Nikan K Namiri¹, Jinhee Lee¹, Bruno Astuto¹, Felix Liu¹, Rutwik Shah¹, Sharmila Majumdar¹, and Valentina Pedoia^1
1Department of Radiology and Biomedical Imaging and Center for Intelligent Imaging, University of California, San Francisco, San Francisco, CA, United States

We built an end-to-end deep learning model to rapidly stratify knees into morphological phenotypes using a large, longitudinal cohort with knee osteoarthritis (OA). We examined associations of phenotypes with odds of concurrent OA and OA progression. Bone, meniscus/cartilage, and inflammatory phenotypes were strongly associated with current structural OA and symptomatic OA. Hypertrophy phenotype was only weakly associated with structural OA. Among those who did not have baseline OA, bone and meniscus/cartilage phenotypes were strongly associated with developing both structural and symptomatic OA in 48 months. Only bone phenotype increased risk of undergoing total knee replacement surgery within 96 months.

Jinwoo Han¹, Suk-Joo Hong¹, Zepa Yang¹, Woo Young Kang¹, Yoonmi Choi¹, Chang Ho Kang², Kyung-sik Ahn², Baek Hyun Kim³, and Euddeum Shim^3
1Radiology, Korea University Guro Hospital, KUGH-MIDC, Seoul, Korea, Republic of, ²Korea University Anam Hospital, Seoul, Korea, Republic of, ³Korea University Ansan Hospital, Ansan, Korea, Republic of

Cartilage loss is fundamental pathology of knee osteoarthritis (OA). Quantitative analysis of cartilage thickness and volume is very time consuming by manual measurement. We proposed development of deep learning based cartilage segmentation at three dimensional knee magnetic resonance images, which can measure thickness and volume of knee joint cartilage, automatically and accurately. To evaluate the performance, we used Dice Similarity Coefficient (DSC) respect to the manual segmentation, and visual inspection. The accuracy DSC values were higher than 0.9. We expect deep learning program can be useful in future study for knee joint osteoarthritis.

Yongsheng Chen¹, Daniel Moiseev¹, Wan Yee Kong¹, Alexandar Bezanovski¹, and Jun Li^1,2
1Department of Neurology, Wayne State University School of Medicine, Detroit, MI, United States, ²John D. Dingell VA Medical Center, Detroit, MI, United States

Axonal loss determines the final disability in patients with peripheral neuropathies. Consequently, axonal loss results in intramuscular fat accumulation. Therefore, measuring muscle fat fraction through Dixon MRI has been a promising biomarker for monitoring disease progression. However, the responsiveness is yet to be improved, particularly in the early phase of the disease. In this study, we developed a deep learning-based method to automate the quantification of individual muscle fat fraction, which mitigates the laborious manual segmentations and enables the use of individual muscle fat fraction as outcome measures to track axonal loss in patients with neuropathies. 
 

Laura Carretero¹, Pablo García-Polo¹, Suryanarayanan Kaushik ², Maggie Fung², Bruno Astuto^3,4, Rutwik Shah^3,4, Pablo F Damasceno^3,4, Valentina Pedoia^3,4, Sharmila Majumdar^3,4, and Mario Padrón^5
1Global Research Organization, GE Healthcare, Madrid, Spain, ²GE Healthcare, Waukesha, WI, United States, ³Department of Radiology and Biomedical Imaging, UCSF, San Francisco, CA, United States, ⁴Center for Digital Health Innovation, UCSF, San Francisco, CA, United States, ⁵Department of Radiology, Clínica Cemtro, Madrid, Spain

This study evaluates the clinical accuracy of a deep learning (DL)-based tool to segment articular cartilage and menisci on 50 knee MRI exams; detect lesions and stage its severity. An experienced MSK radiologist assessed independently the images for the presence of any lesions on the different compartments and checked the accuracy of its segmentation, resulting in no disagreement with the segmentation output in 92.8% of the compartments and correspondence in the detection of lesions in 75.94% of them. The shown results assessed the clinical potential of this tool and present a step forward into structured MSK imaging reports.  

Lee-Ren Yeh¹, Yang Zhang², Jeon-Hor Chen², An-Chi Wang³, JieYu Yang³, Peter Chang², Daniel Chow², and Min-Ying Su^2
1Radiology, E-Da Hospital, Kaohsiung, Taiwan, ²University of California Irvine, Irvine, CA, United States, ³Radiology, Chi-Mei Medical Center, Tainan, Taiwan

This study compared the reading of three radiologists with different level of experience, and also investigated the potential of deep learning to differentiate between benign and malignant vertebral fractures based on T1W and T2W MRI. The results showed that deep learning using ResNet50 achieved a satisfactory diagnostic accuracy of 92%, although inferior to 98% made by a senior MSK radiologist and 96% made by a R4 resident, much higher compared to 66% made by a R1 resident. The inferior performance of ResNet50 might be partly explained by the very limited information when only considering a small bounding box.
 

