Yuchen Pei¹, Fenqiang Zhao¹, Liangjun Chen¹, Zhengwang Wu¹, Tao Zhong¹, Ya Wang¹, Li Wang¹, He Zhang², and Gang Li^1
1Department of Radiology and BRIC, University of North Carolina at Chapel Hill, USA, Chapel Hill, NC, United States, ²Department of Radiology, Obstetrics and Gynecology Hospital, Fudan University, Shanghai, China, Shanghai, China

Brain atlases are of fundamental importance for analyzing the dynamic neurodevelopment in fetal brains. Since the brain size, shape, and structure change rapidly during the prenatal development, it is essential to construct a spatiotemporal (4D) atlas with tissue probability maps for accurately characterizing dynamic changes in fetal brains and providing tissue prior for segmentation of fetal brain MR images. We propose a novel unsupervised learning framework for building multi-channel atlases by incorporating tissue segmentation. Based on 98 healthy fetuses from 22 to 36 weeks, the learned 4D fetal brain atlas includes intensity templates, corresponding tissue probability maps and parcellation maps.

Yue Sun¹, Kun Gao¹, Shihui Ying¹, Weili Lin¹, Gang Li¹, Sijie Niu¹, Mingxia Liu¹, and Li Wang^1
1Department of Radiology and Biomedical Research Imaging Center, UNC at Chapel Hill, Chapel Hill, NC, United States

This study develops a self-supervised transfer learning (SSTL) framework to generate reliable cerebellum segmentations for infant subjects with multi-domain MRIs, aiming to alleviate the domain shift between different time-points/sites and improve the generalization ability. Experiments demonstrate that by transferring limited manual labels from late time-points (or a specific site) with high tissue contrast to early time-points (or other sites) with low contrast, our method achieves improved performance and can be applied to other tasks, especially for those with multi-site data. 

Sumeet Shinde¹, Abhilasha Indoria², Jitender Saini², Manish Beniwal², Vani Santosh², and Madhura Ingalhalikar^1
1Symbiosis centre for medical image analysis, Symbiosis International University, Pune, India, ²Dept of Radiology, National Institute of Mental Health and Neurosciences, Bengaluru, India

Recent developments in glioma subtyping suggest that IDH genotype as well as the histological grading are both crucial factors. However, earlier classification studies based on MRI features have focused either only on grade or IDH. In this work we employ an automated deep learning based technique to delineate the grade as well as the IDH status on a dataset of 178 subjects. Our classifier performs with a superior accuracy of 93.5% and the model explanability is achieved through class activation maps that illustrate the areas important in the classification.  

Eduardo Diniz¹, Karim Helmet², Tales Santini², Howard Eizenstein², and Tamer Ibrahim^2
1Electrical and Computer Engineering, University of Pittsburgh, Pittsburgh, PA, United States, ²Bioengineering, University of Pittsburgh, Pittsburgh, PA, United States

We implemented a generative adversarial network (CycleGAN) to tackle the problem of MRI data harmonization across scanner strength. We leveraged a large dataset of unpaired 3T and 7T MR images for training and evaluated our model in a dataset of paired 3T and 7T data by generating synthetic 7T images and comparing them with their real counterparts. Dice scores and volumetric measures showed strong agreement between the synthetic and real 7T images. This approach allows for research studies to transition from 3T to 7T MR systems, thereby harnessing 7T systems advantages without losing the prior wave’s 3T MR data.

Lavanya Umapathy¹, Gloria Guzman Perez-Carillo², Blair Winegar³, Srinivasan Vedantham⁴, Maria Altbach⁴, and Ali Bilgin^1,4,5
1Electrical and Computer Engineering, University of Arizona, Tucson, AZ, United States, ²Mallinckrodt Institute of Radiology, St Louis, MO, United States, ³Radiology and Imaging Sciences, University of Utah, Salt Lake, UT, United States, ⁴Medical Imaging, University of Arizona, Tucson, AZ, United States, ⁵Biomedical Engineering, University of Arizona, Tucson, AZ, United States

The relationship between cognition and white matter hyperintensities (WMH) volumes often depends on accuracy of the lesion segmentation algorithm used. As such, accurate detection and quantification of WMH is of great interest. Here, we use a deep learning-based WMH segmentation algorithm, StackGen-Net, to detect and quantify WMH on 3D-FLAIR images from ADNI. We used a subset of subjects (n=20) and obtained manual WMH segmentations by an experienced neuro-radiologist to demonstrate the accuracy of our algorithm. On a larger cohort of subjects (n=290), we observed larger WMH volumes correlated with worse performance on executive function (P=.004), memory (P=.01), and language (P=.005).

