Jie Zhang^1,2, Xiaojing Long¹, Xin Feng², and Dong Liang^1
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhen, China, ²Chongqing University of Technology, ChongQing, China

The early diagnosis of AD is important for patient care and disease management. However, early diagnosis of AD is still challenging. In this work, we proposed a 3D attention model based densely connected Convolution Neural Network  to learn the multilevel features of MR brain images for AD classification and prediction. The proposed network was constructed with the emphasis on the interior resource utilization and introduced the attention mechanism into the classification of AD for the first time. Our results showed that the proposed model is effective for AD classification.

Jinseo Lee¹, Daniel McClement², Glen Pridham¹, Olayinka Oladosu¹, and Yunyan Zhang^1
1University of Calgary, Calgary, AB, Canada, ²University of British Columbia, Vancouver, BC, Canada

As deep learning technologies continue to advance, the availability of reliable methods to accurately interpret these models is critical. Based on a trained deep learning model (VGG19) for image classification, we have shown that methods using class activation mapping (CAM) and Grad-CAM have the potential to detect the most critical MRI feature patterns associated with relapsing remitting and secondary progressive multiple sclerosis, and healthy controls, and that these patterns seem to differentiate the two continuing subtypes of MS. This can help further understand the mechanisms of disease development and discover new biomarkers for clinical use.

Harri Merisaari^1,2,3, Pekka Taimen⁴, Otto Ettala⁵, Marko Pesola¹, Jani Saunavaara⁶, Anant Madabhushi³, Peter J Boström⁵, Hannu Aronen^1,6, and Ivan Jambor^1,7
1Department of Diagnostic Radiology, University of Turku, Turku, Finland, ²Department of Future Technologies, University of Turku, Turku, Finland, ³Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, ⁴Department of Pathology, Institute of Biomedicine, Turku, Finland, ⁵Department of Urology, Turku University Hospital, Turku, Finland, ⁶Medical Imaging Centre of Southwest Finland, Turku University Hospital, Turku, Finland, ⁷Department of Radiology, Icahn School of Medicine at Mount Sinai, New York, NY, United States

Acquisition time of a routine prostate MRI can be up to 20-25 minutes leading to significant financial burden on healthcare systems as the number of prostate MRI continue to increase.  We developed, validated and tested an open-source radiomics/texture tools for 5-minute biparametric prostate MRI (T2-weighed imaging and DWI obtained using 4 b-values (0, 900, 1100, 2000 s/mm2)) using whole mount prostatectomy sections of 157 men with prostate cancer, PCa, (244 PCa lesions). Best features were corner detectors with AUC (clinically insignificant vs insignificant prostate cancer) in the range of 0.82-0.89.  Code and data are available at: https://github.com/haanme/ProstateFeatures and http://mrc.utu.fi/data .

Yang Zhang¹, Yezhi Lin^1,2, Siwa Chan³, Jeon-Hor Chen^1,4, Jiejie Zhou², Daniel Chow¹, Peter Chang¹, Meihao Wang², and Min-Ying Su^1
1Department of Radiological Science, University of California, Irvine, CA, United States, ²Department of Radiology, First Affiliate Hospital of Wenzhou Medical University, Wenzhou, China, ³Department of Medical Imaging, Taichung Tzu-Chi Hospital, Taichung, Taiwan, ⁴Department of Radiology, E-Da Hospital and I-Shou University, Kaohsiung, Taiwan

A total of 244 patients were analyzed, 99 in Training, 83 in Testing-1 and 62 in Testing-2. Patients were classified into 3 molecular subtypes: TN, HER2+ and (HR+/HER2-). Deep learning using CNN and Convolutional Long Short Term Memory (CLSTM) were implemented. The mean accuracy in Training dataset evaluated using 10-fold cross-validation was higher using CLSTM (0.91) than CNN (0.79). When the developed model was applied to testing datasets, the accuracy was very low, 0.4-0.5. When transfer learning was applied to re-tune the model using one testing dataset, it could greatly improve accuracy in the other dataset from 0.4-0.5 to 0.8-0.9.

Jiaxin Li¹, Cheng Li², Xiran Jiang¹, Zhenkun Peng², Chaohe Zhang³, Qiegen Liu⁴, and Shanshan Wang^2
1China Medical University, Shenyang, China, ²Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China, ³Cancer Hospital of China Medical University, Shenyang, China, ⁴Department of Electronic Information Engineering, Nanchang University, Nanchang, China

End-to-end deep learning methods, such as the well-known U-Net, have achieved great successes in biomedical image segmentation tasks. These models are often fed with the full field of view images which may contain irrelevant organs or tissues influencing the segmentation performance. In this study, targeting at the accurate segmentation of rectal cancer lesions in T1-weighted MR images, we propose a two-stage deep learning method that is composed of a detection stage and a segmentation stage. Experimental results show that under the guidance of the detected bounding boxes, better segmentation performance is achieved. 

Yihang Zhou¹, Jing Yuan¹, Oi Lei Wong¹, Kin Yin Cheung¹, and Siu Ki Yu^1
1Medical Physics & Research Department, Hong Kong Sanatorium & Hospital, Hong Kong, China

Respiratory induced organ motion reduces radiation delivery accuracy of radiotherapy in thorax and abdomen. MR-guided-radiotherapy (MRgRT) is capable of continuous MRI acquisition during treatment. However, the latency due to MRI acquisition and reconstruction, the detection of tumor position change, and the interaction with multileaf collimator (MLC) have been identified as the major challenges for real-time MRgRT.  In this study, we proposed a deep-learning based 3D motion prediction technique to predict liver motion from volumetric dynamic MR images. Our algorithm showed promising results (< 0.3 cm prediction error on average) , suggesting its possibility of real-time motion tracking in the future MRgRT.

Yi Zhu¹, Rong Wei¹, Ge Gao², Jue Zhang¹, and Xiaoying Wang^2
1Peking University, Beijing, China, ²Peking University First Hospital, Beijing, China

In the wake of population aging, prostate cancer has become one of the most important diseases in elderly men. The low specificity in only image-based diagnosis may lead to unnecessary biopsies. Therefore, clinicians need to consider other variables to make diagnosis, such as age, PSA, and prostate volume. In this study we developed a novel 3D CNN model which combined clinical parameters and MR images for differentiating benign and malignant prostate lesions. The area under the receiver operating characteristics (ROC) of our proposed model (0.84) is significantly higher than that of traditional prediction model (0.71, P < 0.001).

