Yael Jacob¹, Gaurav Verma¹, Lara Marcuse¹, Madeline Fields¹, and Priti Balchandani^1
1Icahn School of Medicine at Mount Sinai, New York, NY, United States

Epilepsy patients (EP) endure harmful effects both on their health and quality of life. Early identification of these individuals would be incredibly helpful to gauge management and expectations. Implementing a novel multilayer network analysis, considering communication within functional and structural networks as well as the interactions between them, we tested whether this whole-brain comprehensive network hierarchy can be used as predictors of epilepsy. Using multilayer network features as predictors in a machine learning algorithm we were able to classify EP and controls with overall accuracy of 84%, demonstrating the applicability of multi-modal imaging for diagnostics of epilepsy.

Ke Lei¹, Shreyas Vasanawala², and John Pauly^1
1Electrical Engineering, Stanford University, Stanford, CA, United States, ²Radiology, Stanford University, Stanford, CA, United States

We proposed a CNN model that automatically assesses image quality within seconds after a scan is finished to reduce the number of patient recalls and inadequate images. Our model is deployed to the clinics where it alerts technicians to take action for low-quality images while the patient is still in the scanner. Our model achieves super-human performance on assessing perceptual noise level in 2D fast spin echo (FSE) MRIs. It can also be used to automatically guide other computational processes, like training of a denoising model or choice of a regularization weight for reconstruction. 

Jun Yang^1,2, Cailei Zhao³, Shi Su⁴, Zhanqi Hu⁵, Jianxiang Liao⁵, Dong Liang^1,2,4, and Haifeng Wang^2,4
1Research Centre for Medical AI, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, ²University of Chinese Academy of Sciences, Beijing, China, ³Department of Radiology, Shenzhen Children’s Hospital, Shenzhen, China, ⁴Paul C. Lauterbur Research Centre for Biomedical Imaging, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, ⁵Department of Neurology, Shenzhen Children’s Hospital, Shenzhen, China

Predicting epilepsy drug treatment outcome is important for treating children with tuberous sclerosis complex (TSC). Here, the best performing model was selected to explore the contribution of the features, using permutation importance (PIMP). An approach similar to PIMP was used to compare the magnetic resonance imaging (MRI) and non-MRI features. The best model multilayer perceptron (MLP) with a hidden layer size of 60 and 30 features selected by F-test achieved the best performance. The results based on 103 children patients showed that some features were more important than others, and MRI features contributed more than non-MRI features in prediction.

Alexander German¹, Angelika Mennecke¹, Jan Martin², Jannis Hanspach¹, Andrzej Liebert¹, Jürgen Herrler¹, Tristan Anselm Kuder³, Manuel Schmidt¹, Armin Nagel¹, Michael Uder¹, Arnd Dörfler¹, Jürgen Winkler¹, Moritz Zaiss^1,4, and Frederik Laun^1
1University Hospital Erlangen, Erlangen, Germany, ²Lund University, Lund, Sweden, ³German Cancer Research Center, Heidelberg, Germany, ⁴Max Planck Institute for Biological Cybernetics, Tübingen, Germany

We acquired diffusion-weighted and CEST data of the brain of 38 healthy volunteers. A MPRAGE and SWI based segmentation into 102 brain regions revealed unique diffusion and chemical MR signals on average. More importantly, we could infer these tissue classes form individual voxel data using a neural network. The revival of this old paradigm for tissue characterization from the 1990s points to the fact that unique MR signals of different brain regions exist and can be used to determine the tissue type voxel-wise. The approach as such is general and could unify the ever-growing diversity of MR contrasts.

