Francesco La Rosa^1,2, Ahmed Abdulkadir³, Mário João Fartaria^1,2,4, Reza Rahmanzadeh^5,6, Riccardo Galbusera^5,6, Jean-Philippe Thiran^1,2, Cristina Granziera^5,6, and Meritxell Bach Cuadra^1,2
1LTS5, École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland, ²Radiology Department, Center for Biomedical Imaging, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ³Universitäre Psychiatrische Dienste and University of Bern, Bern, Switzerland, ⁴Siemens Healthcare AG Switzerland, Lausanne, Switzerland, ⁵Neurologic Clinic and Policlinic, Departments of Medicine, Clinical Research and Biomedical Engineering, University Hospital Basel and University of Basel, Basel, Switzerland, ⁶Translational Imaging in Neurology (ThINk) Basel, Department of Medicine and Biomedical Engineering, University Hospital Basel and University of Basel, Basel, Switzerland

Multiple Sclerosis cortical lesions are not readily visible in conventional MRI, but they are clinically highly relevant and have been recently included in the MS diagnostic criteria. However, advanced MRI sequences such as the MP2RAGE are needed in order to identify them visually. In this work, we propose an automatic method based on a convolutional neural network to automatically detect cortical lesions. In a cohort of 84 patients with FLAIR and MP2RAGE acquisitions our framework achieves a 77% cortical lesion detection rate with a 26% lesion-wise false positive rate.

Wenhao Jiang^1,2, Jian Zhang¹, Peng Cao³, Jing Gu¹, and Silun Wang^1
1YiWei Medical Technology Co.,Ltd, Shenzhen, China, ²School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China, ³Department of Diagnostic Radiology, The University of Hong Kong, Hong Kong, Hong Kong

In this work, deep-learning-based methods were utilized to segment white matter hyperintensities (WMH) on T2-weighted MRI from 213 patients diagnosed with ischemia and lacune. Vulnerability maps of each disease were generated regarding the prevalence of WMH registered to the standard MNI template. The WMH were allocated into 68 regions of interest using a Hammers atlas. Correlation among the region-specific WMH was analyzed for a lesion-symptom study.

Veronica Ravano^1,2,3, Michaela Andelova⁴, Mazen Fouad A-Wali Mahdi¹, Reto Meuli², Tomas Uher⁴, Jan Krasensky⁵, Manuela Vaneckova⁵, Dana Horakova⁴, Tobias Kober^1,2,6, and Jonas Richiardi^1,2
1Advanced Clinical Imaging Technology, Siemens Healthcare, Lausanne, Switzerland, ²Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ³Medical Imaging Processing, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland, ⁴Department of Neurology and Center of Clinical Neuroscience First Faculty of Medicine, Charles University and General University Hospital in Prague, Prague, Czech Republic, ⁵MR unit, Department of Radiology First Facutly of Medicine, Charles University and General University Hospital in Prague, Prague, Czech Republic, ⁶LTS5, Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland

In multiple sclerosis, the correlation between clinical scores and classical radiological metrics is poor (“clinico-radiological paradox”). To improve the prediction of future disease course, we suggest to study structural brain disconnectivity resulting from white matter lesions. We proposed an atlas-based approach to quantify structural disconnectomes without diffusion imaging, as it is typically not part of clinical routine MR protocols for multiple sclerosis. The disconnectome was modelled as a graph where brain regions are vertices and affected connections edges. Our method provides a new representation of brain disconnectivity that enables to stratify multiple sclerosis patients in two groups with different prognosis.

Daniel McClement¹, Jinseo Lee², Glen Pridham², Olayinka Oladosu², Zahra Hosseinpour², and Yunyan Zhang^2
1University of British Columbia, Vancouver, BC, Canada, ²University of Calgary, Calgary, AB, Canada

Transfer learning and greedy layer-wise training are two potential approaches to advance the performance of deep learning, particularly in fields with limited sample size including medical imaging. Taking the advantage of both, we have implemented a novel strategy that allows progressive training with transfer learning using the VGG19 network. Based on clinical MRI of 19 patients only, our approach achieved 88% accuracy in classifying relapsing remitting from secondary progressive multiple sclerosis (MS), 6% greater than training with the traditional approach. This innovative method may help provide new insight into the pathogenesis and progression mechanisms in MS.

Luu-Ngoc Do¹, Byung-Hyun Beak², Seul-Kee Kim³, Hyung-Jeong Yang⁴, Woong Yoon⁵, and Ilwoo Park^5
1Department of Radiology, Chonnam National University, Gwangju, Korea, Republic of, ²Department of Radiology, Chonnam National University Hospital, Gwangju, Korea, Republic of, ³Department of Radiology, Chonnam National University Hwasun Hospital, Gwangju, Korea, Republic of, ⁴Department of Electronics and Computer Engineering, Chonnam National University, Gwangju, Korea, Republic of, ⁵Department of Radiology, Chonnam National University Medical School and Hospital, Gwangju, Korea, Republic of

The purpose of this study was to demonstrate the feasibility of using deep learning algorithms for automatic classification of DWI-ASPECTS from patients with acute ischemic stroke. DWI data from 319 patients with acute anterior circulation stroke were used to train and validate recurrent residual convolutional neural network models for binary task of classifying low- vs high- DWI-ASPECTS. Our model produced the accuracy of 84.9 ± 1.5% and the AUC of 0.925 ± 0.009, suggesting that this algorithm may provide an important ancillary tool for clinicians in a time-sensitive assessment of DWI-ASPECTS from acute ischemic stroke patients.

Simon Hubertus¹, Sebastian Thomas¹, Junghun Cho², Shun Zhang^3,4, Yi Wang^2,3, and Lothar R. Schad^1
1Computer Assisted Clinical Medicine, Heidelberg University, Mannheim, Germany, ²Department of Biomedical Engineering, Cornell University, Ithaca, NY, United States, ³Department of Radiology, Weill Cornell Medical College, New York, NY, United States, ⁴Department of Radiology, Tongji Hospital, Wuhan, China

MRI-based mapping of the oxygen extraction fraction (OEF) is a valuable addition to diagnosis and treatment planning of various diseases; yet, it often lacks robustness and suffers from elaborate, time-consuming reconstructions. We trained an artificial neural network (ANN) on simulated QSM values and qBOLD data, tested it in 7 healthy volunteers and compared it to a standard quasi-Newton approach. The ANN reduced the intersubject variability of OEF by regularizing the reconstruction. Moreover, it lowered the reconstruction time from approximately one hour to one second and removed the necessity of accurate parameter initialization through an additional acquisition.