Linfang Xiao^1,2, Yilong Liu^1,2, Peiheng Zeng^1,2, Mengye Lyu^1,2, Xiaodong Ma^1,2, Alex T. Leong^1,2, and Ed X. Wu^1,2
1Laboratory of Biomedical Imaging and Signal Processing, The University of Hong Kong, Hong Kong, China, ²Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong, China

Present MRI diagnosis comprises two steps: (i) reconstruction of multi-slice 2D or 3D images from k-space data; and (ii) pathology identification from images. In this study, we propose a strategy of direct pathology detection and characterization from MR k-space data through deep learning. This concept bypasses the traditional MR image reconstruction prior to pathology diagnosis, and presents an alternative MR diagnostic paradigm that may lead to potentially more powerful new tools for automatic and effective pathology screening, detection and characterization. Our simulation results demonstrated that this image-free strategy could detect brain tumors and their sizes/locations with high sensitivity and specificity.

Jeffrey Ma¹, Ukash Nakarmi¹, Cedric Yue Sik Kin¹, Joseph Y. Cheng¹, Christopher Sandino², Ali B Syed¹, Peter Wei¹, John M. Pauly², and Shreyas S Vasanawala^1
1Department of Radiology, Stanford University, Stanford, CA, United States, ²Department of Electrical Engineering, Stanford University, Stanford, CA, United States

In this abstract we investigate deep learning frameworks for medical image quality assessment and automatic classification of diagnostic and non-diagnostic quality images. 

Glen Pridham¹, Olayinka Oladosu², and Yunyan Zhang^1
1Department of Radiology, Department of Clinical Neurosciences, Hotchkiss Brain Institute, University of Calgary, Calgary, AB, Canada, ²Department of Neurosciences, Hotchkiss Brain Institute, University of Calgary, Calgary, AB, Canada

The Stockwell Transform (ST) is an advanced local spectral feature estimator, that is prohibitively large for use in machine learning applications for typical MR images. We compared two memory-efficient variants: the Polar ST (PST) and the Discrete Orthogonal ST (DOST) as feature extraction steps in competing random forest classifiers, built to classify white matter regions-of-interest as: lesion or normal-appearing. The DOST failed to out-perform guessing, whereas the PST: out-performed guessing, and improved the accuracy of an intensity-based random forest, achieving 88.8% accuracy. We conclude that the PST can complement MR intensity, whereas the DOST may not.

Jin Fang¹, Bin Zhang¹, Shuo Wang², and Shuixing Zhang^1
1The First Affiliated Hospital of Jinan University, Guangzhou, China, ²Key Laboratory of Molecular Imaging, Institute of Automation, Chinese Academy of Sciences, Beijing, China

 Radiomics is a promising methodology that automatically extracts high-dimensional features from imaging data for supplementary evaluation of prognosis. Herein, we developed radiomic signature based on pretreatment MRI, which can be used as a biomarker for risk stratification for disease-free survival (DFS) in patients with early-stage cervical cancer. This study provides a non-invasive and cost-effective preoperative predictive tool to identify the early stage cervical cancer patients with high possibility of recurrence or metastasis; and they may help to determine whether more intensive observation and aggressive treatment regimens should be administered, aim at assisting clinical treatment and healthcare decisions. 

Michael Liu^1,2, Richard Ha¹, Yu-Cheng Liu¹, Tim Duong², Terry Button², Pawas Shukla¹, and Sachin Jambawalikar^1
1Radiology, Columbia University, New York, NY, United States, ²Stony Brook University, Stony Brook, NY, United States

Quantitative measures of breast functional tumor volume are important response predictors of breast cancer undergoing chemotherapy. Automated segmentation networks have difficulty excluding non tumoral enhancing structures from their segmentations. Using a small small DCE-MRI dataset with coarse slice level labels to weakly supervised segmentation was able to exclude large portions of non tumor structures. Without manual pixel wise segmentation, our Class activation map based region proposer excluded 67% of non-tumoral voxels in a sagittal slice from downstream segmentation networks while maintaining 94% sensitivity.

R. Marc Lebel¹ and Daniel Litwiller^2
1Applications and Workflow, GE Healthcare, Calgary, AB, Canada, ²Applications and Workflow, GE Healthcare, New York, NY, United States

Convolutional neural networks are an emerging tool in medical imaging. Conventional CNNs accept an image as input and return a task-specific output (e.g., a filtered image, a disease probability). Conventional CNNs struggle to generalize or perform poorly when image data alone is insufficient to solve the problem. We propose three ways to incorporate relevant scan information into a CNN. The value of this method is demonstrated on rSOS image denoising, a previously unstable problem.

Abhejit Rajagopal¹, Andrew P. Leynes¹, Thomas Hope¹, and Peder E.Z. Larson^1
1Department of Radiology and Biomedical Imaging, University of California, San Francisco, San Francisco, CA, United States

We introduce a deep neural network scheme for predicting FDG-PET activity from a T1-weighted MR volume. This is useful for creating realistic anatomy-conforming synthetic PET data for prototyping of PET reconstruction algorithms, e.g. from abundant MR-only exam data. While deep networks can learn the average or nominal uptake patterns, in most cases, MR is ultimately incapable of fully predicting PET activity due to fundamental differences in the sensing modalities. We show, however, that these MR-derived “zero-dose” images can aid in differential contrast enhancement and visualization of PET by localizing and highlighting activity uniquely detected by the PET radiotracers.

Jin Zhu¹, Guang Yang^2,3, Tom Wong^2,3, Raad Mohiaddin^2,3, David Firmin^2,3, Jennifer Keegan^2,3, and Pietro Lio^1
1Department of Computer Science and Technology, University of Cambridge, Cambridge, United Kingdom, ²Cardiovascular Research Centre, Royal Brompton Hospital, London, United Kingdom, ³National Heart and Lung Institute, Imperial College London, London, United Kingdom

Three-dimensional late gadolinium enhanced (LGE) CMR plays an important role in scar tissue detection in patients with atrial fibrillation. Although high spatial resolution and contiguous coverage lead to a better visualization of the thin-walled left atrium and scar tissues, markedly prolonged scanning time is required for spatial resolution improvement. In this paper, we propose a convolutional neural network based unsupervised super-resolution method, namely USR-Net, to increase the apparent spatial resolution of 3D LGE data without increasing the scanning time. Our USR-Net can achieve robust and comparable performance with state-of-the-art supervised methods which require a large amount of additional training images.