Sabrina Gjerswold-Selleck¹, Nanyan Zhu^2,3,4, Haoran Sun¹, Dipika Sikka¹, Jie Shi¹, Chen Liu^3,4,5, Tal Nuriel^4,6, Scott A. Small^4,7,8, and Jia Guo^3,9
1Biomedical Engineering, Columbia University, New York, NY, United States, ²Biological Science, Columbia University, New York, NY, United States, ³Mortimer B. Zuckerman Mind Brain Behavior Institute, Columbia University, New York, NY, United States, ⁴Taub Institute for Research on Alzheimer’s Disease and the Aging Brain, Columbia University, New York, NY, United States, ⁵Electrical Engineering, Columbia University, New York, NY, United States, ⁶Department of Pathology and Cell Biology, Columbia University, New York, NY, United States, ⁷Gertrude H. Sergievsky Center, Columbia University, New York, NY, United States, ⁸Radiology, Columbia University, New York, NY, United States, ⁹Psychiatry, Columbia University, New York, NY, United States

Brain extraction plays an integral role in image processing pipelines in both human and small animal preclinical MRI studies. Due to lack of state-of-the-art tools for automated brain extraction in rodent research, this step is often performed semi-supervised with manual correction, making it prone to inconsistent results. Here, we perform a multi-model brain extraction study and present a semi-automated preprocessing workflow and deep neural network with a 3D Residual Attention U-Net architecture as the optimal network for automated skull-stripping in neuroimaging analysis pipelines, achieving a DICE score of 0.987 and accuracy of 99.7%.

Nanyan Zhu^1,2,3, Chen Liu^2,3,4, Sabrina Gjerswold-Selleck⁵, Xinyang Feng⁵, Dipika Sikka⁵, Scott A. Small^2,6,7, and Jia Guo^3,8
1Biological Science, Columbia University, New York, NY, United States, ²Taub Institute for Research on Alzheimer's Disease and the Aging Brain, Columbia University, New York, NY, United States, ³Mortimer B. Zuckerman Mind Brain Behavior Institute, Columbia University, New York, NY, United States, ⁴Electrical Engineering, Columbia University, New York, NY, United States, ⁵Biomedical Engineering, Columbia University, New York, NY, United States, ⁶Radiology, Columbia University, New York, NY, United States, ⁷Gertrude H. Sergievsky Center, Columbia University, New York, NY, United States, ⁸Psychiatry, Columbia University, New York, NY, United States

Alzheimer’s disease (AD) is a neurodegenerative disorder where functional deficits precede structural deformations. Various studies have demonstrated the efficacy of deep learning in diagnosing AD using imaging data, and that functional modalities are more helpful than structural counterparts over comparable sample size. To deal with the lack of large-scale functional data in the real world, we used a structure-to-function translation network to artificially generate a previously non-existent spatially-matched functional neuroimaging dataset from existing large-scale structural data. The artificial functional data, generated with little cost, complemented the authentic structural data to further improve the performance of AD classification.

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

A subset of 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 preoperative MR radiomics and apparent diffusion coefficient (ADC) value for prediction of P/R in meningiomas. The four most significant radiomic features selected by support vector machine (SVM) were used to calculate SVM score for each patient. High SVM score and low ADC value were associated with early P/R and short progression-free survival in meningiomas.

Ruihao Liu¹, Ziyu Meng¹, Wenli Li¹, Yao Li¹, Yiping P. Du¹, and Zhi-Pei Liang^2,3
1Institute for Medical Imaging Technology, School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai, China, ²Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, Urbana, IL, United States, ³Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign, Urbana, IL, United States

Accurate segmentation of brain tissues is essential for brain imaging applications. Classical Bayesian methods rely on good probability functions to produce good segmentation results. This paper presents a new method to synergistically integrate classical Bayesian segmentation with deep learning-based classification. A cluster of patch-based position-dependent neural networks were trained to effectively capture the joint spatial-intensity distributions of brain tissues. This cluster of patch networks significantly extends the capability of classical Markov Random field models and conventional statistical brain atlases. By combining the classical Bayesian classifier with the proposed networks, our method significantly improved segmentation results compared with the state-of-art methods.

Kattie Sepehri¹, Xiaowei Song², Ryan Proulx³, Sujoy Ghosh Hajra⁴, Brennen Dobberthien⁵, Careesa Liu⁶, Ryan D'Arcy⁷, Don Murray³, and Andra Krauze^5
1UBC, Vancouver, BC, Canada, ²Surrey Memorial Hospital, Vancouver, BC, Canada, ³Safe Software, Vancouver, BC, Canada, ⁴National Research Council, Vancouver, BC, Canada, ⁵BC Cancer, Vancouver, BC, Canada, ⁶Baycrest Health Sciences Centre, Vancouver, BC, Canada, ⁷HealthTech Connex, Vancouver, BC, Canada

Robust machine learning algorithms for tumor identification require ground truth data sets. Ground truth data sets require expert input, are difficult and inefficient to produce. Feature Manipulation Engine (FME) allows for specific and complex data manipulation. We have created an FME workflow to produce simulated tumors that resemble realistic gliomas as rated by experts. 