Nazanin Makkinejad¹, Ashish A. Tamhane², Carles Javierre Petit¹, Arnold M. Evia², David A. Bennett², Julie A. Schneider², and Konstantinos Arfanakis^1,2
1Biomedical Engineering, Illinois Institute of Technology, Chicago, IL, United States, ²Rush Alzheimer's Disease Center, Rush University Medical Center, Chicago, IL, United States

Arteriolar sclerosis is common in the brains of older adults and has been shown to be associated with cognitive decline and dementia. Definitive diagnosis of arteriolar sclerosis is only possible at autopsy. The purpose of this work was to develop an end-to-end deep learning model to predict the presence of severe arteriolar sclerosis from MR images without the need to extract hand-engineered features. The model was developed by combining ex-vivo MRI and pathology data in a large community-based cohort of older adults.

James R Golden¹, Yicheng Chen¹, Melanie A Morrison¹, Kate Nelson¹, and Janine M Lupo^1
1Radiology, UCSF, San Francisco, CA, United States

Cerebral microbleeds (CMBs) are known risk factors of stroke and hemorrhage that can be a marker of cognitive impairment. Although CMBs are easily visualized with susceptibility weighted imaging, they are burdensome to localize and quantify manually. Image processing algorithms based on the radial symmetry transform have previously been used to identify candidate CMBs, and convolutional neural networks have been effective at distinguishing real CMBs from mimics with high sensitivity and specificity. A deep neural network was trained to carry out this entire pipeline and to predict CMB voxel masks using a dataset of radiation therapy-induded CMBs from patients with gliomas.     

Jinwoo Hong^1,2, HyukJin Yun¹, Gilsoon Park², Jong-Min Lee², and Kiho Im^1
1Division of Newborn Medicine, Boston Children's Hospital, BOSTON, MA, United States, ²Hanyang University, Seoul, Korea, Republic of

Quantitative measurements about sulcal patterns are a basic approach to characterize the possibility of neurological disorders. In this study, we used recurrent residual U-Net. Also, to improve the accuracy of segmentation, the models of each plain in the MR image were produced separately and the results combined. As a result, the cortical plate segmentation showed an average dice coefficient of 0.901 ± 0.030, and segmentation for the internal region of the cortical plate showed 0.978 ± 0.010. The proposed method is expected to be useful as a quantitative developmental measurement of the fetus.

Aymen Ayaz¹, Jouke Smink², Tom Geraedts³, Cristian Lorenz⁴, Juergen Weese⁴, and Marcel Breeuwer^1,2
1Biomedical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands, ²MR R&D – Clinical Science, Philips Healthcare, Best, Netherlands, ³MR R&D – Collaboration Office, Philips Healthcare, Best, Netherlands, ⁴Philips Research Laboratories, Hamburg, Germany

A pipeline based on the XCAT phantom, the JEMRIS software for simulating the MR signal and a commercial reconstruction pipeline has been set-up for simulating realistic brain MRI images. Using this pipeline, an anatomically variable brain MRI population is simulated across age and gender. Anatomical variation is generated by means of changing individual brain sizes, and cortical gray matter volumes to mimic aging brain. MS lesions are simulated to mimic diseased brain as well. Significant contrast is generated across detailed brain structures. The commercial reconstruction pipeline increased the realism of simulated data.

Bonian Lu¹ and Gopikrishna Deshpande^1
1AU MRI Research Center, Electrical and Computer Engineering, Auburn University, Auburn, AL, United States

Overfitting, the main issue that constrains the validity and generalizability of machine-learning in neuroimaging-based diagnostic-classification, is in part due to small sample-sizes in relation to what is required for generalization. Even with data aggregation (such as in ABIDE), the relatively smaller sample-sizes are a result of the fact that it is difficult/expensive to acquire data from clinical-populations. With healthy-controls, we have comparatively larger samples available. Therefore, we propose to address overfitting by using larger healthy-samples (HCP) to learn the neural signature of healthy-controls, with the aim of transferring that learning into the context of discriminating Autism from healthy-controls.

Filip Sobczak^1,2, Patricia Pais-Roldán^1,3, Xiaoning Zhao¹, and Xin Yu^1,4
1Translational Neuroimaging and Neural Control Group, High Field Magnetic Resonance Department, Max Planck Institute for Biological Cybernetics, Tuebingen, Germany, ²Graduate Training Centre of Neuroscience, International Max Planck Research School, University of Tuebingen, Tuebingen, Germany, ³Forschungszentrum Jülich, Jülich, Germany, ⁴Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, Charlestown, MA, United States

Lately, we have acquired the resting state fMRI (rs-fMRI) signal with pupillometry from anesthetized rats to investigate specific resting-state network correlations with brain state-specific pupil dynamics. Here we used the acquired data to estimate the instantaneous arousal index based on the rs-fMRI signal. We evaluated predicting pupil dynamics using three methods: linear regression (LR), gated recurrent unit (GRU) neural networks and a previously proposed correlation-based (CC) approach. LR and GRU provided much better predictions than CC method. Also, using weighted PCA components, we can identify specific regions of the brain related to pupil dynamics as the brain state index.

Filip Sobczak^1,2, Yi He^1,3, Terrence J. Sejnowski^4,5, and Xin Yu^1,6
1Translational Neuroimaging and Neural Control Group, High Field Magnetic Resonance Department, Max Planck Institute for Biological Cybernetics, Tuebingen, Germany, ²Graduate Training Centre of Neuroscience, International Max Planck Research School, University of Tuebingen, Tuebingen, Germany, ³Danish Research Centre for Magnetic Resonance, Hvidovre, Denmark, ⁴Howard Hughes Medical Institute, Computational Neurobiology Laboratory, Salk Institute for Biological Studies, La Jolla, CA, United States, ⁵Division of Biological Sciences, University of California, San Diego, CA, United States, ⁶Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital and Harvard Medical School, Charlestown, MA, United States

Previously we developed an echo-state network (ESN) to predict the future temporal evolution of the rs-fMRI slow oscillatory feature from both rodent and human brains. In particular, rs-fMRI signals from individual blood vessels that were strongly correlated with neural calcium oscillations were used to train an ESN to predict brain state-specific rs-fMRI signal fluctuations. Here, the ESN-based predictive model was applied to classify rs-fMRI datasets from the Human Connectome Project (HCP). The ESN enables to decouple the brain state-dependent global rs-fMRI signal fluctuation from the intrinsic activity of the default-mode network.