Xiaoyang Liu^1,2 and Paul Bottomley^1,2
1Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD, United States, ²MR Research Division, Department of Radiology, Johns Hopkins University, Baltimore, MD, United States

Parameterized image synthesis is desirable for region-specific image contrast and protocol optimization. While this could arguably be performed with quantitative mapping and Bloch equation simulations, it is not well-suited for real-time cine MRI or MRI endoscopy in a real-time interventional application with views changing from frame-to-frame. Here we propose to use empirical data to calibrate signal behavior patterns from a specific MRI endoscopic device and pulse sequences to fit new acquisitions from which extra images and ultimately guide automated local contrast optimization.

Mahmut Yurt^1,2, Salman Ul Hassan Dar^1,2, Aykut Erdem³, Erkut Erdem³, and Tolga Çukur^1,2,4
1Department of Electrical and Electronics Engineering, Bilkent University, Ankara, Turkey, ²National Magnetic Resonance Research Center, Bilkent University, Ankara, Turkey, ³Department of Computer Engineering, Hacettepe University, Ankara, Turkey, ⁴Neuroscience Program, Sabuncu Brain Research Center, Bilkent University, Ankara, Turkey

For synthesis of a single target contrast within a multi-contrast MRI protocol, current approaches perform either one-to-one or many-to-one mapping. One-to-one methods take as input a single source contrast and learn representations sensitive to unique features of the given source. Meanwhile, many-to-one methods take as input multiple source contrasts and learn joint representations sensitive to shared features across sources. For enhanced synthesis, we propose a novel multi-stream generative adversarial network model that adaptively integrates information across the sources via multiple one-to-one streams and a many-to-one stream. Demonstrations on neuroimaging datasets indicate superior performance of the proposed method against state-of-the-art methods.

Alessandro Sciarra^1,2, Max Dünnwald^1,3, Hendrik Mattern², Oliver Speck^2,4,5,6, and Steffen Oeltze-Jafra^1,4
1MedDigit, Department of Neurology, Medical Faculty, Otto von Guericke University, Magdeburg, Germany, ²BMMR, Biomedical Magnetic Resonance, Otto von Guericke University, Magdeburg, Germany, ³Faculty of Computer Science, Otto von Guericke University, Magdeburg, Germany, ⁴Center for Behavioral Brain Sciences (CBBS), Magdeburg, Germany, ⁵German Center for Neurodegenerative Disease, Magdeburg, Germany, ⁶Leibniz Institute for Neurobiology, Magdeburg, Germany

In clinical routine acquisitions, the resolution in the slice direction is often worse than the in-plane resolution. Super-resolution techniques can help to retrieve the lack of information. Employing a conditional generative adversarial network (c-GAN), known as pix2pix for T1-w brain images, we were able to reconstruct downsampled images till 10-fold. The neural network was compared with the traditional bivariate interpolation method, and the results show that pix2pix is a valid alternative.

Li Huang¹, Xueming Zou^1,2,3, and Tao Zhang^1,2,3
1School of Life Science and Technology, University of Electronic Science and Technology of China, Chengdu, China, ²Key Laboratory for Neuroinformation, Ministry of Education, Chengdu, China, ³High Field Magnetic Resonance Brain Imaging Laboratory of Sichuan, Chengdu, China

The existing deep learning networks for MR super-resolution image reconstruction using standard 3D convolutional neural networks typically require a huge amount of parameters and thus excessive computational complexity. This has restricted the development of deeper neural networks for better performance. Here we propose a lightweight separable 3D convolution neural network for MR image super-resolution. Results show that our method can not only greatly reduce the amount of parameters and computational complexity but also improve the performance of image super-resolution. 

Francesco Cognolato^1,2, Steffen Bollmann^1,3, and Markus Barth^1,3
1The University of Queensland, Brisbane, Australia, ²Technical University of Munich, Munich, Germany, ³ARC Training Centre for Innovation in Biomedical Imaging Technology, The University of Queensland, Brisbane, Australia

In our abstract we present QSMResGAN, a conditional Generative Adversarial Network (cGAN) with a novel architecture for the generator (ResUNet), trained only with simulated data of different shapes to solve the dipole inversion problem for quantitative susceptibility mapping (QSM). The network has been compared with other state-of-the-art QSM methods on the QSM challenge 2.0 and on in vivo data.

Haining Wei¹, Jiang Liu², Enhao Gong³, Stefan Skare⁴, and Greg Zaharchuk^5
1Tsinghua University, Beijing, China, ²The Johns Hopkins University, Baltimore, MD, United States, ³Subtle Medical Inc., Menlo Park, CA, United States, ⁴Karolinska Institutet, Stockholm, Sweden, ⁵Stanford University, Stanford, CA, United States

EPImix is a one-minute full brain magnetic resonance exam using a multicontrast echo-planar imaging (EPI) sequence, which can generate six contrasts at a time. However, the low resolution and signal-to-noise ratio impeded its application. And EPI distortion makes it harder to improve the quality of imaging. In this study, we applied topup for distortion correction and propose a supervised deep learning model to enhance the EPImix images. We test the GRE, T2, T2FLAIR images on network and analyze the peak signal to noise ratio and structural similarity results. The results suggest that the proposed method can effectively enhance EPImix images.

Gregory Kuling¹, Matt Hemsley ^1,2, Geoff Klein¹, Philip Boyer³, and Marzyeh Ghassemi^4
1Medical Biophysics, University of Toronto, Toronto, ON, Canada, ²Physical Sciences Platform, Sunnybrook Research Institute, Toronto, ON, Canada, ³Institute of Biomaterials and Biomedical Engineering, University of Toronto, Toronto, ON, Canada, ⁴Computer Science and Medicine, University of Toronto, Toronto, ON, Canada

Performance of machine learning models for medical image segmentation is often hindered by a lack of labeled training data. We present a method for data augmentation wherein additional training examples are synthesized using a conditional generative adversarial network (cGAN) conditioned on a ground truth segmentation mask. The mask is later used as a label during the segmentation task. Using a dataset of N=48 T2-weighted MR volumes of the prostate, our results demonstrate the mean DSC score of a U-Net prostate segmentation model increased from 0.74 to 0.76 when synthetic training images are included with real data.

Hanbyol Jang¹, Sewon Kim¹, Jinseong Jang¹, Young Han Lee², and Dosik Hwang*^1
1School of Electrical and Electronic Engineering, Yonsei University, Seoul, Republic of Korea, ²Yonsei University College of Medicine, Seoul, Republic of Korea

The goal of this study is to make new contrast image from multiple contrast Magnetic Resonance Image (MRI) using deep learning with loss function specialized for multiple image processing. Our contrast-conversion deep neural network (CC-DNN) is an end-to-end architecture that trains the model to create one image (STIR image) from three images (T1-weighted, T2-weighted, and GRE images). And we propose a new loss function to take into account intensity differences, misregistration, and local intensity variations.