Jian Wang¹, Guohui Ruan^2,3, Yingjie Mei⁴, Yanjun Chen¹, Jialing Chen¹, Yanqiu Feng^2,3, and Xiaodong Zhang^1
1Department of Medical Imaging, The Third Affiliated Hospital of Southern Medical University, Guangzhou, China, ²Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, Guangzhou, China, ³School of Biomedical Engineering, Southern Medical University, Guangzhou, China, ⁴China International Center, Philips Healthcare, Guangzhou, China

Accurate regional segmentation of the Lumbosacral Plexus (LSP) on magnetic resonance neurography (MRN) images is a fundamental requirement before LSP related disorders diagnosis can be achieved. In this paper, we utilize U-Net to segment LSP trunk and branch from three-dimensional fast field echo(3D-FFE) with principle of selective excitation technique (Proset) images. The results show that a U-Net deep learning framework expresses highly performance and less time-consumption for LSP segmentation in patients with degenerative spinal diseases and healthy subjects.

Rupsa Bhattacharjee^1,2, Rakesh Kumar Gupta³, Suhail P Parvaze⁴, Rana Patir⁵, Sandeep Vaishya⁵, Sunita Ahlawat⁶, and Anup Singh^1,7
1Center for Biomedical Engineering, Indian Institute of Technology (IIT) Delhi, New Delhi, India, ²Philips Health Systems, Philips India Limited, Gurugram, India, ³Department of Radiology, Fortis Memorial Research Institute, Gurugram, India, ⁴Philips Health Systems, Philips Innovation Campus, Bangalore, India, ⁵Department of Neurosurgery, Fortis Memorial Research Institute, Gurugram, India, ⁶SRL Diagnostics, Gurugram, India, ⁷Department of Biomedical Engineering, All India Institute of Medical Sciences, New Delhi, India

Intra-tumoral-susceptibility-signal (ITSS) has been increasingly proven to play a major role in glioma grading, progression assessment and follow-up. Quantitative ITSS assessment involves segmentation of ITSS from SWI images, separating vasculature ITSS from hemorrhage ITSS and finally quantifying the ITSS-vasculature-volume (IVV) to grade the glioma non-invasively. This study involves radiomic feature extraction, random-forest based feature selection and classification to indicate that radiomic features can significantly differentiate between 3D_vasculature and 3D_Hemorrhage mask regions in SWI-magnitude images. This is also one of the first studies that explores the vasculature and hemorrhage radiomic properties extracted from SWI-magnitude images through machine-learning in grade-IV GBM patients.

Roberto Souza^1,2, Jordan McEwen³, Carissa Chung³, and Ashley D. Harris^2,4
1Electrical and Computer Engineering, University of Calgary, Calgary, AB, Canada, ²Hotchkiss Brain Institute, Calgary, AB, Canada, ³Biomedical Engineering, University of Calgary, Calgary, AB, Canada, ⁴Radiology, University of Calgary, Calgary, AB, Canada

Magnetic Resonance Spectroscopy (MRS) non-invasively acquires in-vivo data on the chemical composition of localized tissue samples. MRS acquisitions are lengthy because they often require the acquisition of several averages to obtain a spectrum with a sufficient signal to noise ratio (SNR). This issue is augmented in J-difference edited MRS in which the analyzed spectrum is generated as the difference between sub-spectra in which editing pulses have been applied to selectively refocus the coupling of the target metabolite. In this work, we investigate the reduction of J-difference edited MRS acquisition times using deep learning.

Zuojun Wang¹, Peng Xia¹, Wenming Cao², Kui Kai, Gary Lau¹, Henry Ka Fung Mak³, and Peng Cao^1
1Diagnostic Radiology, Department of Diagnostic Radiology, HKU, Hong Kong, China, ²Department of Diagnostic Radiology, HKU, HongKong, China, ³Department of Diagnostic Radiology, HKU, Hong Kong, China

MRI anonymizations, including face removal, are necessary for clinical data archiving and sharing. Segmentation based methods have been developed for semi-automated face removal on brain MRI. Meanwhile, the conventional methods are inefficient and unreliable, as the images have to be pre-processed and fed in the software manually. Deep learning-based methods are highly efficient in image-to-image translation on large scale databases. In this study, we utilized a cycle generative adversarial network to anonymize brain MRI data. The model showed reliable performance when testing on T1-weighted images, and we also extend it to the unseen MPRAGE images, targeting different brain MRI contrasts.