Can Wu^1,2 and Qi Peng^3
1Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY, United States, ²Philips Healthcare, Andover, MA, United States, ³Department of Radiology, Albert Einstein College of Medicine and Montefiore Medical Center, Bronx, NY, United States

3D T1ρ mapping is a promising technique for quantitative assessment of biochemical changes in knee cartilage. However, synovial fluid, if not suppressed, may compromise T1ρ quantification, particularly in clinical conditions like osteoarthritis where cartilage is usually irregular and synovial fluid is increased. A long-T2-selective inversion approach can be used to suppress the synovial fluid signal at the cost of increased scan time by 50%. This study demonstrated that deep learning can be used to effectively eliminate synovial fluid from T1ρ data acquired without active fluid suppression, potentially leading to improved T1ρ quantification of knee cartilage accuracy without adding scan time.

Xing Lu¹, Yajun Ma¹, Saeed Jerban¹, Hyungseok Jang¹, Yanping Xue¹, Xiaodong Zhang¹, Mei Wu¹, Amilcare Gentili^1,2, Chun-nan Hsu³, Eric Y Chang^1,2, and Jiang Du^1
1Department of Radiology, University of California, San Diego, San Diego, CA, United States, ²Radiology Service, Veterans Affairs San Diego Healthcare System, San Diego, CA, United States, ³Department of Neurosciences, University of California, San Diego, San Diego, CA, United States

In this study, we proposed end-to-end deep learning convolutional neural networks to perform simultaneous segmentation and quantification (MSQ-Net) on the knee without and with physical constraint networks (pcMSQ-Net). Both networks were trained and tested for the feasibility of simultaneous segmentation and quantitative evaluation of multiple knee joint tissues from 3D ultrashort echo time (UTE) magnetic resonance imaging. Results demonstrated the potential of MSQ-Net and pcMSQ-Net for fast and accurate UTE-MRI analysis of the knee, a “whole-organ” approach which is impossible with conventional clinical MRI.

Emma Bahroos¹, Misung Han¹, Cynthia Chin¹, David Shin², Javier Villanueva-Meyer¹, Thomas Link¹, Valentina Pedoia¹, and Sharmila Majumdar^1
1Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States, ²Applications and Workflow, GE Healthcare, Menlo Park, CA, United States

Lower back pain is one of the most common health problems, for which MRI is extensively used. Standard clinical, and fast acquisition images of lumbar spine were acquired for 18 patients.  A (DL)-based image reconstruction was applied to the raw data of the fast images, with 25%, 50%, and 75% noise reduction factors.  Evaluation of fast images with DL algorithm, for image quality, diagnostic capability, and SNR to standard images was conducted by three experienced radiologists. Our results show SNR improvement with higher noise reduction factor without a severe degradation in the ability to discern anatomical structures.

Upasana Upadhyay Bharadwaj¹, Cynthia T Chin¹, Valentina Pedoia¹, and Sharmila Majumdar^1
1Radiology, University of California, San Francisco, San Francisco, CA, United States

Lumbar facet arthropathy is frequently observed along with other degenerative changes of the spine in patients presenting with chronic low back pain. Deep learning has demonstrated unprecedented success in automated assessment of many spine degenerative changes, but heretofore not applied to facet arthropathy. This study presents binary classification of facet arthropathy (normal/mild vs moderate/severe) on T2-weighted axial MRI slices using a two-staged approach: zero-shot facet detection followed by classification. Our model achieves an AUC of 0.916 [0.911, 0.921] with sensitivity and specificity of 97.8% [97.4, 98.3] and 64.1% [63.1, 65.1], respectively and can potentially enhance the clinical workflow.

Jing Cheng¹, Yuanyuan Liu¹, Xin Liu¹, Hairong Zheng¹, Yanjie Zhu¹, and Dong Liang^1
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China

In this work, we propose a novel deep learning-based framework DEMO for fast and robust MR parametric mapping. Different from current deep learning-based methods, DEMO trains the network in an unsupervised way. Specifically, a CS-based loss function is used in DEMO to avoid the necessity of using fully sampled k-space data as the label, and thus make it an unsupervised learning approach. DEMO reconstructs the parametric weighted images and generates the parametric map simultaneously, which enables multi-tasking learning. Experimental results show the promising performance of the proposed DEMO framework in quantitative MR T1ρ mapping.