Max W.K. Law¹, Oilei Wong¹, Jing Yuen¹, and S.K. Yu^1
1Hong Kong Sanatorium & Hospital, Hong Kong, Hong Kong

This study presented a hyperparameter-free deep network modal for cranial pseudo-CT generation. The model was potentially universal to various scanning machines without the need of network hyperparameter adjustment and could handle testing images from MR- and CT-simulators different from the training data. It is beneficial to perform clinical trial in institutions where multiple MR- and CT-machines are in operations,  without supervision by deep learning experts. The proposed model was examined using training and testing datasets acquired from two sets of MR- and CT-simulators, showing promising accuracy, <79 mean-absolute-error and <170 root-mean-squared-error.

Liangdong Zhou¹, Jinwei Zhang^1,2, Qihao Zhang^1,2, Pascal Spincemaille¹, Thanh D Nguyen¹, Yi Wang^1,2, and Liangdong Zhou^3
1Weill Medical School of Cornell University, New York, NY, United States, ²Cornell University, Ithaca, NY, United States, ³Radiology, Weill Medical School of Cornell University, New York, NY, United States

Perfusion parameters, including blood flow (BF), apparent blood velocity (V), blood volume (BV) and arterial transit time (ATT) are useful for the disgnosis of many dieases. Typically, perfusion quantification methods utilize the tracer concentration (ASL, DEC, DSC, etc.) as input and blood flow map as output. We proposed a deep learning-based perfusion parameters mapping (DL-PPM), which uses 4D time-revolved tracer concentration as input and perfusion parameters (BF, V, BV, ATT) as output. We tested the propose method using simulated data and in vivo data in kidney.

Andrew P. Leynes^1,2, Abhejit Rajagopal¹, Valentina Pedoia^1,2, and Peder E.Z. Larson^1,2
1Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States, ²UC Berkeley - UC San Francisco Joint Graduate Program in Bioengineering, Berkeley and San Francisco, CA, United States

We investigated the use of a Bayesian deep auto-encoder to visualize intrinsic variations within a dataset in image-space. The variations were visualized by calculating a voxel-wise standard deviation over the predictions of the Bayesian deep auto-encoder. The low mutual information that was measured between the MRI and the standard deviation maps suggests that new information is contained in the standard deviation maps. This may be useful in the training of deep learning models for anomaly detection.

Bragi Sveinsson^1,2, Akshay Chaudhari³, Bo Zhu^1,2, Neha Koonjoo^1,2, and Matthew Rosen^1,2
1Massachusetts General Hospital, Boston, MA, United States, ²Harvard Medical School, Boston, MA, United States, ³Stanford University, Stanford, CA, United States

The osteoarthritis initiative (OAI) performed several morphological MRI scans on both knees of a large patient cohort, but only acquired T2 maps in the right knee of most patients. We train a conditional GAN to use the morphological scans acquired in both knees to predict the T2 map, using the acquired T2 map in the right knee as a training target. Post-training, we apply the network to predict T2 values in the left knee, without an acquired T2 map.

Mario Serrano-Sosa¹, Christine DeLorenzo^1,2, and Chuan Huang^1,2,3
1Biomedical Engineering, Stony Brook University, Stony Brook, NY, United States, ²Psychiatry, Stony Brook University, Stony Brook, NY, United States, ³Radiology, Stony Brook University, Stony Brook, NY, United States

We developed a new PET denoising model by utilizing a dilated CNN (dNet) architecture with PET/MRI inputs (dNet_PET/MRI) and compared it to three other deep learning models with objective imaging metrics Structural Similarity index (SSIM), Peak signal-to-noise ratio (PSNR) and mean absolute percent error (MAPE). The dNet_PET/MRI performed the best across all metrics and performed significantly better than uNet_PET/MRI (p_SSIM=0.0218, p_PSNR=0.0034, p_MAPE=0.0305). Also, dNet_PET performed significantly better than uNet_PET (p<0.001 for all metrics). Trend-level improvements were found across all objective metrics in networks using PET/MRI compared to PET only inputs within similar networks (dNet_PET/MRI vs. dNet_PET and uNet_PET/MRI vs. uNet_PET).

Junshen Xu¹, Molin Zhang¹, Esra Turk², Polina Golland^1,3, P. Ellen Grant^2,4, and Elfar Adalsteinsson^1,5
1Department of Electrical Engineering & Computer Science, MIT, Cambridge, MA, United States, ²Fetal-Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Boston, MA, United States, ³Computer Science and Artificial Intelligence Laboratory (CSAIL), MIT, Cambridge, MA, United States, ⁴Harvard Medical School, Boston, MA, United States, ⁵Institute for Medical Engineering and Science, MIT, Cambridge, MA, United States

Estimating fetal pose from 3D MRI has a wide range of applications including fetal motion tracking and prospective motion correction. Fetal pose estimation is challenging since fetuses may have different orientation and body configuration in utero. In this work, we propose a method for fetal pose estimation from low-resolution 3D EPI MRI using generative adversarial network. Results show that the proposed method produces a more robust estimation of fetal pose and achieves higher accuracy compared with conventional convolution neural network.

Karim Armanious^1,2, Sherif Abdulatif¹, Vijeth Kumar¹, Tobias Hepp², Bin Yang¹, and Sergios Gatidis^2
1University of Stuttgart, Stuttgart, Germany, ²University Hospital Tübingen, Tübingen, Germany

MRI suffers from incomplete information and localized deformations due to a manifold of factors. For example, metallic hip and knee replacements yield local deformities in the resultant scans. Other factors include, selective reconstruction of data and limited fields of views. In this work, we propose a new deep generative framework, referred to as IPA-MedGAN, for the inpainting of missing or complete information in brain MR. This framework aims to enhance the performance of further post-processing tasks, such as PET-MRI attenuation correction, segmentation or classification. Quantitative and qualitative comparisons were performed to illustrate the performance of the proposed framework.

Edgar Rios Piedra^1,2, Morteza Mardani^1,2, and Shreyas Vasanawala^1,2
1Department of Radiology, Stanford University, Stanford, CA, United States, ²Department of Electrical Engineering, Stanford University, Stanford, CA, United States

Automated segmentation of kidneys and their sub-components is a challenging problem, particularly in pediatric patients and in the presence of anatomical deformations or pathology. We present an improved segmentation framework using a multi-channel U-Net with added attention block that allows for the automated segmentation of the multi-phase DCE-MRI of kidneys as well as a functional evaluation of the glomerular filtration rate. Results achieve an average Dice similarity coefficient of 0.912, 0.853, and 0.917 for kidney cortex, medulla, and collector system, respectively.