Farnaz Orooji¹, Xinyang Wang¹, Mohammed Ayoub Alaoui Mhamdi¹, and Russell Butler^1
1Computer Science, Bishop's University, Sherbrooke, QC, Canada

One of the main strengths of MRI is the wide range of soft tissue contrasts which can be obtained using different sequence parameters. T1-weighted-images provide exquisite contrast between gray/white matter but fail to capture arterial vessels. Here we experiment with deep-learning for image synthesis, to synthesize a time-of-flight-(TOF) angiogram based on T1 contrast image. We then compare arterial diameters from synthesized TOF with ground-truth-TOF diameters. We also synthesize a T1 from a T2, and compare cortical metrics such as thickness, curvature etc. We show that it is possible to obtain vessel diameters from T1, and cortical thick/vol/curv measures from T2.  

Rahul Gaurav^1,2,3, Romain Valabregue^1,2, Nadya Pyatigorskaya^1,2,3,4, Lydia Yahia-Cherif^1,2, Emma Biondetti^1,2,3, Graziella Mangone^2,5, R. Matthew Hutchison⁶, Jean-Christophe Corvol^2,5,7, Marie Vidailhet^2,3,7, and Stephane Lehericy^1,2,3,4
1CENIR, ICM Paris, Paris, France, ²Paris Brain Institute (ICM), Sorbonne University, UPMC Univ Paris 06, Inserm U1127, CNRS UMR 7225, Paris, France, ³ICM Team “Movement Investigations and Therapeutics” (MOV’IT), Paris, France, ⁴Department of Neuroradiology, Pitié-Salpêtrière Hospital, AP-HP, Paris, France, ⁵INSERM, Clinical Investigation Center for Neurosciences, Pitié-Salpêtrière Hospital, Paris, France, ⁶Biogen Inc., Cambridge, MA, United States, ⁷Department of Neurology, APHP, Pitié-Salpêtrière Hospital, Paris, France

There is a need of accurate imaging biomarkers of dopaminergic cell neurodegeneration to facilitate drug trials in Parkinson’s disease (PD). PD demonstrates neurodegenerative substantia nigra pars compacta (SNc) changes that can be detected efficiently using neuromelanin-sensitive MRI. Characterizing neuromelanin signal variations using manual SNc segmentation is an operator-dependent and time-consuming task. Hence, in this cross-sectional, observational, case-control study, we investigated neuromelanin SNc abnormalities in the early PD patients using convolutional neural network-based fully automatic segmentation of SNc. We found a highly significant difference in SNc volume and signal intensity between early PD and healthy volunteers.

Hui Zhang¹, Junqi Xu¹, Xutong Kuang², Shuai Xu², Xuchen Yu¹, Weibo Chen³, Chengyan Wang², and He Wang^1
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, ²Human Phenome Institute, Fudan University, Shanghai, China, ³Philips Healthcare, Shanghai, China

This study aimed to propose a reliable deep learning based high resolution intravoxel incoherent motion (IVIM) parameter mapping for the assessment of lacunar infarction patients.

Maira Siqueira Pinto^1,2, Stefan Winzeck^3,4, Marta M. Correia⁵, Evgenios N. Kornaropoulos^4,6, David K. Menon⁴, Ben Glocker³, Arnold J. den Dekker², Jan Sijbers², Pieter-Jan Guns⁷, Pieter Van Dyck¹, and Virginia F. J. Newcombe^4
1Radiology, UZA - Antwerp University Hospital, Antwerpen, Belgium, ²imec-Vision Lab, University of Antwerp, Antwerpen, Belgium, ³BioMedIA Group, Department of Computing, Imperial College London, London, United Kingdom, ⁴Division of Anaesthesia, Department of Medicine, University of Cambridge, Cambridge, United Kingdom, ⁵MRC Cognition and Brain Sciences Unit, University of Cambridge, Cambridge, United Kingdom, ⁶Clinical Sciences, Diagnostic Radiology, Lund University, Lund, Sweden, ⁷Physiopharmacology, University of Antwerp, Antwerpen, Belgium

