Yasuhiko Tachibana¹, Masataka Nishimori², Masaaki Chiku³, Naoyuki Kitamura², Kensuke Umehara⁴, Junko Ota⁴, Takayuki Obata¹, and Tatsuya Higashi^5
1Applied MRI Research, Department of Molecular Imaging and Theranostics, National Institute of Radiological Sciences, QST, Chiba, Japan, ²MNES corporation, Tokyo, Japan, ³Medical Check Studio Ginza Clinic, Tokyo, Japan, ⁴Medical Informatics Section, QST Hospital, QST, Chiba, Japan, ⁵Department of Molecular Imaging and Theranostics, National Institute of Radiological Sciences, QST, Chiba, Japan

The neural network model was designed to judge the existence of aneurysms from brain MR angiography images. On the hypothesis that each radiologist (annotator) has a unique bias for decision, the network was designed so that it accepts input of who the annotator was as an additional information to compute the output. The hypothesis might be reasonable, and the model design might be useful because the accuracy of the trained model (area under the curve (AUC) in receiver operating characteristic (ROC) analysis) elevated significantly (P<.0001, DeLong test) by adding the information of who the annotator was.

Fernanda Lenita Ribeiro^1,2, Steffen Bollmann³, and Alexander M Puckett^1,2
1School of Psychology, University of Queensland, Brisbane, Australia, ²Queensland Brain Institute, University of Queensland, Brisbane, Australia, ³Centre for Advanced Imaging, University of Queensland, Brisbane, Australia

Whether it be in a man-made machine or a biological system, form and function are often directly related. In the latter, however, this particular relationship is often unclear due to the intricate and involved nature of biology. Here we developed a geometric deep learning model capable of exploiting the actual structure of the cortex to learn the complex relationship between brain function and anatomy from structural and functional MRI data. Our model was not only able to predict the functional organization of human visual cortex from anatomical properties alone, but it was also able to predict nuanced variations across individuals.

Srivathsa Pasumarthi Venkata¹, Jonathan Tamir¹, Enhao Gong¹, Greg Zaharchuk¹, and Tao Zhang^1
1Subtle Medical Inc., Menlo Park, CA, United States

Gadolinium-based contrast agents (GBCAs) create unique image contrast to facilitate identification of various clinical findings. However, recent discovery of gadolinium deposition after contrast-enhanced MRI raises new safety concerns of GBCAs. Deep learning (DL) has recently been used to predict the contrast-enhanced images using only a fraction of the standard dose. However, challenges remain in generalizing the DL methods across different protocols/vendors/institutions. In this work, we propose comprehensive technical solutions to improve DL model robustness and obtain high quality low-dose contrast-enhanced MRI across multiple scanners and institutions.

Jonghyun Bae^1,2, Li Feng³, Krzysztof Geras⁴, Florian Knoll⁴, Yulin Ge^4,5, and Sungheon Gene Kim^4,5
1Sackler Institute of Graduate Biomedical Science,NYU School of Medicine, New York, NY, United States, ²Radiology, Center for Advanced Imaging Innovation and Research, New York, NY, United States, ³Icahn School of Medicine at Mount Sinai, New York, NY, United States, ⁴Center for Advanced Imaging Innovation and Research, New York, NY, United States, ⁵Center for Biomedical Imaging, NYU, New York, NY, United States

This study proposes a deep learning approach of estimating the capillary level of input function for kinetic model analysis on dynamic contrast enhanced (DCE)-MRI data. Our deep-learning network was trained with the numerically synthesized data generated with a wide range of contrast kinetic dynamics with different arterial input function (AIF). We hypothesize that the voxel level capillary input functions would be more accurate input functions for pharmacokinetic analysis. This hypothesis was tested with the DCE-MRI data of healthy subjects.

Kai Wang¹, Qinyang Shou¹, Samantha Ma¹, David Liebeskin², Xin Qiao², Jeffrey Saver², Noriko Salamon², Songlin Yu³, Hosung Kim¹, Yannan Yu⁴, Yuan Xie⁴, Greg Zaharchuk⁴, Fabien Scalzo², and Danny Wang^1
1University of Southern California, Los Angeles, CA, United States, ²University of California, Los Angeles, Westwood, CA, United States, ³Beijing Tiantan Hospital, Capital Medical University, Beijing, China, ⁴Stanford University, Stanford, CA, United States

A deep learning (DL)-based algorithm was developed to automatically identify the hypoperfusion lesion and penumbra in ASL images of arterial ischemic stroke (AIS) patients. A total of 167 3D pCASL datasets from 137 AIS patients on Siemens MR were used for training, using concurrently acquired DSC MRI as the label. The DL model achieved a voxel-wise area under the curve (AUC) of 0.958, and 92% accuracy for retrospective determination for subject-level endovascular treatment eligibility. The DL-model was cross validated on 12 GE pCASL data with 92% accuracy without fine-tuning of parameters.