Yanlin Wang¹, Ping Jiang¹, Shi Tang¹, Lu Lu¹, Xuan Bu¹, Hailong Li¹, Yingxue Gao¹, Lianqing Zhang¹, Lingxiao Cao¹, Jing Liu¹, Xinyue Hu¹, Xinyu Hu¹, Qiyong Gong¹, and Xiaoqi Huang^1
1Huaxi MR Research Center (HMRRC), Functional and molecular imaging Key Laboratory of Sichuan Province, Department of Radiology, West China Hospital, Sichuan University, Chengdu 610041, China, Chengdu, China

It remains unclear whether cortical changes in poor sleep quality are modulated by emotional distress in the general population. An Elastic-Net Regularized Generalized Linear Models based on bootstrapping was implemented to acquire candidate regions for sleep quality. The moderation effects of emotional distress on the relationship between sleep quality and cortical morphometry were acquired by a hierarchical regression. The low level of emotional distress could offset positive correlation between sleep quality and dorsal anterior cingulate cortex (dACC) volume. The result indicates dACC volume enlargement might be a compensatory response to poor sleep quality, or a marker of resilience against recurrent emotional distress.

Ken-Pin Hwang¹, Xinzeng Wang², Marc Lebel², Peter Johansson³, Catharina Petersen³, Marcel Warntjes³, Ersin Bayram², Suchandriam Banerjee², Jingfei Ma¹, and Jason M Johnson^4
1Department of Imaging Physics, The University of Texas M.D. Anderson Cancer Center, Houston, TX, United States, ²MR Applications and Workflow, GE Healthcare, Waukesha, WI, United States, ³SyntheticMR, Linkoping, Sweden, ⁴Department of Radiology, The University of Texas M.D. Anderson Cancer Center, Houston, TX, United States

Images from a multiparameter mapping sequence were reconstructed with a novel deep learning based reconstruction (DL Recon) method trained to remove noise and enhance edges. Mean T1, T2, and PD values as measured in a system phantom differed by less than 0.6% between the DL and conventional reconstructions, while noise was lower in all measurements on DL Recon images. In vivo synthetic images also exhibited reduced noise and increased definition of structures. We find that the SNR and resolution benefits of DL Recon applied to raw MR data extend to improve the fitted relaxation maps and subsequent synthetic images.

Long Wang¹, Suzie Bash², Sara Dupont¹, Sebastian Magda³, Chris Airriess³, Enhao Gong¹, Greg Zaharchuk^1,4, Ajit Shankaranarayanan¹, and Tao Zhang^1
1Subtle Medical Inc, Menlo Park, CA, United States, ²RadNet, Encino, CA, United States, ³CorTechs Labs Inc, San Diego, CA, United States, ⁴Stanford University, Stanford, CA, United States

3D T1-weighted MRI is valuable for providing high resolution structural information and is commonly used in brain MRI exams despite long scan times. The recent development of deep learning (DL) has shown great potential for scan time reduction. However, the generalizability of DL methods is of concern for actual clinical deployment. In this study, we apply FDA-cleared DL software to accelerate 3D T1-weighted MRI scans by two folds and evaluate the quantification accuracy using FDA-cleared image analysis software. This study provides insight into the generalizability and accuracy of DL in clinical settings.

Valentin H. Prevost¹, Bei Zhang², Clemence Bal³, and Wolter de Graaf^2
1Canon Medical Systems Corporation, Tochigi, Japan, ²Canon Medical Systems Europe, Zoetermeer, Netherlands, ³Bordeaux University, Bordeaux, France

In MRI, signal noise ratio is the key point, determining the image quality and its medical relevance. Different ways exist to significantly increase it and then to access to high resolution imaging. Preliminary works introduced deep learning based denoising on several contexts and conclude to a significant signal noise ratio on qualitative images. However, it has not been tested yet on quantitative imaging sequences, questioning its feasibility and potential in this context. In this study, we investigated the DLR impact on calculated T₁ and T₂ relaxation times and diffusion imaging in healthy human brain areas.

Frederik Luca Sandig^1,2, Julian Emmerich^3,4, Edris El-Sanosy^1,2, Mark Ladd^1,2,3, Heinz-Peter Schlemmer¹, and Sina Straub^3
1Division of Radiology, German Cancer Research Center (DKFZ), Heidelberg, Germany, ²Faculty of Medicine, Heidelberg University, Heidelberg, Germany, ³Division of Medical Physics in Radiology, German Cancer Research Center (DKFZ), Heidelberg, Germany, ⁴Faculty of Physics and Astronomy, Heidelberg University, Heidelberg, Germany

Multiple sclerosis is a chronic inflammatory disease characterized by demyelination. Magnetic resonance imaging (MRI) is an important method for diagnosis and prognosis predictions. The ongoing study presented here shows the use of deep learning algorithms for white and grey matter lesion segmentation in 7T MRI images. Results show high accuracy for patients with high lesion load. Furthermore, it is demonstrated that it is possible to train a neural net to find small cortical lesions, which can be used as a potential biomarker.

Kaikai Shen^1,2,3, Lee Reid¹, Samantha Burnham¹, and Jurgen Fripp^1
1Australian eHealth Research Centre, CSIRO, Herston, Australia, ²Department of Biomedical Sciences, Macquarie University, Sydney, Australia, ³Rapiscan Systems, Sydney, Australia

The distribution of fiber population in the whole brain can be inferred from the samples generated by tractography on diffusion MRI. In this paper, we modeled the distribution of fiber population globally based on representation learning method. Using deep neural networks, we performed dimension reduction on the fiber tracts, and modeled the fiber population by a probability distribution over a latent space in lower dimension. This method enabled us to identify tracts distributed with different densities when compared with another tractogram, and can thus be used to identify structural difference or to detect spurious tracts caused by probabilistic tractography.

Hui Zhang¹, ZiYing Feng², Fei Dai¹, WeiBo Chen³, YiShi Wang⁴, ChengYan Wang², and He Wang^1,2
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China, ²Human Phenome Institute, Fudan University, Shanghai, China, ³Philips Healthcare, Shanghai, China, ⁴Philips Heathcare, Beijing, China

This work tried to optimize MUSE for high resolution multi-shot EPI DWI reconstruction by using Convolutional neural network (CNN). By using multi-scale U-net learning neural network, final reconstructed images showed less artifacts due to the improved phase estimation. Besides, CNN can improve the computational efficiency for the image reconstruction process.