Over 40% of patients after a mild traumatic brain injury (mTBI) may have persisting symptoms. This study investigated a Support Vector Machine (SVM) approach for outcome prediction after mTBI from multi-modal MRI. The datasets included 77 mTBI patients from the CENTER-TBI study with acute T2w, SWI, FA and MD scans and outcome scores six month post-injury. Benefits of data harmonization were tested and Z-scoring reduced site-specific biases yielding 67.7% prediction accuracy. Our data-driven approach revealed that predictive signal was retrieved mainly from diffusion maps rather than conventional images, and was located in the superior fronto-occipital fascicle and the corticospinal tract.

Zuojun Wang¹, Peng Xia¹, Henry Ka Fung Mak², and Peng Cao^1
1Diagnostic Radiology, Department of Diagnostic Radiology, HKU, Hong Kong, China, ²Department of Diagnostic Radiology, HKU, Hong Kong, China

Quantitative susceptibility mapping (QSM) obtained from the MRI phase images is valuable in neurological disease diagnoses. Meanwhile, the role of thumb MRI scan probing susceptibility contrast is susceptibility weighting imaging (SWI), which might contain blooming artifacts that would affect the hypointensity appearance. Many conventional methods have been developed for QSM reconstruction, including the deep learning-based approach that is applicable in clinical diagnoses. Here, we apply the cycle generative adversarial network with a perceptual loss to synthesize QSM images from SWI images. The predicted QSM images showed their application in brain microbleed detection.

Yi Duan¹, Yida Wang¹, Naying He², Yan Li², Zenghui Cheng², Yu Liu², Zhijia Jin², Pei Huang³, Shengdi Chen³, Ewart Mark Haacke^2,4, Fuhua Yan², and Guang Yang^1
1East China Normal University, Shanghai Key Laboratory of Magnetic Resonance, Shanghai, China, ²Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, ³Department of Neurology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, ⁴Department of Biomedical Engineering, Wayne State University, Detroit, MI, United States

Diagnosing Parkinson’s disease (PD) is still a clinical challenge. Deep grey matter is involved in the pathophysiological changes of PD. We built a radiomics model to distinguish PD from normal controls (NC) based on five brain nuclei in multiple quantitative images derived from STrategically Acquired Gradient Echo (STAGE) imaging. This model combined features from the caudate nucleus, globus pallidus, putamen, red nucleus, and substantia nigra regions in QSM, T1and proton density maps and achieved a test AUC of 0.948. Features from the SN region as seen in the QSM images were found to be the most important ones for classification.

Quan Dou¹, Xue Feng¹, Sohil Patel², and Craig H. Meyer^1
1Biomedical Engineering, University of Virginia, Charlottesville, VA, United States, ²Radiology & Medical Imaging, University of Virginia, Charlottesville, VA, United States

Non-invasive MRI-based survival prediction for glioblastoma patients is potentially valuable for informing prognostic and treatment counseling. In this study, we analyzed the relationships between overall survival and several automatic segmentation-based MR imaging features. Simple logistic regression models to classify 1-year survival with clinical factors and selected imaging features were trained and tested. Results showed that combining imaging features with clinical factors improved the survival prediction.

Thomas Lilieholm¹, Matt Henningsen², Azam Ahmed³, Alan McMillan^1,4, and Walter F Block^1,4,5
1Medical Physics, University of Wisconsin at Madison, Madison, WI, United States, ²Electrical Engineering, University of Wisconsin at Madison, Madison, WI, United States, ³Neurological Surgery, University of Wisconsin at Madison, Madison, WI, United States, ⁴Radiology, University of Wisconsin at Madison, Madison, WI, United States, ⁵Biomedical Engineering, University of Wisconsin at Madison, Madison, WI, United States

Previous work has shown that minimally-invasive reduction of hematoma volume in intracerebral hemorrhage to a threshold of 15mL is indicative of improved long term patient outcome. To attain this goal, image-guided minimally-invasive surgical techniques are applied to both lyse clot material and drain from the site of hemorrhage via a porous catheter. We propose a Convolutional Neural Network to identify and autonomously segment clot and peripheral edema in MR images of the brain for volumetric analysis, and image-guidance during evacuation. Quantitative measurements produced in this way can be used for superior clot visualization and direct measurement of remaining clot volume.

