims is enabled using convolutional neural networks.

Hanyu Wei¹, Le He¹, Rongsong Zhou², Shuo Chen¹, Miaoqi Zhang¹, Wenwen Chen¹, Xuesong Li³, Yu Ma², and Rui Li^1
1Center for biomedical imaging research, Tsinghua University, Beijing, China, ²Department of Neurosurgery, Tsinghua University Yuquan Hospital, Beijing, China, ³School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China

Deep brain stimulation(DBS) is routine treatment in Parkinson’s disease(PD) but its outcome varies among patients. It’s of benefit to investigate predictive value of pre-operative MR perfusion patterns in DBS outcome. In this study, regional cerebral blood flow(CBF) were measured by ASL both medication “on” and “off” before DBS surgery. Machine learning models were applied to select features and to predict the DBS outcome. Pre-operative MR perfusion patterns shows valuable to predict the DBS outcome in de novo PD patients. The Ridge regression performed best with good correlation(r=0.87，p<0.001) and good agreement with the measured DBS outcome at error level of 9.0%.

Kellen Mulford¹, Sean Moen², Andrew W. Grande², Donald R. Nixdorf³, and Pierre-Francois Van de Moortele^1
1Center for Magnetic Resonance Imaging, University of Minnesota, Minneapolis, MN, United States, ²Department of Neurosurgery, University of Minnesota, Minneapolis, MN, United States, ³Department of Diagnostic and Biological Science, University of Minnesota, Minneapolis, MN, United States

Due to a lack of objective markers, trigeminal neuralgia is difficult to diagnose and classify. This difficulty results in frequent pain recurrences following medication and invasive therapies. In this work, we show that a radiomics based model with features extracted from MRI images can be used to identify painful nerves. Using a cohort of trigeminal neuralgia patients, our predictive model achieves an accuracy of 78% and an AUC of 0.84 in distinguishing between nerves affected and nonaffected by trigeminal neuralgia. 

SUBIN ERATTAKULANGARA¹, KARTHIKA KELAT², JUNJIE LIU³, and SAJAN GOUD LINGALA^1,4
1Roy J Carver Department of Biomedical Engineering, University of Iowa, Iowa City, IA, United States, ²Government Engineering College Kozhikode, Kozhikode, India, ³Department of Neurology, University of Iowa, Iowa City, IA, United States, ⁴Department of Radiology, University of Iowa, Iowa City, IA, United States

We propose a stacked U-NET architecture to automatically segment the tongue, velum, and airway in speech MRI based on hybrid learning. Three separate U-nets are trained to learn the mapping between the input image and their specific articulator. The two U-NETs to segment the velum, and tongue are based on transfer learning, where we leverage open-source brain MRI segmentation. The third U-NET for airway segmentation is based on classical training methods. We demonstrate the utility of our approach by comparing against manual segmentations.  

Yibo Dan¹, Hongyue Tao², Yida Wang¹, Chengxiu Zhang¹, Chenglong Wang¹, Shuang Chen², and Guang Yang^1
1Shanghai Key Laboratory of Magnetic Resonance, East China Normal University, shanghai, China, ²Department of Radiology, Huashan Hospital, Fudan University, shanghai, China

Radiomics has been used widely in image-based diagnosis and prognosis. Since radiomics studies often involve a small number of samples, effective dimension reduction and feature selection are crucial to the successful modeling. In this study, we proposed a heuristic method for effective dimension reduction and feature selection, which built a scout model for each category of features to select features from the category for the final model building. The approach was applied to the modeling with two different datasets, including the BraTS 2019 open data, and achieved results better than those of traditional methods on both datasets.

Margaux Verdier^1,2, Justine Belko¹, Jeremy Deverdun¹, Nicolas Menjot de Champfleur^1,3, Thomas Troalen², Bénédicte Maréchal^4,5,6, Emmanuelle Le Bars¹, and Till Huelnhagen^4,5,6
1I2FH , Neuroradiology, CHU Montpellier, Montpellier University, France, Montpellier, France, ²Siemens Healthcare, Saint Denis, France, ³Laboratoire Charles Coulomb, University of Montpellier, France, Montpellier, France, ⁴Advanced Clinical Imaging Technology, Siemens Healthcare AG, Lausanne, Switzerland, ⁵LTS5, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland, ⁶Radiology Department, Lausanne University Hospital and University of Lausanne, Switzerland, Lausanne, Switzerland

Tumor growth exceeding 8mm/year is the main indication for surgical intervention in low-grade gliomas (LGG). As manual growth assessment is very time-consuming, automated segmentation is desirable. We trained a Convolutional Neural Network (CNN) to segment LGG on 277 MRI-exams (T1+T2-FLAIR) and tested its performance on 9 unknown exams. The mean Dice Similarity Coefficient for automated segmentation was 0.72. The algorithm correctly segmented low T1 and high FLAIR values but tended to underestimate heterogeneous gliomas. Results were independent of cavity or tumor volume. Automated segmentation using CNNs seems promising for clinical practice. Performance might be improved using 3D FLAIR sequences.

Ozan Genc¹, Sivakami Avadiappan¹, Yicheng Chen², Christopher Hess¹, and Janine M. Lupo^1
1Department of Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States, ²Facebook Inc., Mountain View, CA, United States

This study trains a deep convolutional neural network to learn the SWI contrast from T2*-weighted magnitude images using a LSGAN deep learning model and assesses the performance of the resulting network on CMB detection. Our predicted SWI images were able to improve CMB contrast over T2* magnitude images and minimize residual artifacts from standard SWI processing.

Xiaoyuan Hou^1,2, Hui Zhang^1,2, Yan Tan³, Zhenchao Tang^1,2, Hui Zhang³, and Jie Tian^1,2
1Beijing Advanced Innovation Center for Big Data-Based Precision Medicine(BDBPM) ,Beihang University,100083, Beijing, China, ²Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences,100190, Beijing, China, ³Department of Radiology, First Clinical Medical College, Shanxi Medical University,030001, Taiyuan, China

In order to preoperatively predict the multiple genotype mutation for gliomas, we proposed an end-to-end multi-task deep learning model based on MR images analysis for simultaneously predicting IDH and MGMT mutation. Best-performed model was obtained by changing the number of sharing layers in the network, achieving accuracy of 79.78% for MGMT, 78.88% for IDH in the test dataset. Our results indicated that multi-task deep learning model provided a potential solution for simultaneously prediction of multiple genotype in gliomas.