Robert Wujek^1,2, Melissa Prah³, Mona Al-Gizawiy³, and Kathleen Schmainda^3
1Graduate School, Medical College of Wisconsin, Milwaukee, WI, United States, ²Graduate School, Marquette University, Milwaukee, WI, United States, ³Biophysics, Medical College of Wisconsin, Milwaukee, WI, United States

There is a clinical need for a biomarker that can distinguish between infiltrative tumor and non-tumor within FLAIR non-enhancing brain lesions. In this study, we used multi-parametric MRI and FLAIR non-enhancing biopsy ground truths to train a 3D convolutional neural network to address this need.

Ai Nakao¹, Daiki Tamada², Tomohiro Takamura², Shintaro Ichikawa², Utaroh Motosugi², and Yasuhiko Terada^1
1Institute of Applied Physics, University of Tsukuba, Tsukuba, Japan, ²Department of Radiology, University of Yamanashi, Chuo, Japan

Shortening scan time has been a long-standing goal in MRI, and image synthesis and superresolution using deep learning (DL) are promising tools for achieving this goal. However, most of studies use datasets with healthy volunteers for network training, and the clinical evaluation has not yet been fully performed. Here we trained networks using a large, clinical dataset of patients with brain lesions, and evaluated the generated images in terms of the diagnostic image quality and performance. Our results showed that FLAIR superresolution outperformed FLAIR image synthesis. Our results could also provide useful guidelines for evaluating diagnostic performance of DL-based networks.

Danhui Fu¹, Guanqiao Jin¹, Lidong Liu¹, Wenjuan Deng¹, Wei Zhang¹, Qianlin Ding¹, Weiyin Vivian Liu², Long Qian², and Danke Su^1
1GuangXi Medical University Cancer Hospital, Nanning, China, ²MR Research, GE Healthcare, Beijin, China

Conventional neuropsychometric tests could not early detect HIV-associated neurocognitive disorder (HAND) on HIV-patients. More objective indicators with neuroimaging techniques have offered to sub-classify patients with specific symptoms. Significant differences of regional gray matter volumes (rGMV) betweenHIV-positive group and age-matched healthy group was found in our preliminary study. Ten major rGMV well distinguished HIV-positive patients with HAND from those without HAND. Our study showed rGMV might be an indicator to explain for neurocognitive impairments.

Kyle Murray¹, Igor B Titoff², Henry Wang³, Jianhui Zhong^1,3, and Giovanni Schifitto^2,3
1Physics and Astronomy, University of Rochester, Webster, NY, United States, ²Neurology, University of Rochester, Rochester, NY, United States, ³Imaging Sciences, University of Rochester, Rochester, NY, United States

HIV-infection is known to be related to vascular diseases, which can be explored via cerebral imaging techniques such as magnetic resonance angiography (MRA) and arterial spin labeling (ASL). In this abstract, we use quantitative features extracted from demographic information and vascular imaging data, only, to predict HIV-status in adults using a support vector machine (SVM). This is the first SVM to reasonably predict HIV-status in an aging HIV-population on combination antiretroviral therapy, which may have future biological implications in HIV research.

Ceren Tozlu¹, Keith Jamison¹, Susan Gauthier^1,2,3, and Amy Kuceyeski^1,4
1Department of Radiology, Weill Cornell Medicine, New York City, NY, United States, ²Judith Jaffe Multiple Sclerosis Center, Weill Cornell Medicine, New York City, NY, United States, ³Department of Neurology, Weill Cornell Medicine, New York City, NY, United States, ⁴Brain and Mind Research Institute, Weill Cornell Medicine, Ithaca, NY, United States

No study to date has performed a rigorous analysis of the relative contributions of multi-modal imaging data including the brain’s functional (FC) and structural connectivity (SC) in the task of classifying high and low adapting MS patients for a deeper understanding of the connectome-level mechanism contributing to variability in MS-related impairment. We built a machine learning based ensemble model that can accurately classify MS patients as high and low adapters (AUC> 0.626). We observed that SC and FC networks can be used to identify the most discriminative regions and to accurately classify MS patients regarding their impairment level.

Saba Momeni^1,2, Amir Fazlollahi¹, Pierrick Bourgeat¹, Paul Yate³, Nawaf Yassi⁴, Patricia Desmond⁵, Yongsheng Gao², Alan Wee Chung Liew⁶, and Olivier Salvado^7
1CSIRO Health and Biosecurity, Brisbane, Australia, ²Griffith University, Brisbane, Australia, ³Department of Nuclear Medicine and Centre for PET, Brisbane, Australia, ⁴University of Melbourne, Parkville, Australia, Melbourne, Australia, ⁵Austin Health Heidelberg, Australia, Melbourne, Australia, ⁶Griffith University, Gold coast, Australia, ⁷Data61, Brisbane, Australia

The lack of clinical dataset with enough examples of rare lesions challenges supervised machine learning methods. Here we propose to generate synthetic lesions for training a classifier to identify microbleeds from MRI QSM. We show that the performance of the classifier is improved compare to standard data augmentation using actual data alone. Our synthetic dataset can have unlimited size allowing to perform validation experiment, while keeping the actual data for testing. Moreover, many aspects of the data can be investigated, which would not be possible when using actual lesions: lesions can be synthetize on any location, size, shape, and intensity.  

Jun-Cheng Weng^1,2,3, Tung-Yeh Lin¹, Man Teng Cheok¹, Yuan-Hsiung Tsai⁴, Yi-Peng Eve Chang⁵, and Vincent Chin-Hung Chen^3,6
1Department of Medical Imaging and Radiological Sciences, Chang Gung University, Taoyuan, Taiwan, ²Medical Imaging Research Center, Institute for Radiological Research, Chang Gung University and Chang Gung Memorial Hospital at Linkou, Taoyuan, Taiwan, ³Department of Psychiatry, Chang Gung Memorial Hospital, Chiayi, Taiwan, ⁴Department of Diagnostic Radiology, Chang Gung Memorial Hospital, Chiayi, Taiwan, ⁵Department of Counseling and Clinical Psychology, Columbia University, New York City, NY, United States, ⁶School of Medicine, Chang Gung University, Taoyuan, Taiwan

It is estimated that at least one million people die by suicide every year, showing the importance of suicide prevention and detection. An autoencoder and machine learning model was employed to predict people with suicidal ideation based on their brain structural imaging. Our results showed that the best pattern of structure across multiple brain locations can classify suicidal ideates from NS and HC with a prediction accuracy of 85%, a specificity of 100% and a sensitivity of 75%. The algorithms developed here might provide an objective tool to help identify suicidal ideation risk among depressed patients alongside clinical assessment.