Peng Xia¹, Zuojun Wang¹, Han Yu¹, Fan Huang¹, Henry Ka-Fung MAK¹, Edward Sai-am HUI², and Peng Cao^1
1Department of Diagnostic Radiology, The University of Hong Kong, Hong Kong, China, ²Department of Rehabilitation Science, The Hong Kong Polytechnic University, Hong Kong, China

Cerebral microbleeds (CMBs) detection has been applied to many brain diseases but not well studied on stroke cohort. We developed and analyzed a deep learning-based CMB detection pipeline on quantitative susceptibility mapping (QSM) images from 138 stroke patients cohort. 

Maximilian Sackl¹, Alina Dima², Christian Payer², Darko Štern³, Reinhold Schmidt¹, and Stefan Ropele^1
1Department of Neurology, Medical University of Graz, Graz, Austria, ²Institute of Computer Graphics and Vision, Graz University of Technology, Graz, Austria, ³Department of Biophysics, Medical University of Graz, Graz, Austria

Segmentation of the hippocampal formation on T1-weighted structural MR scans is a prerequisite for most imaging studies in Alzheimer’s disease. In this work, we evaluated the performance and accuracy of deep learning-based hippocampus segmentation combined with manual ground truth (GT) data that originates from high-resolution T2-weighted MR images. Results were evaluated against the GT-labels and compared to segmentation results obtained with FreeSurfer. All learning approaches outperformed FreeSurfer in terms of accuracy and speed, where experiments utilizing the T2-based GT-labels yielded the best results. Thus, using T2-weighted images for training a deep learning model can improve automated HC segmentation.

Fan Huang¹, Peng Xia¹, Varut Vardhanabhuti¹, Edward Sai-Kam Hui², Gary Kui-Kai Lau³, Henry Ka-Fung Mak¹, and Peng Cao^1
1Department of Diagnostic Radiology, The University of Hong Kong, Hong Kong, Hong Kong, ²Department of Rehabilitation Science, The Hong Kong Polytechnic University, Hong Kong, Hong Kong, ³Department of Medicine, The University of Hong Kong, HongKong, Hong Kong

We propose a semi-supervised training scheme for white matter hyperintensity (WMHs) segmentation using V-Net on FLAIR images. The training procedure does not require manual labeling data but only a few domain knowledge of WMHs. The segmentation result obtained by the V-Net with the proposed scheme outperformed that obtained by the supervised loss with manual labels, showing great potential and generalizability in medical image applications.

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

A subset of nonfunctioning pituitary macroadenomas (NFPAs) may show early progression/recurrence (P/R) after surgery. In clinical practice, one of the main challenges in the treatment of NFPAs is to determine factors associated with P/R. This study investigated the role of preoperative MR radiomics based on support vector machine (SVM) for the prediction of P/R in NFPAs. The three most significant radiomic features selected by SVM were used to calculate SVM score for each patient. High SVM scores were associated with P/R and short progression-free survival in NFPAs.

Jacob Ellison¹, Francesco Caliva¹, Pablo Damasceno², Tracy Luks¹, Marisa LaFontaine¹, Julia Cluceru¹, Anil Kemisetti¹, Yan Li¹, Valentina Pedoia², Javier Villanueva-Meyer¹, and Janine M Lupo^1
1Radiology and Biomedical Imaging, UCSF, San Francisco, CA, United States, ²Center for Intelligent Imaging, UCSF, San Francisco, CA, United States

Routine monitoring of response to therapy in patients with glioma greatly benefits from using volumetrics quantified from lesion segmentation. Yet, the vast majority of deep learning models developed for this task have been trained using data from treatment-naïve, newly-diagnosed patients, whose T2-lesions have different appearance on imaging. We found that increasing the proportion of treated patients in training, incorporating a cross-entropy loss term that takes into account the spatial distance from surgical resection cavity and leading tumor edge, and transfer learning from newly-diagnosed to post-treatment imaging domains were effective strategies to improve the generalizability of segmentation of the T2-lesion post-treatment.