Neha Vats¹, Virendra Kumar Yadav¹, Manish Awasthi¹, Dinil Sasi¹, Mamta Gupta², Rakesh Kumar Gupta², and Anup Singh^1,3
1Centre for Biomedical Engineering, Indian Institute of Technology, Delhi, New Delhi, India, ²Department of Radiology, Fortis Memorial Research Institute, Gurugram, India, ³Biomedical Engineering, AIIMS, New Delhi, India

Segmentation of contrast enhancing tumor region from post-contrast T1-W MR images is sometime difficult due to low enhancement or presence of infarct tissue around or inside tumor, which exhibits similar intensity as contrast enhancement. Relative difference map obtained from pre-and post-contrast T1-weighted images can increase sensitivity to enhancement visualization as well as clearly differentiate infarct tissue from enhancing lesion. The objective of the current study was to evaluate accuracy of segmentation of contrast enhancing lesion using Support Vector Machine (SVM) classifier developed on relative difference map intensities. Optimized SVM classifier enabled accurate segmentation of contrast enhancing tumor lesion.

Zihao Tang¹, Kain Kyle², Michael H Barnett^2,3, Ché Fornusek⁴, Weidong Cai¹, and Chenyu Wang^2,3
1School of Computer Science, University of Sydney, Sydney, Australia, ²Sydney Neuroimaging Analysis Centre, Sydney, Australia, ³Brain and Mind Centre, University of Sydney, Sydney, Australia, ⁴Discipline of Exercise and Sport Science, Faculty of Medicine and Health, University of Sydney, Sydney, Australia

Robust and accurate MRI-based thigh muscle segmentation is critical for the study of longitudinal muscle volume change. However, the performance of traditional approaches is limited by morphological variance and often fails to exclude intramuscular fat. We propose a novel end-to-end semantic segmentation framework to automatically generate muscle masks that exclude intramuscular fat using longitudinal T1-weighted MRI scans. The architecture of the proposed U-shaped network follows the encoder-decoder network design with integrated residual blocks and attention gates to enhance performance. The proposed approach achieves a performance comparable with human imaging experts.        

Dhanunjaya Mitta¹, Soumick Chatterjee^1,2,3, Oliver Speck^2,4,5,6, and Andreas Nürnberger^1,3
1Faculty of Computer Science, Otto von Guericke University, Magdeburg, Germany, ²Department of Biomedical Magnetic Resonance, Otto von Guericke University, Magdeburg, Germany, ³Data & Knowledge Engineering Group, Otto von Guericke University, Magdeburg, Germany, ⁴Center for Behavioral Sciences, Magdeburg, Germany, ⁵German Center for Neurodegenerative Disease, Magdeburg, Germany, ⁶Leibniz Insitute for Neurobiology, Magdeburg, Germany

Image segmentation is a process of dividing an image into multiple coherent regions. Segmentation of biomedical images can assist diagnosis and decision making. Manual segmentation is time consuming and requires expert knowledge. One solution is to segment medical images by using deep neural networks, but traditional supervised approaches need a large amount of manually segmented training data. A possible solution for the above issues is unsupervised medical image segmentation using deep neural networks, which our work tries to solve with our proposed 3D Attention W-Net.

Andre Maximo¹, Chitresh Bhushan², Dattesh D. Shanbhag³, Radhika Madhavan², Desmond Teck Beng yeo², and Thomas Foo^2
1GE Healthcare, Rio de Janeiro, Brazil, ²GE Research, Niskayuna, NY, United States, ³GE Healthcare, Bengaluru, India

It is common practice to use dropout layers on U-net segmentation deep-learning models, and it is usually desirable to measure uncertainty of a deployed model while inferencing in clinical scenario. We present a method to convert a pre-trained model to a Bayesian model that can estimate uncertainty by posterior distribution of its trained weights. Our method uses both regular dropouts and converted Monte-Carlo dropouts to estimate uncertainty via cosine similarity of fixed and stochastic predictions.  It can identify cases differing from training set by assigning high uncertainty and can be used to ask for human intervention with tough cases.

Hariharan Ravishankar¹, Chitresh Bhushan², Arathi Sreekumari¹, and Dattesh D Shanbhag^1
1GE Healthcare, Bangalore, India, ²GE Global Research, Niskayuna, NY, United States

Most of the advancements with deep learning have come from mapping the reconstructed MRl images to outcomes (e.g. tumor segmentation, survival rate, pathology risk map). In this work, we present methods to arrive at critical medical imaging tasks like segmentation, classification directly from raw k-space data without image reconstruction. We specifically demonstrate that from k-space MRI data, we can perform hippocampus segmentation as well as detection of motion affected scans with similar performance to that obtained from imaging data. We also demonstrate that such an approach is more resilient to localized artifacts (e.g signal loss in hippocampus due to metal).

Daming Shen^1,2, Justin J Baraboo^1,2, Brandon C Benefield³, Daniel C Lee^2,3, Michael Markl^1,2, and Daniel Kim^1,2
1Biomedical Engineering, Northwestern University, Evanston, IL, United States, ²Radiology, Northwestern University Feinberg School of Medicine, Chicago, IL, United States, ³Feinberg Cardiovascular Research Institute, Northwestern University Feinberg School of Medicine, Chicago, IL, United States

Late gadolinium enhanced (LGE) CMR is the gold standard test for assessment of myocardial scarring. While quantifying scar volume is helpful to clinical decision making, its lengthy image segmentation time limits its use in practice. The purpose of this study is to enable fully automated LGE image segmentation using deep learning (DL) and explore a more efficient way of using annotation.

Yasmina Al Khalil¹, Cristian Lorenz², Jürgen Weese², and Marcel Breeuwer^1,3
1Biomedical Engineering Department, Eindhoven University of Technology, Eindhoven, Netherlands, ²Philips Research Laboratories, Hamburg, Germany, ³Philips Healthcare, MR R&D - Clinical Science, Best, Netherlands

The versatility of MRI acquisition parameters and sequences can have a substantial impact on the design and performance of medical image segmentation algorithms. Even though recent studies report excellent results of deep-learning (DL) based algorithms for tissue segmentation, their generalization capability and sequence dependence is rarely addressed, while being crucial for inclusion in clinical settings. This study attempts to demonstrate the lack of adaptation of such algorithms to unseen data from different sites and scanners. For this purpose, we use a 3D U-Net trained for brain tumor detection and test it site-wise to evaluate how well generalization can be achieved.