Ho-Joon Lee¹, Yeonah Kang¹, Marc Lebel², Min Soo Park³, and Joonsung Lee^3
1Department of Radiology, Haeundae Paik Hospital, Busan, Republic of Korea, ²MR Collaboration and Development, GE Healthcare, Calagary, AB, Canada, ³GE Healthcare Korea, Seoul, Republic of Korea

With the increased scientific interest in hippocampal subfields, a clinically applicable T2 mapping method is demanding. With a dual echo fast spin echo based approach, we were able to acquire high resolution T2 maps <5 minutes scan time. With recently introduced deep learned reconstruction, improved image quality of T2 weighted image, and T2 map is observed. Differences in measurements regarding T2 values and volumes of hippocampal subfields will be presented.

Matthew Joseph Kuhn¹, Julia Patriarche², Douglas Patriarche², Annabel Kuhn³, and Andrew Scott Kuhn^4
1Radiology, University of Illinois College of Medicine, Peoria, IL, United States, ²A.I. Analysis, Inc., Seattle, WA, United States, ³Medical College of Wisconsin, Milwuakee, WI, United States, ⁴Radiology, Yale University, New Haven, CT, United States

The use of half dose gadolinium coupled with artificial intelligence algorithms  is effective in visualizing primary and metastatic disease to the brain and reduces gadolinium retention concerns. There is potential cost savings from the use of lower dose gadolinium.

Huijin Song¹, Eunji Kim², Hyunsil Cha², Moon Jung Hwang³, Yongmin Chang^2,4, and Chul-Ho Sohn^5
1Biomedical Research Institute, Seoul National University Hospital, Seoul, Korea, Republic of, ²Medical & Biological Engineering, Kyungpook National University, Daegu, Korea, Republic of, ³GE Health Korea, Seoul, Korea, Republic of, ⁴Radiology, College of Medicine, Kyungpook National University, Daegu, Korea, Republic of, ⁵Radiology, Seoul National University Hospital, Seoul, Korea, Republic of

Deep learning-based brain tumor segmentation requires multi-modality dataset for a precision accuracy. However, multi-modality data acquisition has a limitation due to several reasons. In this study, we propose that a 3D Mask R-CNN network model could provide a reliable accuracy in individual modality dataset for brain tumor segmentation.

Abhilasha Indoria¹, Madhura Ingalhalikar², and Jitender Saini^1
1Neuroimaging and Interventional Radiology, NIMHANS, Bengaluru, India, ²Symbiosis Centre for Medical Image Analysis (SCMIA), Symbiosis International University, Pune, India

High grade extraventricular supratentorial ependymoma’s in adults are uncommon neoplasms with imaging features that can mimic cortical tumors if small and high-grade gliomas (HGG) if large. No previous work has tried to discriminate ependymoma from high grade gliomas using MRI. Our work evaluates preoperative diffusion weighted imaging for discrimination of ependymomas from grade III and grade IV gliomas using textural analysis. Results demonstrate significant differences in the histogram and first order textural features derived from diffusion weighted imaging in cases of ependymomas and high-grade gliomas.

Anup Singh^1,2, Neha Vats¹, Virendra Kumar Yadav¹, Anirban Sengupta³, Rakesh Kumar Gupta⁴, Sumeet Agarwal⁵, Mamta Gupta⁶, Rana Patir⁷, Sunita Ahlawat⁶, and Jitender Saini^8
1Centre for Biomedical Engineering, Indian Institute of Technology Delhi, New Delhi, India, ²Biomedical Engineering, AIIMS, New Delhi, India, ³Vanderbilt University Medical Center, Nashville, TN, United States, ⁴Department of Radiology, Fortis Memorial Research Institute, Gurugram, India, ⁵Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, India, ⁶Fortis Memorial Research Institute, Gurugram, India, ⁷Neurosurgery, Fortis Memorial Research Institute, Gurugram, India, ⁸Department of Neuroimaging and Interventional Radiology (NIIR), NIMHANS Bangalore, Bangalore, India

Imaging based diagnosis of Pilocytic Astrocytoma (PA) is quite important for better prognosis. PA can easily be misdiagnosed since its location, growth pattern, and contrast enhancement often mimic a more aggressive high-grade glioma(HGG) tumor. In the current study, quantitative analysis of T1-Perfusion(DCE) MRI data was performed followed by extraction of various features from tumor region and development of an optimized support-vector-machine(SVM) classifier for automatic differentiation of PA vs HGG. The proposed machine learning based approach which uses features derived from quantitative T1 perfusion MRI and tumor volume fraction can enable accurate diagnosis of PA and HGG tumors.

Agata Sularz¹, Fulvio Zaccagna¹, Dimitri A Kessler¹, Fraser Tonnard¹, Sonia Benitez¹, Thomas Santarius², Fiona J Gilbert¹, Tomasz Matys¹, and Joshua D Kaggie^1
1Department of Radiology, University of Cambridge, Cambridge, United Kingdom, ²Department of Neurosurgery, University of Cambridge, Cambridge, United Kingdom

We describe a deep learning method for fully-automated brain meningioma MRI segmentation. A conditional generative adversarial network (cGAN) was trained on T1 contrast-enhanced (T1ce) MRI of 37 patients. We explored the effect of batch size, transfer learning and histogram equalization on segmentation accuracy. The highest results for T1ce images were achieved for meningioma dataset of batch size = 1 (DSC = 0.347). Histogram equalization improved segmentation accuracy for batch size = 1 (DSC = 0.364) and batch size = 200. Transfer learning on a publicly available glioma dataset did not improve segmentation results.

David D Shin¹, David Yen-Ting Chen², Audrey P Fan², Moss Y Zhao², Jia Guo³, Suchandrima Banerjee¹, and Greg Zaharchuk^2
1GE Healthcare, Menlo Park, CA, United States, ²Stanford University, Stanford, CA, United States, ³University of California, Riverside, CA, United States

Deep learning ASL Pipeline is presented that automatically retrieves relevant MR images from multiple clinical scanners, performs preprocessing/inferencing and returns the derived synthesized PET-like CBF maps back to the scanners for rapid clinical evaluation and prototyping. 