Ritu Lahoti¹, Lakshay Agarwal¹, Neelam Sinha¹, and Vinod Reddy^1
1International Institute of Information Technology Bangalore (IIITB), Bengaluru, India

Brain tumor segmentation is one of the challenging image segmentation problem. Among the various architectures of CNN used for brain tumor segmentation from MRI images, we compare multi-view 2D CNN and multi-view 3D anisotropic CNN on the popular BraTS dataset. We computed four metrics to measure their performance on 10 test data. The first approach showed a mean sensitivity of 0.816 whereas second approach outperforms with mean sensitivity of 0.9217. Other metrics for both approaches achieved comparable results. Although both approaches consider all orthogonal planes, anisotropic CNN takes into account both global and local features efficiently thereby giving better results.

Peter Hsu¹, Sindhuja Govindarajan¹, Nikhil Chettipally¹, Lev Bangiyev², Robert Peyster², Giuseppe Cruciata², Patricia Coyle², Haifang Li², Hasan Saffiudin¹, Ryan Merritt¹, Eric Wei¹, Almighty Ironnah¹, and Kwan Chen^1
1Stony Brook University, Stony Brook, NY, United States, ²Stony Brook University Hospital, Stony Brook, NY, United States

Multiple Sclerosis lesions in the spinal cord are associated with more debilitative disease outcomes and have predictive value for prognosis and diagnosis. However, these lesions are difficult to detect from MRI scans and this process is susceptible to inter-rater and intra-rater variability. Machine Learning techniques have the ability to assist in this problem. We propose a Convolutional Neural Network that can perform accurate identification and segmentation of MS lesions in the spinal cord. This method achieves high overlap with the segmentations of attending radiologists and is robust to imaging artifacts, showcasing the potential to be a tool for clinical practice.

Maxime DRAI¹, GILLES BRUN¹, Nadine GIRARD^1,2, Benoit TESTUD^1,3, and Jan-Patrick STELLMANN^1,3
1Neuroradiology, APHM, Marseille, France, ²CRMBM-CEMEREM, Aix-Marseille Université, Marseille, France, ³CNRS, CRMBM-CEMEREM, UMR 7339, Aix-Marseille Université, Marseille, France

AI brain tumor segmentation and brain extraction algorithms are an essential step in image processing, however they are mainly developed in adults. Here, we aimed to explore the usability of these algorithms in a heterogenous pediatric population. In 42 brain pediatric tumor MRI, we compared manual mask with mask generated by the algorithms. Results were excellent for brain extraction, moderate for segmentation of contrast-enhancing tumors, and weak for non-enhancing T2-signal abnormalities. Some improvements are necessary to adapt this algorithm to pediatric brain tumors. However, borrowing strength from adults might be a feasible approach for AI implementation in rare pediatric populations.

Atsuro Suzuki¹, Chizue Ishihara¹, Yukio Kaneko¹, Tomoki Amemiya¹, Yoshitaka Bito¹, and Toru Shirai^1
1Healthcare Business Unit, Hitachi, Ltd., Kokubunji-shi, Japan

To reduce inhomogeneous noise caused by parallel imaging, we developed a deep-learning-based noise reduction method that incorporates spatial distribution of noise. For noise distribution we used a g-factor map segmented into high and low g-factor regions. We reduced the noise by using a different optimized network in each region. Finally, a denoised image was generated by combining the two denoised regions. Denoised brain images demonstrated improved signal to noise ratio (SNR) and mean square error (MSE) between denoised and full sampling images throughout the brain regions. Our method was able to reduce the inhomogeneous noise proportional to the noise intensity.