Yanping Xue^1,2, Hyungseok Jang¹, Zhenyu Cai¹, Hoda Shirazian¹, Mei Wu¹, Michal Byra¹, Yajun Ma¹, Eric Y Chang^1,3, and Jiang Du^1
1University of California, San Diego, San Diego, CA, United States, ²Beijing Chao-Yang Hospital, Beijing, China, ³VA San Diego Healthcare System, San Diego, CA, United States

The existence of short T2 tissues and high ordered collagen fibers in cartilage render it “invisible” to conventional MR and sensitive to the magic angle effect. Segmentation is the first step to obtain parameters of cartilage, which is often performed manually (time-consuming and variable). Automatic segmentation and providing a biomarker that visualizes both short and long T2 tissues and insensitive to the magic angle effect is desideratum. U-Net is based on CNN to process images. The purpose of this study is to describe and evaluate the pipeline of fully-automatic segmentation of cartilage and extraction of MMF in 3D UTE-Cones-MT modeling.

Chitresh Bhushan¹, Dattesh D Shanbhag², Andre de Alm Maximo³, Arathi Sreekumari², Dawei Gui⁴, Uday Patil², Brandon Pascual⁴, Rakesh Mullick², Teck Beng Desmond Yeo¹, and Thomas Foo^1
1GE Global Research, Niskayuna, NY, United States, ²GE Healthcare, Bangalore, India, ³GE Healthcare, Rio de Janeiro, Brazil, ⁴GE Healthcare, Waukesha, WI, United States

We demonstrate a deep learning-based workflow for intelligent slice placement (ISP) in MR knee imaging: meniscus plane, femoral condyle plane, tibial plane, sagittal plane and ACL plane, based on standard 2D tri-planar localizer images. We leveraged a previously described generalized architecture for ISP planning in brain, with only the training data and plane definitions adapted for knee. The mean absolute distance error between GT plane and predicted plane was < 0.5 mm for all planes except tibial plane (~ 1 mm). The results indicate the generalization of deep-learning ISP framework and its suitability for ISP in any anatomy of interest.

Rong Wei¹, Yi Zhu¹, Xiaoying Wang², Ge Gao², and Jue Zhang^1
1Peking University, Beijing, China, ²Peking University First Hospital, Beijing, China

In the wake of population aging, prostate cancer has become one of the most important diseases in western elderly men. T2-weighted-images (T2WIs) provides the best depiction of the prostate’s anatomy, and diffusion‐weighted images (DWIs) provides clues of cancer as prostate cancer appears as an area of high signal on DWIs. However, high b-value DWIs is affected by the susceptibility effects. In this paper，we describe and utilize a fully automatic 3D deep neutral network to correct misalignment between T2WIs and DWIs. The average increasement of Dice coefficient of before-and-after registration is 73% and 84%, which reflects the removal of susceptibility effects.

Molin Zhang¹, Junshen Xu¹, Esra Turk², Borjan Gagoski^2,3, P. Ellen Grant^2,3, Polina Golland^1,4, and Elfar Adalsteinsson^1,5
1Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ²Fetal-Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Boston, MA, United States, ³Harvard Medical School, Boston, MA, United States, ⁴Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology, Cambridge, MA, United States, ⁵Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA, United States

Fetal pose estimation could play an important role in fetal motion tracking or automatic fetal slice prescription by real-time adjustments of the prescribed imaging orientation based on fetal pose and motion patterns. In this abstract, we used a multiple image scale deep reinforcement learning method (DQN) to train an agent finding the target landmark of fetal pose by optimizing searching policy based on landmark features and its surroundings. Under an error tolerance of 15-mm, the detection accuracy reaches 58%.

Miller Fawaz¹, Saifeng Liu¹, David Utriainen¹, Sean Sethi¹, Zhen Wu¹, and E. Mark Haacke^1
1Magnetic Resonance Innovations, Inc., Bingham Farms, MI, United States

Automatic cerebral microbleed detection is attainable with our two step model for many disease states. We attributed previously shown lower performance in stroke data to different scenarios unique to stroke, including asymmetrically prominent cortical veins. We improved our existing pipeline for this detection by retraining the deep learning step of our model using stroke cases both in the acute and subacute stages. The results were improved performance in validation data in stroke cases as well as our previously tested data (multiple diseases). This makes our pipeline a viable and versatile real time automatic microbleed detection procedure.

Punith B Venkategowda¹, Asha K Kumaraswamy^1,2, Jonas Richiardi^3,4,5, Sanjeev Krishnan Thampi¹, Tobias Kober^3,4,5, Bénédicte Maréchal^3,4,5, and Ricardo A. Corredor-Jerez^3,4,5
1Siemens Healthcare Pvt. Ltd., Bangalore, India, ²Vidyavardhaka College of Engineering, Mysuru, India, ³Advanced Clinical Imaging Technology, Siemens Healthcare AG, Lausanne, Switzerland, ⁴Department of Radiology, Centre Hospitalier Universitaire Vaudois (CHUV) and University of Lausanne (UNIL), Lausanne, Switzerland, ⁵Signal Processing Laboratory (LTS 5), École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland

Deep learning techniques have proved their robustness in solving medical image analysis problems. This study proposes a conservative approach to benefit from the use of these methods to incrementally improve the performance of a well-established brain segmentation method. For this purpose, convolutional neural networks are trained to perform a reliable skull-stripping, based on weak labels of the original algorithm. The performance of the new pipeline is evaluated in a large cohort of dementia patients and healthy controls. The results present significant improvements in reproducibility and computation speed, while preserving accuracy and power of discrimination between groups.

Eugene Ozhinsky¹, Valentina Pedoia¹, and Sharmila Majumdar^1
1Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States

High quality scan prescription that optimally covers the area of interest with scan planes aligned to relevant anatomical structures is crucial for error-free radiologic interpretation. The goal of this project was to develop a machine learning pipeline for oblique scan prescription that could be trained on localizer images and metadata from previously acquired MR exams. To achieve that, we have developed a novel multislice rotational region-based convolutional neural network (MS-R²CNN) architecture and evaluated it on dataset of knee MRI exams.

Wilfred W Lam¹, Wendy Oakden¹, Elham Karami^1,2,3, Margaret M Koletar¹, Leedan Murray¹, Stanley K Liu^1,2,4, Ali Sadeghi-Naini^1,2,3,4, and Greg J Stanisz^1,2
1Sunnybrook Research Institute, Toronto, ON, Canada, ²University of Toronto, Toronto, ON, Canada, ³York University, Toronto, ON, Canada, ⁴Sunnybrook Health Sciences Centre, Toronto, ON, Canada

Saturation transfer-weighted images along with T₁ and T₂ maps at 7 T for 31 tumour xenografts in mice were used to automatically segment 1) tumour, 2) necrosis/apoptosis, 3) edema, and 4) muscle. Independent component analysis and Gaussian mixture modeling were used to segment these regions. Qualitatively excellent agreement was found between MRI and histopathology. An nine-image subset was identified that resulted in a 96% match in voxel labels compared to those found using the entire 24-image dataset. This subset had positive and negative predictive values of 96% and 97%, respectively, for tumour and 88% and 97%, respectively, for necrosis/apoptosis voxels.