Bragi Sveinsson^1,2 and Matthew S Rosen^1,2,3
1Martinos Center, Massachusetts General Hospital, Boston, MA, United States, ²Harvard Medical School, Boston, MA, United States, ³Physics, Harvard University, Cambridge, MA, United States

Deep learning networks allow the creation of new images based on a separate set of reference image data. This can be used to synthesize a specific MRI contrast from other image contrasts sharing the same anatomy. A particularly successful approach uses a conditional generative adversarial network with a patch-based discriminator, processing image patches of a fixed size. In this work, we investigate the benefits of using multiple patch sizes to improve image quality.

Jeffrey Dominic¹, Arjun Desai¹, Andrew Schmidt¹, Elka Rubin¹, Garry Gold¹, Brian Hargreaves¹, and Akshay Chaudhari^1
1Stanford University, Stanford, CA, United States

Deep learning (DL)-based approaches have shown promise for automating medical image segmentation with high efficacy. However, current state-of-the-art DL supervised methods require large extents of labeled training images, which are difficult to curate at scale. In this work, we propose a self-supervised training scheme to reduce dependence on labeled data by pretraining networks in an unsupervised manner. We show that our method can improve segmentation performance, especially in the context of very limited data scenarios (only 10-25% scans available) and can achieve or surpass the accuracy of state-of-the-art supervised networks with approximately 50% fewer labeled scans.

Firas Khader¹, Gustav Müller-Franzes¹, Johannes Stegmaier², Martin Pixberg³, Jonas Müller-Hübenthal³, Christiane Kuhl ¹, Sven Nebelung⁴, and Daniel Truhn^1
1Department of Diagnostic and Interventional Radiology, Aachen University Hospital, Aachen, Germany, ²Institute of Imaging and Computer Vision, RWTH Aachen University, Aachen, Germany, ³Praxis im Köln Triangle, Cologne, Germany, ⁴Department of Diagnostic and Interventional Radiology, Düsseldorf University Hospital, Dusseldorf, Germany

In this study we aimed to analyze the capability of neural networks to accurately diagnose the presence of ACL and meniscus tears in our in-house dataset comprised of 3887 manually annotated knee MRI exams. To this end we trained the MRNet architecture on a varying number of training exams that included proton density-weighted axial, sagittal and coronal planes for each knee exam. Additionally, we compared the performance of the architecture when trained on expert vs non-expert annotations. This study demonstrates that while our neural network benefits from a larger dataset, expert annotations do not considerably improve the performance.

Kenneth T Gao^1,2,3, Radhika Tibrewala^1,2, Madeline Hess^1,2, Upasana Bharadwaj^1,2, Gaurav Inamdar^1,2, Cynthia T Chin¹, Valentina Pedoia^1,2, and Sharmila Majumdar^1,2
1Department of Radiology and Biomedical Imaging, University of California, San Francisco, San Francisco, CA, United States, ²Center for Intelligent Imaging, Department of Radiology and Biomedical Imaging, University of California, San Francisco, San Francisco, CA, United States, ³University of California, Berkeley-University of California San Francisco Graduate Program in Bioengineering, San Francisco, CA, United States

Modic changes are common degenerative lesions seen in spinal MRI and are strongly linked to lower back pain. However, detection of Modic changes suffers from poor inter-operator and inter-scanner reliabilities. We present a fully automatic, quantitative model that leverages deep learning and signal-based clustering for mapping Modic changes from clinically acquired MRI. The model achieves an identification rate of 85.7% and substantial agreement with radiologists. More importantly, the mapping technique classifies detected lesions on a voxel-wise basis, allowing for assessment of sensitive, local pathologies.

Madeline Hess¹, Kenneth Gao¹, Radhika Tibrewala¹, Gaurav Inamdar¹, Upasana Bharadwaj¹, Cynthia Chin¹, Valentina Pedoia¹, and Sharmila Majumdar^1
1Center for Intelligent Imaging, University of California, San Francisco, San Francisco, CA, United States

Lumbar spine segmentation serves as an important first step for automated disease classification and monitoring, but manual segmentation is costly and time consuming. We present a deep learning-based pipeline to automatically segment the vertebral bodies, intervertebral discs, and paraspinal muscles in the lumbar spine. We leverage the results of this method to quickly and accurately extract disc height with a mean absolute error of 2.09 mm, muscle CSA with mean absolute errors of less than 1.46 cm2, and muscle centroid position with a mean absolute error of less than 7.23mm.