Nicholas J. Luciw^1,2, Zahra Shirzadi², Sandra E. Black², Maged Goubran², and Bradley J. MacIntosh^1,2
1Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada, ²Sunnybrook Research Institute, Toronto, ON, Canada

Arterial spin-labelled (ASL) MRI is used to quantify cerebral blood flow and arterial transit time. Currently, these parameters are not calculated at the scanner given the time-consuming processing required. Fast, automated parameter estimation is therefore desirable to radiology clinics. Here, we trained a convolutional neural network to estimate cerebral blood flow and arterial transit time from multiple post-label delay ASL. The network produces estimates comparable to other approaches and was designed to evaluate model uncertainty. This fast, automated method is suitable for scan-time generation of accurate hemodynamic maps, important in the assessment of neurological disorders and neurodegeneration. 

Soozy Jung¹, Kanghyun Ryu¹, Jae Eun Song¹, Mina Park², and Dong-Hyun Kim^1
1Department of Electrical and Electronic Engineering, Yonsei University, Seoul, Korea, Republic of, ²Department of Radiology, Gangnam Severance Hospital, Seoul, Korea, Republic of

 Recently, magnitude-based artificial neural network (ANN) method was implemented to estimate myelin water fraction (MWF) mapping using multi-echo gradient-echo (mGRE) data. However, MWF mapping in mGRE data requires phase information with the demand of considering frequency shifts in white matter. Here, we developed a complex-valued ANN for MWF mapping which could learn the phase information of the mGRE signal. According to simulation and in vivo analysis, complex-valued ANN is more robust to fiber orientation and noise than magnitude-based ANN and conventional fitting method.

Zifei Liang¹, Choong Heon Lee¹, Tanzil M. Arefin¹, Piotr Walczak², Song-Hai Shi³, Florian Knoll¹, Yulin Ge¹, Leslie Ying⁴, and Jiangyang Zhang^1
1Radiology, NYU Langone Health, New York, NY, United States, ²Diagnostic Radiology & Nuclear Medicine, University of Maryland School of Medicine, Baltimore, MD, United States, ³Center for Molecular Imaging & Nanotechnology, Memorial Sloan Kettering Cancer Center, New York, NY, United States, ⁴Electrical Engineering, University at Buffalo, Buffalo, NY, United States

We developed a deep learning network that can generate new tissue contrasts from MRI data to match the contrasts of several histological methods. The network was trained using the carefully curated histological data from the Allen Institute mouse brain atlas and co-registered MRI data. In our tests, the new contrasts, which resembled Nissl, neurofilament, and myelin-basic-protein stained histology, demonstrated higher sensitivity and specificity than commonly used diffusion MRI markers to characterize neuronal, axonal, and myelin structures in the mouse brain. The contrasts were further validated using two mouse models with abnormal neuronal structures and dysmyelination.  

Tommaso Di Noto¹, Guillaume Marie¹, Sebastien Tourbier¹, Guillaume Saliou¹, Meritxell Bach Cuadra^1,2,3, Patric Hagmann¹, and Jonas Richiardi^1,4
1Faculty of Biology and Medicine, Department of Radiology, Lausanne University Hospital (CHUV) and University of Lausanne (UNIL), Lausanne, Switzerland, ²Medical Image Analysis Laboratory (MIAL), Centre d’Imagerie BioMédicale (CIBM), Lausanne, Switzerland, ³Signal Processing Laboratory (LTS 5), Ecole Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland, ⁴Advanced Clinical Imaging Technology, Siemens Healthcare, Lausanne, Switzerland

One recurrent problem for applying deep learning models in medical imaging is the reduced availability of labelled training data. A common approach is therefore to focus on image patches rather than whole volumes, thus increasing the number of samples. However, for many diseases anomalous patches (positive samples) are outnumbered by negative patches showing no anomaly. Here, we explore different strategies for negative sampling in the context of brain aneurysm detection. We show that classification performances can vary drastically with respect to negative sampling, and that real-world disease or anomaly prevalence can further degrade performance estimates. 

Peter A. Wijeratne^1,2, Sara Garbarino³, Eileanoir B. Johnson², Sarah Gregory², Rachael I. Scahill², Sarah J. Tabrizi², Marco Lorenzi³, and Daniel C. Alexander^1
1Department of Computer Science, University College London, London, United Kingdom, ²Department of Neurodegenerative Disease, University College London, London, United Kingdom, ³Université Côte d’Azur, Valbonne, France

Longitudinal measurements of brain atrophy using structural T1-weighted MRI (sMRI) can provide powerful biomarkers for clinical trials in neurodegenerative diseases. Here we use the latest advances in disease progression modelling, specifically the Gaussian Process Progression Model (GPPM), to untangle the effects of inter-subject variability, measurement noise and individual disease stage on longitudinal sMRI measurements in Huntington’s disease (HD). We use GPPM to estimate, for the first time, the relative timescale of sub-cortical atrophy in HD, and identify when sMRI provides additional information to genetics. We conclude that GPPM could increase power over standard imaging biomarkers for clinical trials in HD.