Andre de Alm Maximo¹, Chitresh Bhushan², Dawei Gui³, Uday Patil⁴, and Dattesh D Shanbhag^4
1GE Healthcare, Rio de Janeiro, Brazil, ²GE Global Research, Niskayuna, NY, United States, ³GE Healthcare, Waukesha, WI, United States, ⁴GE Healthcare, Bangalore, India

In this work, we demonstrate a novel automated MRI scan plane prescription workflow by making use of the pre-scan calibrations scans to generate prescription planes for knee MRI planning. Using large-FOV, low-resolution 3D calibration data, we find the meniscus plane with very-high accuracy (angle error = 0.76, distance error = 0.07 mm).  The approach obviates the need to acquire any localizer images with potential benefits: (1) avoiding subsequent retakes for correct planning of plane prescription; (2) reducing total scan time; and (3) easing the MRI scanning experience for both patient and technologist by enabling single push scan.

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

We propose accessing the MRI quality, perceptual noise level in particular, during a scan to stop it when the image is good enough. A convolutional neural network is trained to map an image to a perceptual score. The label score for training is a statistical estimation of error standard deviation calibrated with radiologist inputs. Image rulers for different scan types are used in the inference phase to determine a flexible classification threshold. Our proposed training and inference methods achieve a 89% classification accuracy. The same framework can be used to tune the regularization parameter for compressed-sensing reconstructions.

Hongli Fan^1,2, Pan Su¹, Yang Li¹, Peiying Liu¹, Jay J. Pillai^1,3, and Hanzhang Lu^1,2,4
1The Russell H. Morgan Department of Radiology & Radiological Science, Johns Hopkins School of Medicine, Baltimore, MD, United States, ²Department of Biomedical Engineering, Johns Hopkins School of Medicine, Baltimore, MD, United States, ³Department of Neurosurgery, Johns Hopkins School of Medicine, Baltimore, MD, United States, ⁴F. M. Kirby Research Center for Functional Brain Imaging, Kennedy Krieger Institute, Baltimore, MD, United States

Arterial-Spin-Labeling (ASL) MRI has not been used widely in clinical practice because of lower SNR and the lack of ability to resolve cerebral-blood-flow (CBF) from bolus-arrival-time (BAT) effects¹. MR fingerprinting (MRF) ASL is a recently developed technique which has the potential to provide multiple parameters such as CBF, BAT, T₁ and cerebral-blood-volume (CBV) in one single scan^2-6. However, it still suffers from low SNR. The present work proposes a multi-band MRF-ASL in combination with deep learning, which can improve the reliability of MRF-ASL parametric maps up to 3-fold and provide whole-brain mapping of CBF and BAT in 4 minutes.

Peng Cao¹, Di Cui¹, Vince Vardhanabhuti¹, and Edward S. Hui^1
1Diagnostic Radiology, The University of Hong Kong, Hong Kong, China

We proposed a multi-layer perceptron deep learning method to achieve 100-fold acceleration for MRF quantification. 

Yong Chen^1,2, Zhenghan Fang³, and Weili Lin^1,2
1Radiology, University of North Carolina at Chapel Hill, Chapel Hill, NC, United States, ²Biomedical Research Imaging Center, University of North Carolina at Chapel Hill, Chapel Hill, NC, United States, ³Biomedical Engineering, University of North Carolina at Chapel Hill, Chapel Hill, NC, United States

In this study, we proposed a framework to generate simulated training dataset to train a convolutional neural network, which can be applied to highly undersampled MR Fingerprinting images to extract quantitative tissue properties. This eliminates the necessity to acquire training dataset from multiple subjects and has the potential to enable wide applications of deep learning techniques in quantitative imaging using MR Fingerprinting.

Jong Bum Son¹, Mark D. Pagel², and Jingfei Ma^1
1Imaging Physics Department, The University of Texas MD Anderson Cancer Center, Houston, TX, United States, ²Cancer Systems Imaging Department, The University of Texas MD Anderson Cancer Center, Houston, TX, United States

3D deep-learning neural networks can help ensure the slice-to-slice consistency. However, the performance of 3D networks may be degraded due to limited hardware. In this work, we developed a video domain transfer framework for 3D MRI processing to combine benefits of 2D and 3D networks with less graphical processing unit memory demands and slice-by-slice coherent outputs. Our approach consists of first translating “3D MRI images” to “a time-sequence of 2D multi-frame motion pictures,” then applying the video domain transfer to create temporally coherent multi-frame video outputs, and finally translating the output back to compose “spatially consistent volumetric MRI images.”

Radhika Madhavan¹, Andre Maximo², Chitresh Bhushan¹, Soumya Ghose¹, Dattesh D Shanbhag³, Uday Patil³, Matthew Frick⁴, Kimberly K Amrami⁴, Desmond Teck Beng Yeo¹, and Thomas K Foo^1
1GE Global Research, Niskayuna, NY, United States, ²GE Healthcare, Rio de Janeiro, Brazil, ³GE Healthcare, Bangalore, India, ⁴Mayo Clinic, Rochester, MN, United States

On model deployment, ideally deep learning models should be able learn continuously from new data, but data privacy concerns in medical imaging do not allow for ready sharing of training data. Retraining with incremental data generally leads to catastrophic forgetting. In this study, we evaluated the performance of a knee plane prescription model by retraining with incremental data from a new site. Increasing the number of incoming training data sets and transfer learning significantly improved test performance. We suggest that partial retraining and distributed learning frameworks may be more suitable for retraining of incremental data.

Or Perlman¹, Christian T Farrar¹, and Ouri Cohen^2
1Radiology, Athinoula A. Martinos Center for Biomedical Imaging, Charlestown, MA, United States, ²Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, United States

Chemical exchange saturation transfer MR fingerprinting (CEST-MRF) enables quantification of multiple tissue parameters. Optimization of the acquisition schedule can improve tissue discrimination and reduce scan times but is highly challenging because of the large number of acquisition and tissue parameters. The goal of this work is to demonstrate a scalable deep learning based global optimization method that provides schedules with improved discrimination. The benefits of our approach are demonstrated in an in vivo mouse tumor model.