Jeremiah W Sanders¹, Jason M Johnson², Jong Bum Son¹, Zijian Zhou¹, Henry Szu-Meng Chen¹, Joshua Yung¹, Jason Szu-Meng Stafford¹, Melissa Chen², Maria Gule-Monroe², Ho-Ling Liu¹, Mark D Pagel³, and Jingfei Ma^1
1Imaging Physics, University of Texas MD Anderson Cancer Center, Houston, TX, United States, ²Neuroradiology, University of Texas MD Anderson Cancer Center, Houston, TX, United States, ³Cancer Systems Imaging, University of Texas MD Anderson Cancer Center, Houston, TX, United States

Neurocognitive function is often associated with structural differences in the brain for patients with neurofibromatosis type-1 (NF1), and studies have shown that NF1 is associated with larger subcortical volumes and thicker cortices of certain brain structures. Routine monitoring of NF1 patients would be possible with tools that enable rapid whole-brain segmentation in standard of care T1w MRI. Modern machine learning techniques, including fully convolutional networks (FCNs), have demonstrated the ability to rapidly perform segmentation tasks across a range of applications. In this work, we investigate the performance of different FCNs for rapid whole-brain segmentation in pediatric T1w brain MRI.

Patrick Quarterman¹, Gul Moonis², and Marc Lebel^3
1GE Healthcare, New York, NY, United States, ²Radiology, Columbia University, New York, NY, United States, ³GE Healthcare, Calgary, AB, Canada

The purpose of this study was to determine the effectiveness of using deep learning reconstruction (DL Recon) for improved signal-to-noise (SNR) allowing for higher resolution imaging of the lateral rectus-superior rectus band of the orbit without substantial increase in scan/exam time. Evaluation of this new reconstruction technique was performed on a group of 5 volunteers with results indicating that higher resolution protocols can produce images with increased SNR and removal of noise leading to higher confidence in identifying the lateral rectus-superior rectus band.

Danhui Fu¹, Guanqiao Jin¹, Lidong Liu¹, Wenjuan Deng¹, Sen Hong¹, Qianlin Ding¹, Long Qian², Weiyin Vivian Liu², and Danke Su^1
1GuangXi Medical University Cancer Hospital, Nanning, China, ²MR Research, GE Healthcare, Beijin, China

Early detection of cognitive impairments in HIV carriers is essential. Neuroimaging studies has shown brain structure and function of HIV-positive patients are associated with cognition malfunction. However, the conclusions reported from previous studies are mainly based on the univariate analysis, such as t-test. In recent days, increasing attention have focused on the multivariate pattern analysis. Hence, in current study, we demonstrated the aberrant degree centrality (DC) using a machine learning approach. Our results suggested that DC might be an indicator to early detect neurocognitive impairments.

xiao hu¹, chaoyong xiao², xiangrong zhang², sidong liu³, zaixu cui⁴, weiguo liu⁵, and long qian^6
1radiology, nanjing brain hospital, nanjing, China, ²nanjing brain hospital, nanjing, China, ³Clinical Medicine, Macquarie University, Sydney, Australia, ⁴Psychiatry, University of Pennsylvania, Philadelphia, PA, United States, ⁵neurology, nanjing brain hospital, nanjing, China, ⁶Biomedical Engineering, Peking University, beijing, China

Parkinson's disease (PD) is a common neurodegenerative disorder characterized by disabling motor and non-motor symptoms.1 The abnormalities of white-matter (WM) tracts/regions have been demonstrated in PD. However, previous studies have largely dependent on univariate analysis, such as t-test, which may result in Type-1 error. Further, it remains unclear whether the disruption of WM tracts/regions provided worthwhile information to identify PD from HC. Hence, in current study, a machine learning approach was applied to investigate the white matter profiles of PD.

Yan Fu¹, Tianjing Zhang², Haixia Li³, Yichen Tong³, Xiangchuang Kong⁴, and Dingxi Liu^4
1EPFL, Lausanne, Switzerland, ²Philips Healthcare, Guangzhou, China, ³Sun Yat-Sen University, Guangzhou, China, ⁴Radiology, Union Hospital,Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China

In routine MRI clinical application of compressed sensing, the image quality is often not useful for diagnosis when its CS(Compressed SENSE) acceleration factor is beyond a certain level(e.g. 8 or 10). It is desirable to further accelerate MR sequences in multiple applications such as brachial plexus nerve, coronary artery, and so on. It is possible to use generative adversarial network(GAN) models to further optimize the imaging workflow by improving the image quality of data acquired with high CS factors. 

Richard Michael Gagen¹, Carmen Dragos², and Charles Hutchinson^1
1Populations, Evidence and Technology, Warwick Medical School, University of Warwick, Coventry, United Kingdom, ²Imaging, University Hospital Covenry and Warwickshire, Coventry, United Kingdom

A case-referant type study assessing the performance of MRI lumbar spine based measurements of the spinal canal in classifying participants as either having neurogenic claudication or non-specific low-back pain. Three measurements of the central lumbar canal are assessed: the dural-sac cross-sectional area, anteroposterior canal diameter and A-D grading system proposed by Schizas et al. The performance of these measurements is compared to a machine learning model incorporating a range of other MRI measurements and demographic information.

Bo Pang¹, Bhaswati Roy², Milena Lai², Luke Ehlert², Ravi S. Aysola ³, Daniel W. Kang³, Ariana Anderson^1,4, and Rajesh Kumar^2,5,6,7
1Statistics, University of California at Los Angeles, Los Angeles, CA, United States, ²Anesthesiology, University of California at Los Angeles, Los Angeles, CA, United States, ³Medicine, University of California at Los Angeles, Los Angeles, CA, United States, ⁴Psychiatry and Biobehavioral Sciences, University of California at Los Angeles, Los Angeles, CA, United States, ⁵Radiology, University of California at Los Angeles, Los Angeles, CA, United States, ⁶Bioengineering, University of California at Los Angeles, Los Angeles, CA, United States, ⁷Brain Research Institute, University of California at Los Angeles, Los Angeles, CA, United States

Deep learning has demonstrated impressive performance in a wide range of complex and high-dimensional imaging data, including medical image classification and segmentation. One major challenge of harnessing the power of neural networks in image analysis is the small sample size. The present work utilizes deep learning models to classify high-resolution T1-weighted images of obstructive sleep apnea patients (OSA) from healthy controls. Using 193 participants and with adopted model regularization and exponential moving averaging of model weights, we showed 65% testing accuracy and 80% sensitivity. The findings demonstrate the potential for applying neural network models in assisting image-based OSA diagnoses.