Yibo Dan¹, Hongyue Tao², Chengxiu Zhang¹, Chenglong Wang¹, Yida Wang¹, Shuang Chen², and Guang Yang^1
1Shanghai Key Laboratory of Magnetic Resonance, East China Normal University, shanghai, China, ²Department of Radiology, Huashan Hospital, Fudan University, shanghai, China

The naked eye can only recognize the morphological changes of cartilage and subchondral bone on conventional MRI, but cannot recognize the subtle changes in their internal structure. The aim is to use radiomics to evaluate the cartilage and subchondral bone changes in patients with chronical ankle joint instability (CAI) on conventional MRI images¹. We built a pipeline to automatically identify CAI from FS-PD images. The pipeline automatically segmented cartilage regions and subchondral bone (5mm) regions, then used SVM based on radiomics features extracted from these regions for classification. In the test dataset, the proposed model achieved an AUC of 0.965.

Nicole Wake^1,2, Stephanie Shamir¹, Beverly Thornhill¹, Nogah Haramati¹, Graeme McKinnon³, Mathias Engstrom⁴, Florian Wiesinger⁴, Michael Carl⁵, Fraser Robb⁶, and Maggie Fung^7
1Department of Radiology, Montefiore Medical Center, Bronx, NY, United States, ²Center for Advanced Imaging Innovation and Research, Department of Radiology, NYU Langone Health, New York, NY, United States, ³GE Healthcare, Waukesha, WI, United States, ⁴GE Healthcare, Munich, Germany, ⁵GE Healthcare, San Diego, CA, United States, ⁶GE Healthcare, Aurora, OH, United States, ⁷GE Healthcare, New York, NY, United States

Patient-specific three-dimensional (3D) printed anatomic models are valuable clinical tools which are generally created from computed tomography (CT).  However, magnetic resonance imaging (MRI) is an attractive alternative, since it offers exquisite soft-tissue characterization and flexible image contrast mechanisms while avoiding the use of ionizing radiation.  The purpose of this study was to evaluate the image quality and assess the feasibility of creating 3D printed models using a 3D Zero echo time (ZTE) MR images which were reconstructed with a deep learning reconstruction method.

Jiaping Hu¹, Zhao Wang², Lijie Zhong¹, Keyan Yu¹, Yanjun Chen¹, Yingjie Mei³, Qi Dou⁴, and Xiaodong Zhang^1
1Department of Medical Imaging, The Third Affiliated Hospital of Southern Medical University, Guangzhou, China, ²College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, China, ³China International Center, Philips Healthcare, Guangzhou, China, ⁴Department of Computer Science & Engineering, The Chinese University of Hong Kong, Hong Kong, China

The presence of a bone marrow lesion is associated with incident and progressive knee osteoarthritis (KOA) and joint replacement. Since the ill-defined boundary and various signal strength, identification of bone marrow lesions (BMLs) requires professional diagnostic ability and is subjective. Therefore, we utilize a model to assess whether there exists BMLs in every subregion and their severity on 3D-dual echo steady state (DESS) images according to MRI Osteoarthritis Knee Score (MOAKS). The initiatory results showed that deep learning framework performed well on discrimination of BMLs with good reproducibility.

Yan Wu¹, Yajun Ma², Jiang Du², and Lei Xing^1
1Stanford University, Palo Alto, CA, United States, ²University of California San Diego, San Diego, CA, United States

While versatile soft tissue contrasts are achievable in MRI, contrast attainable from each scan is predetermined by the imaging protocol. A retrospective tuning of contrast will provide an opportunity to normalize MRI data for radiomics analysis. In this study, we present a new paradigm to obtain a spectrum of contrasts from a single T1-weighted image. Using deep learning, T1 map, proton density map, and B1 map are predicted from every T1-weighted image, and new contrasts can be obtained with the application of Bloch equations. The method has been validated in knee MRI with high accuracy achieved.

Mingrui Yang¹, Ceylan Colak¹, Mercan Aslan¹, Sibaji Gaj¹, Morgan Jones¹, Carl Winalski¹, Naveen Subhas¹, and Xiaojuan Li^1
1Cleveland Clinic, Cleveland, OH, United States

Early diagnosis and effective detection of cartilage degeneration is an important factor for osteoarthritis prevention and treatment, which are still challenging in routine clinical practice, resulting in poor patient treatment and management plans. The purpose of this study is to assess the feasibility of building an automatic femoral cartilage lesion classification pipeline for heterogenous clinical routine MR scans by combining deep learning segmentation and classification models together.