Zhuoying Ruan¹, Nan Mei², Yiping Lu², Ji Xiong³, Xuanxuan Li², Yajing Zhao², Pu-Yeh Wu⁴, Li Liu⁵, and Bo Yin^2
1Shanghai Institute of Medical Imaging, Fudan University, Shanghai, China, ²Department of Radiology, Huashan Hospital Fudan University, Shanghai, China, ³Department of Pathology, Huashan Hospital Fudan University, Shanghai, China, ⁴GE Healthcare, MR Research China, Shanghai, China, ⁵Department of Radiology, Shanghai Cancer Center, Fudan University, Shanghai, China

The prediction of overall survival in glioblastoma patients can provide great aid to clinical treatment. Therefore, we extracted clinical, VASARI and radiomic features from pre-operative MRI scannings and compared the predictive capacity of Cox algorithm and random forest based on these features. According to our results, tumor without cortical involvement had longer overall survival than those involved; the random forest models outperformed Cox regression models in general and the random forest model consisting of radiomic features was the best one. Ten radiomic features were reproducible in our and other’s studies exhibiting promising value in overall survival prediction of glioblastoma.

Kaiming Li¹ and Xiaoping Hu^1
1Department of Bioengineering, UC Riverside, Riverside, CA, United States

The brain is a complex dynamic system that constantly evolves. Characterization of the spatiotemporal dynamics of brain activity is fundamental to understand how brain works. Current studies with functional connectivity and linear models are limited by sacrificed temporal resolution and insufficient model capacity. With a generative variational auto encoder (VAE), the present study mapped the high-dimensional transient co-activity patterns (CAPs) of large datasets in a low-dimensional latent space. We demonstrated with multiple datasets that VAE can effectively represent the transient CAPs in latent space, paving the way for frame-wise modeling of the complex spatiotemporal dynamics in future. 

Yiwen Bao¹, Patrick Ka-Chun Chiu², Yat-Fung Shea², Joseph SK Kwan³, Felix Hon Wai Chan², and Henry Ka-Fung Mak^1
1Department of diagnostic radiology, University of Hong Kong, Hong Kong, Hong Kong, ²Department of medicine, Queen Mary Hospital, Hong Kong, Hong Kong, ³Department of brain sciences, Imperial College London, London, United Kingdom

Random forest model as a high efficacy classifier was incorporated in our study for supporting clinical diagnosis. We aimed at evaluating the accuracy of RF model in distinguishing HC, MCI from AD and the importance of various neuroradiological features in selection. Additionally, in order to unify quantitative amyloid uptake across three cohorts, we transformed SUVR into standard Centiloid unit. The results indicated that RF model had moderate to high accuracy in differentiating AD from HC and MCI. Regional Ab load had more important effects than other features in distinguishing AD from others. 

Christian Tinauer¹, Stefan Heber¹, Lukas Pirpamer¹, Anna Damulina¹, Reinhold Schmidt¹, Stefan Ropele¹, and Christian Langkammer^1
1Department of Neurology, Medical University of Graz, Graz, Austria

Using T1-weighted images we separated Alzheimer's patients (n=130) from healthy controls (n=375) by using a deep neural network and found that the preprocessing steps might introduce unwanted features to be used by the classifier. We systematically investigated the influence of registration and brain extraction on the learned features using a relevance map generator attached to the classification network. The results were compared to our relevance-guided training method. Relevance-guided training identifies sparser but substantially more relevant voxels, which improves the classification accuracy.

Hironobu Endo¹, Yuhei Takado¹, Kenji Tagai¹, Matsuoka Kiwamu¹, Manabu Kubota², Yasunori Sano¹, Keisuke Takahata¹, Maiko Ono¹, Chie Seki¹, Hideki Matsumoto^1,3, Oya Masaki¹, Yoko Ikoma¹, Kazunori Kawamura¹, Ming-rong Zhang¹, Hitoshi Shinotoh^1,4, Kenichi Oishi⁵, Susumu Mori⁵, Takahiko Tokuda¹, Hitoshi Shimada¹, and Makoto Higuchi^1
1National Institute of Radiological Sciences, National Institutes for Quantum and Radiological Science and Technology, Chiba, Japan, ²Department of Psychiatry, Kyoto University Graduate School of Medicine, Kyoto, Japan, ³Department of Oral and Maxillofacial Radiology, Tokyo Dental College, Tokyo, Japan, ⁴Neurology Clinic Chiba, Chiba, Japan, ⁵Department of Radiology and Radiological Science, Johns Hopkins University School of Medicine, Baltimore, MD, United States

MRI-based diagnostic marker is desired for recruiting patients with progressive supranuclear palsy (PSP) for clinical trials of tau-targeting therapies. Elastic Net analysis was applied to a set of 144 MRI-VOIs (MR images-volumes of interest) to determine the VOIs useful for discriminating tau-positive PSP from healthy controls. The MRI-PSP score, calculated from the analysis, demonstrated 90.0% and 91.1% accuracy for training and validation data. The PSP rating scale correlated well with the MRI-PSP score. The MRI-PSP score calculated by an unbiased analysis system would be a promising diagnostic marker and can potentially predict disease severity in PSP.