Yoon-Chul Kim¹, In-Young Baek², Ji-Eun Lee², Ha Na Song², and Woo-Keun Seo^2
1Clinical Research Institute, Samsung Medical Center, Sungkyunkwan Univ. Sch. of Med., Seoul, Republic of Korea, ²Department of Neurology, Samsung Medical Center, Sungkyunkwan Univ. School of Medicine, Seoul, Republic of Korea

This study demonstrates an automatic method that predicts favorable/unfavorable clinical outcome based on pre-treatment DWI data and machine learning (ML) in acute ischemic stroke. We present the use of ADC histogram information in the brain tissue as features for the ML prediction. In the histogram analysis, the 5 or 10 percentile value of the ADC distribution was indicative of clinical outcome regardless of success/failure of recanalization. The ROC analysis in unseen test subjects resulted in an area under the curve (AUC) of 0.79 with the proposed feature extraction, which was greater than 0.71 with the DWI lesion volume only. 

Joohyun Lee¹, Jongho Lee¹, and Haejin Kim^2
1Department of Electrical and Computer Engineering, Seoul National University, Seoul, Republic of Korea, ²Department of Basic Science and Technology, Hongik University, Sejong, Republic of Korea

Although segmentation using deep learning performs well, it often works poorly on small lesions or boundaries of lesion. This can occur the serious issues when applying in the medical images and the more reliable method is essential. In this study, we developed a deep learning process based on the uncertainty measurements that improves brain tumor segmentation. For selectively maximizing either precision or recall, two types of segmentation methods were presented.

Chia-Feng Lu¹, Cheng-Chia Lee^2,3,4, Hsiu-Mei Wu^3,5, Huai-Che Yang^2,3, Man-Chin Chen¹, Chung-Jung Lin^3,5, Wan-Yuo Guo^3,5, and Wen-Yuh Chung^2,3
1Department of Biomedical Imaging and Radiological Sciences, National Yang-Ming University, Taipei, Taiwan, ²Department of Neurosurgery, Neurological Institute, Taipei Veteran General Hospital, Taipei, Taiwan, ³School of Medicine, National Yang-Ming University, Taipei, Taiwan, ⁴Brain Research Center, National Yang-Ming University, Taipei, Taiwan, ⁵Department of Radiology, Taipei Veteran General Hospital, Taipei, Taiwan

Even though the gamma knife radiosurgery (GKRS) shows promising evidence in treating cavernous malformation (CM), there are still a part of patients will have a recurrent hemorrhage after radiosurgery. We proposed a prediction model based on preradiosurgical MR radiomics to estimate the personalized hemorrhage free survival after GKRS. The satisfactory results of the proposed model can benefit the healthcare in patients with CM by providing a reliable prognosis before treatment.

Yuya Saito^1,2, Akihiko Wada², Shimpei Kato², Koji Kamagata², Masaaki Hori³, and Shigeki Aoki^2
1Department of Radiological Sciences, Graduate School of Human Health Sciences, Tokyo Metropolitan University, Tokyo, Japan, ²Department of Radiology, Graduate School of Medicine, Juntendo University, Tokyo, Japan, ³Department of Radiology, Toho University Omori Medical Center, Tokyo, Japan

In the treatment of acute infarction, the detection of abnormal high signals in diffusion weighted images contributes to early diagnosis and treatment of infarction. In this study, we developed a deep learning neural network model via autoencoder (AE) to diagnosis brain infarction and predict vascular territory automatically from a DWI image. 1582 brain images including normal and abnormal brain which had infarctions were used as a training and test dataset. As a result, our model detected brain infarction and estimated vascular territory with high accuracy. It can be an effective indicator for diagnosing correctly infarction and predicting treatment effect.

Ching-Chung Ko^1,2, Yang Zhang³, Kai-Ting Chang³, Jeon-Hor Chen^3,4, and Min-Ying Lydia Su^3
1Department of Medical Imaging, Chi Mei Medical Center, Tainan, Taiwan, ²Department of Pharmacy, Chia Nan University of Pharmacy and Science, Tainan, Taiwan, ³Department of Radiological Sciences, University of California, Irvine, CA, United States, ⁴E-Da Hospital/I-Shou University, Kaohsiung, Taiwan

A subset of benign meningiomas may show early progression/recurrence (P/R) after surgery. In clinical practice, one of the main challenges in the treatment of meningiomas is to determine factors that correlate with P/R. This study investigated the role of radiomics and machine learning for the prediction of P/R in meningiomas. 128 patients diagnosed with WHO grade I meningioma were studied. Total 214 descriptors were extracted from the various MR sequences. The prediction accuracy of P/R was 74% and the AUC of the prediction model was 0.80.

Ching-Chung Ko^1,2, Kai-Ting Chang³, Yang Zhang³, Jeon-Hor Chen^3,4, and Min-Ying Lydia Su^3
1Department of Medical Imaging, Chi Mei Medical Center, Tainan, Taiwan, ²Department of Pharmacy, Chia Nan University of Pharmacy and Science, Tainan, Taiwan, ³Department of Radiological Sciences, University of California, Irvine, CA, United States, ⁴E-Da Hospital/I-Shou University, Kaohsiung, Taiwan

A subset of nonfunctioning pituitary macroadenomas (NFMAs) show early progression/recurrence (P/R) after surgery. In clinical practice, one of the main challenges in the treatment of NFMAs is to determine factors that correlate with P/R. This study investigated the role of radiomics for the prediction of P/R in NFMAs. 50 patients diagnosed with benign NFMAs were studied. Totally 214 descriptors were extracted from the various MR sequences. The prediction accuracy of P/R was 82% and the AUC of the prediction model was 0.78.

Jinghua Wang¹, Ming Chen^2,3, Lili He^2,4, Hailong Li², Vivek Khandwala¹, David Wang¹, Brady Williamson¹, Daniel Woo⁵, and Achala Vagal^1
1Radiology, University of Cincinnati, Cincinnati, OH, United States, ²The Perinatal Institute and Section of Neonatology, Perinatal and Pulmonary Biology, Cincinnati Children’s Hospital Medical Center, Cincinnati, OH, United States, ³Electrical Engineering and Computer Science, University of Cincinnati, Cincinnati, OH, United States, ⁴Pediatrics, University of Cincinnati, Cincinnati, OH, United States, ⁵Neurology and Rehabilitation Medicine, University of Cincinnati, Cincinnati, OH, United States

Timely and reliable prognostic tools for intracerebral hemorrhage (ICH) have great potential to guide physician decision making. They are potentially useful for targeting patients for interventions and optimizing rehabilitation strategies. The objective of this study is to investigate if a deep transfer learning model can capture individual variability to predict clinical outcome for ICH patients at 3 months using the integration of clinical and T₂-weighted fluid-attenuated inversion recovery (FLAIR) imaging data. Our model was able to correctly identify patients likely to have unfavorable outcomes with an AUC of 0.87 (95% confidence interval: 0.86, 0.89).  