Yi Xiao¹, Muheng Li¹, Ruizhi Liao^2,3, Tingyin Liu³, Junshen Xu², Esra Turk⁴, Borjan Gagoski^4,5, Karen Ying¹, Polina Golland^2,3, P.Ellen Grant^4,5, and Elfar Adalsteinsson^2,6
1Department of Engineering Physics, Tsinghua University, Beijing, China, ²Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, ³Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology, Cambridge, MA, United States, ⁴Fetal-Neonatal Neuroimaging and Developmental Science Center, Boston Children’s Hospital, Boston, MA, United States, ⁵Harvard Medical School, Boston, MA, United States, ⁶Institute for Medical Engineering and Science, Massachusetts Institute of Technology, Cambridge, MA, United States

Artifacts generated by severe and unpredictable fetal and maternal movements during MRI limit the success of imaging during pregnancy. Although modern clinical applications use single-shot imaging sequences, such as HASTE, to partially mitigate this problem, inter-slice motion artifacts are unavoidable and their impact on fetal imaging is not fully characterized. In order to analyze this problem, we exploit a large repository of volumetric EPI over long duration across many pregnant women to estimate the artifact load due to inter-slice motion on single-shot fetal brain MRI.

Bogdan Dzyubak¹, Joel P Felmlee¹, and Richard L. Ehman^1
1Radiology, Mayo Clinic, Rochester, MN, United States

Magnetic Resonance Elastography (MRE) accurately predicts fibrosis by measuring liver stiffness. The subjectivity in human analysis poses the biggest challenge to stiffness measurement reproducibility, and also complicates the training of a neural network to automate the task. In this work, we present a CNN-based stiffness measurement tool, giving special attention to training and validation in context of reader subjectivity. Compared to an older automated tool used by our institution in a reader-verified workflow, the CNN reduces ROI failure rate by 50%, and has an excellent agreement in measured stiffness with reader-verified target ROIs. 

Yuhang Shi^1
1Corporate Research Institute, United Imaging Healthcare, Shanghai, China

The implementation of shimming relies on phase unwrapping techniques to measure the underlying field inhomogeneity. This work presents a simple and effective deep learning based phase unwrapping method to accelerate the field mapping stage for shimming applications. The method significantly reduces the compulational time and shows similar performance compared to the conventional path-following method for shimming applications. 

Jeremy Kim¹, Thanh Nguyen², Pascal Spincemaille², and Yi Wang^2
1Hunter College High School, New York, NY, United States, ²Weill Cornell Medical College, New York, NY, United States

A novel deep neural network architecture, Deep Inversion Net, and a training scheme is proposed to accurately solve the multi-compartmental T2 relaxometry inverse problem for myelin water imaging in multiple sclerosis. Multiple neural networks communicate their outputs to regularize each other — thus better handling the ill-posed nature of this inverse problem. Results in simulated T2 relaxometry data and patients with demyelination show that Deep Inversion Net outperforms conventional optimization algorithms and other neural network architectures.

Kirsten Koolstra¹, Olga Dergachyova², Zidan Yu^2,3, Andrew Webb¹, and Martijn Cloos^2,3
1C.J. Gorter Center for High Field MRI, Department of Radiology, Leiden University Medical Center, Leiden, Netherlands, ²Center for Advanced Imaging Innovation and Research, Department of Radiology, New York University School of Medicine, New York, NY, United States, ³Sackler Institute of Graduate Biomedical Sciences, NYU Langone Health, New York, NY, United States

Simultaneous proton (¹H) and sodium (²³Na) acquisition can provide important metabolic information. However, proton data may suffer from off-resonance artifacts due to the long dwell time required to obtain sufficient SNR for ²³Na. In this work we use center outward and center inward image pairs to train a convolutional neural network that performs an off-resonance correction for the proton data without an additional measured field map.

Peter Stijnman^1,2, Cornelis van den Berg¹, and Alexander Raaijmakers^1,2
1Computational Imaging, UMC Utrecht, Utrecht, Netherlands, ²Biomedical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands

The RF safety assessment of implants is a computationally demanding task. An acceleration method was presented in [1] where the local field enhancement was determined by sparse matrix inversion. In this work, we show how a model-based deep learning approach for unrolled optimization could significantly reduce the number of iterations required. The benefit of this approach is that traditional minimization is still possible afterwards, combining short computation times with high accuracy. We trained 5 iterations with 10.000 randomly generated implants. The hybrid approach finds a numerically equivalent solution in$$$\,\frac{1}{13}^{th}\,$$$of the traditional method. This approach would enable online RF safety assessment.

Jun Li¹, Hongjian He¹, Yi-Cheng Hsu², and Jianhui Zhong^1,3
1Center for Brain Imaging Science and Technology, Department of Biomedical Engineering, Key Laboratory for Biomedical Engineering of Ministry of Education, Zhejiang University, Hangzhou, China, ²MR Collaboration, Siemens Healthcare Ltd, Shanghai, China, ³Department of Imaging Sciences, University of Rochester, Rochester, NY, United States

There is a need to obtain quantitative measures of tissue susceptibility in the form of susceptibility-weighted imaging (SWI). In this study, we used a deep neural network to generate QSM maps from SWI high pass (HP)–filtered phase images. Using the QSM maps reconstructed from mGRE data by iLSQR (mGRE iLSQR) as the ground truth, the QSM maps generated from SWI HP-filtered phase images by UNet (SWI UNet) resulted in lower residuals and a better performance in quantitative metrics compared with the QSM maps reconstructed from SWI HP-filtered phase images by iLSQR (SWI iLSQR).

Rula Amer¹, Jannette Nassar¹, David Bendahan², Hayit Greenspan¹, and Noam Ben-Eliezer^1,3,4
1Department of Biomedical Engineering, Tel Aviv University, Tel Aviv, Israel, ²CNRS, CRMBM, Aix Marseille University, Marseille, France, ³Center for Advanced Imaging Innovation and Research, New York University, New York, NY, United States, NY, United States, ⁴Sagol School of Neuroscience, Tel Aviv University, Ramat-Aviv, Israel

Quantification of subcutaneous fat infiltration in diseased muscle regions holds great prognostic value as it helps monitor the progression of muscular dystrophies. To estimate the infiltration, two stages are performed. The first isolates the region of muscle from the thigh/calf anatomy using U-net architecture for fully supervised segmentation. The second stage classifies muscle into diseased/healthy pixels using a weakly supervised segmentation method, incorporating a deep convolutional auto-encoder with triplet loss, and creates two clusters in the embedded feature space using k-means. The results showed a high Dice coefficient and a strong correlation between the fat infiltration and disease severity level.