Jinghua Wang¹, Ming Chen^2,3, Lili He^2,4, Hailong Li², Vivek Khandwala¹, David Wang¹, Brady Williamson¹, Daniel Woo⁵, and Achala Vagal^1
1Radiology, University of Cincinnati, Cincinnati, OH, United States, ²The Perinatal Institute and Section of Neonatology, Perinatal and Pulmonary Biology, Cincinnati Children’s Hospital Medical Center, Cincinnati, OH, United States, ³Electrical Engineering and Computer Science, Cincinnati Children’s Hospital Medical Center, Cincinnati, OH, United States, ⁴Pediatrics, University of Cincinnati, Cincinnati, OH, United States, ⁵Neurology and Rehabilitation Medicine, University of Cincinnati, Cincinnati, OH, United States

Intracerebral hemorrhage (ICH) accounts for 10% - 30% of all strokes and is associated with high short-term mortality (≤50% at 3 month).  There is a critical unmet need for an effective prognostic tool using imaging markers to identify patients at risk for poor outcome and thereby better facilitating treatments at individual level as well as tailoring personalized interventions and optimizing rehabilitation strategies. In this work, we developed a machine learning method using radiomics features derived from T₂-weighted FLAIR images to predict recovery outcome in patients with ICH at 3 months with a accuracy of 80.8% (95% confidence interval: 78.9%, 82.8%).

Sumeet Shinde¹, Tanay Chougule¹, Jitender Saini², and Madhura Ingalhalikar^1
1Symbiosis Center for Medical Image Analysis, Symbiosis International University, Pune, India, ²Department of Radiology, National Institute of Mental Health and Neurosciences, Bengaluru, India

High grade supratentorial ependymomas in adults are rare neoplasms with imaging features that can often be confused for high grade gliomas. However, the pathogenesis of ependymoma’s differs significantly and it therefore crucial to determine pre-operative non-invasive markers for treatment planning and optimization. Our work creates a multi-variate 3-way classification framework based on multimodal MRI radiomics to discriminate supra-tentorial ependymomas, grade III and IV gliomas. Results demonstrate high accuracy with specific textures that evolve as the top discriminative features and can be pursued for clinical applicability.

Madhura Ingalhalikar¹, Tanay Chougule¹, Sumeet Shinde¹, Vani Santosh², and Jitender Saini^3
1Symbiosis Center for Medical Image Analysis, Symbiosis International University, Pune, India, ²Department of Neuropathology, National Institute of Mental Health and Neurosciences, Bangalore, India, ³Department of Radiology, National Institute of Mental Health and Neurosciences, Bangalore, India

 Radiomics based multi-variate models and state-of-art convolutional neural networks (CNNs) have demonstrated their usefulness for predicting IDH genotype in gliomas from MRI images. However, adaptability and clinical explanability of these models on unseen multi-center datasets has not been investigated. Our work trains radiomics and CNN based classifiers on a large dataset (TCIA) and tests multiple local datasets. Results demonstrate higher adaptability of radiomics than standard CNNs, except for transfer learned CNNs. Better interpretability was obtained from feature ranking (in case of radiomics) and high resolution class activation maps (in case of CNNs).

Xiaoxue Liu¹, Baoming Li², Jianrui Li¹, Li Yu², Jun Xu², Guangming Lu¹, and Zhiqiang Zhang^1
1Department of Medical Imaging, Jinling Hospital, Nanjing University School of Medicine, Nanjing, China, ²Jiangsu Key Laboratory of Big Data Analysis Technique and CICAEET, Nanjing University of Information Science and Technology, Nanjing, China

We tried to establish a model based on features extracted from Pyradiomic framework and Co-occurrence of Local Anisotropic Gradient Orientations (CoLlAGe) in MRI to predict the prognosis of gliomas. Fifty-five pathologically confirmed glioma patients were retrospectively collected. All patients underwent 3DTI enhancement, standardized treatment and follow-up for overall survival. 1781 Pyradiomic features and 260 CoLlAGe features were collected. Three Cox proportional hazards models were fitted with Pyradiomics features, CoLlAGe features and Pyradiomics+CoLlAGe features. The C-index in these models were 0.835, 0.820 and 0.844, respectively. The model with Pyradiomics+CoLlAGe features demonstrated improved survival predictive performance than the single model. 

Suhail Pathan Parvaze¹, Mamta Gupta², Anup Singh³, Rana Patir⁴, Sunita Ahlawat², Madhura Ingalhalikar⁵, Neha Vats³, and Rakesh Kumar Gupta^2
1Philips Innovation Campus, Bangalore, India, ²Fortis Memorial Research Institute, Gurgaon, India, ³Indian Institute of Technology, New Delhi, India, ⁴Department of Neurosurgery, Fortis Memorial Research Institute, Gurgaon, India, ⁵Symbiosis Center for Medical Image analysis, Symbiosis International University, Pune, India

This work aimed at identifying radiomic signatures that discriminate glioblastoma sub-regions ( enhancing tumor, non-enhancing, necrosis and edema) using DCE T1 perfusion based post contrast MRI. Results demonstrated texture features that delineate the four regions as well as separate the tumoral region from normative white matter. These radiomic signatures can be further investigated to gain deeper understanding of tumor progression and recurrence.

Zhe Liu¹, Wenxin Xue¹, Xiaotong Liu², Ting Liang¹, Chao Jin¹, Xiaocheng Wei³, Buyue Qian², and Jian Yang^1
1The first affiliated Hospital of Xi’an Jiaotong University, Xi’an, China, ²Xi’an Jiaotong University, Xi’an, China, ³MR Research China, GE Healthcare,, Xi'an, China

Glioblastoma and brain solitary metastasis from lung cancer have similar peritumoral edema on T2-weighted imaging (T2WI). However, indistinguishable signs between these two tumors embarrass the radiologists and lead to high misdiagnosis rate. To address such issue, radiomics biomarkers were analyzed to detail the tumors’ histologic and morphologic characteristics. Results indicated that radiomics biomarkers including histogram of oriented gradient, shape and grey level co-occurrence matrix, which charaterize the lesion’s shape and signal showed good performance in differentiating these two tumors. Furthermore, using those radiomics biomarkers, a gradient-boosting machine learning model was established and showed good performance (Area under the curve=0.88).