Yuya Saito¹, Peter A. Wijeratne², Koji Kamagata¹, Christina Andica¹, Wataru Uchida^1,3, Toshiaki Akashi¹, Akihiko Wada¹, Masaaki Hori⁴, and Shigeki Aoki^1
1Department of Radiology, Juntendo University Graduate School of Medicine, Tokyo, Japan, ²Centre for Medical Image Computing, Department of Computer Science, University College London, London, United Kingdom, ³Department of Radiological Sciences, Graduate School of Human Health Sciences, Tokyo Metropolitan University, Tokyo, Japan, ⁴Department of Radiology, Toho University Omori Medical Center, Tokyo, Japan

Corticobasal syndrome (CBS) and progressive supranuclear palsy (PSP) are classic clinical syndromes derived from 4R tau pathology. Differential clinical diagnosis remains difficult due to neurodegenerative overlap. Most previous studies have assessed white-matter (WM) degeneration using cross-sectional data. This study applied Subtype & Stage Inference (SuStaIn), a novel unsupervised machine-learning technique for regional WM fractional anisotropy based on cross-sectional brain diffusion MRI to identify differences in temporal progression patterns of WM degeneration between CBS and PSP. Results suggested the utility of SuStaIn for identifying temporal WM degeneration patterns in and classifying patients with CBS and PSP.

Wen Li¹, Ge Ren¹, Tian Li¹, Haonan Xiao¹, Francis Kar-ho Lee², Kwok-hung Au², and Jing Cai^1
1Department of Health Technology and Informatics, The Hong Kong Polytechnic University, Hong Kong, China, ²Department of Clinical Oncology, Queen Elizabeth Hospital, Hong Kong, China

To reduce the usage of gadolinium-based contrast agents (GBCAs), we proposed a deep learning based multi-inputs network (CE-Net) for contrast enhanced T1-weighted MR image synthesis based on pre-contrast T1-weighted and T2-weighted images in nasopharyngeal carcinoma (NPC) cases. When compared with multi-channel input methods, the proposed CE-Net has the ability to extract information from each input modalities separately. Supervision and multi-scale strategies are also applied in the proposed network. Quantitative and qualitative results show that our proposed CE-Net could achieve better performance when compared with the newly proposed Hi-Net and its extensions.

Esra Zihni¹, Bryony McGarry^1,2, and John D. Kelleher^1,3
1PRECISE4Q, Predictive Modelling in Stroke, Information Communications and Entertainment Institute, Technological University Dublin, Dublin, Ireland, ²School of Psychological Science, University of Bristol, Bristol, United Kingdom, ³ADAPT Research Centre, Technological University Dublin, Dublin, Ireland

Applying deep learning models to MRI scans of acute stroke patients to extract features that are indicative of short-term outcome could assist a clinician’s treatment decisions. Deep learning models are usually accurate but are not easily interpretable. Here, we trained a convolutional neural network on ADC maps from hyperacute ischaemic stroke patients for prediction of short-term functional outcome and used an interpretability technique to highlight regions in the ADC maps that were most important in the prediction of a bad outcome. Although highly accurate, the model’s predictions were not based on aspects of the ADC maps related to stroke pathophysiology.

Yuval Buchsweiler^1,2, Orna Aizenstein^3,4, Felix Bokstein^3,5,6, Idan Bressler^1,2, Netanell Avisdris^2,7, Deborah T. Blumenthal^3,5, Dror Limon^3,8, Dafna Ben Bashat^2,3,6, and Moran Artzi^2,3,6
1The Iby and Aladar Fleischman Faculty of Engineering, Tel Aviv University, Tel Aviv, Israel, ²Sagol Brain Institute, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, ³Sackler Faculty of Medicine, Tel Aviv University, Tel Aviv, Israel, ⁴Division of Radiology, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, ⁵Neuro-Oncology Service, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel, ⁶Sagol School of Neuroscience, Tel Aviv University, Tel Aviv, Israel, ⁷School of computer science and engineering, Hebrew University of Jerusalem, Israel, Jerusalem, Israel, ⁸Division of Oncology, Tel Aviv Sourasky Medical Center, Tel Aviv, Israel

Brain tumor segmentation is highly important for clinical management. We propose HUMBLe, a hierarchical 3D U-Net for MRI Brain Lesion segmentation architecture. HUMBLe breaks down the segmentation into its separate classes: enhancing tumor, edema, and necrotic classes, and uses a classifier to merge the different segmentation results into a final segmentation mask. Evaluation was performed on multi-parametric longitudinal local dataset, of patients with Glioblastoma. Segmentation results obtained by HUMBLe on our cohort improved DICE scores by 7%-16% for the different tumor components, compared to segmentation performed using 3D U-Net based architecture trained on BraTS2019 and our cohort.

SeyyedKazem HashemizadehKolowri^1,2, Rong-Rong Chen², Ganesh Adluru^1,3, and Edward V. R. DiBella^1,2,3
1Radiology and Imaging Science, University of Utah, SALT LAKE CITY, UT, United States, ²Electrical and Computer Engineering, University of Utah, SALT LAKE CITY, UT, United States, ³Biomedical Engineering, University of Utah, SALT LAKE CITY, UT, United States

Advanced diffusion models enable characterization of tissue microstructure with higher specificity than conventional DTI. While beyond DTI diffusion imaging have been found valuable in many studies,  their clinical availability have been hampered mainly due to their long scan times. Furthermore, each diffusion model can only extract a few relevant microstructural features. Therefore, using multiple models helps to better understand the brain microstructure, which requires multiple expensive model-fitting.  In this study, we use different deep learning approaches to jointly estimate multiple advanced diffusion maps from highly undersampled q-space data, which can reduce both the scan and processing times significantly.

Spencer D. Christiansen^1,2, Junmin Liu², Maria Bres Bullrich³, Manas Sharma⁴, Sachin K. Pandey⁴, Mel Boulton³, Luciano A. Sposato³, and Maria Drangova^1,2
1Medical Biophyics, Western University, London, ON, Canada, ²Robarts Research Institute, London, ON, Canada, ³Clinical Neurological Sciences, London Health Sciences Centre, London, ON, Canada, ⁴Department of Medical Imaging, Western University, London, ON, Canada

Thrombus red blood cell (RBC) content has been associated with ischemic stroke etiology and responsiveness to recanalization therapies, yet currently can only be analyzed through retrospective histological analysis. We evaluated the ability of a convolutional neural network for predicting thrombus RBC content using multiparametric (R₂^*, QSM, late echo GRE) MR image slices of retrieved stroke thrombi ex vivo. The network predicted thrombus RBC content with an accuracy and mean absolute error of up to 71 and 8%, respectively, when data augmentation was applied. This technique holds potential for in vivo RBC content prediction and improving acute stroke care.

Francesco La Rosa^1,2,3, Germán Barquero^1,2,3, Omar Al-Louzi⁴, Bénédicte Maréchal^1,3,5, Tobias Kober^1,3,5, Jean-Philippe Thiran^1,3, Pascal Sati^4,6, Daniel S. Reich⁴, Pietro Maggi^7,8, Martina Absinta^4,9, Meritxell Bach Cuadra^1,2,3, and Cristina Granziera^10,11,12
1Signal Processing Laboratory (LTS5), École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland, ²Medical Image Analysis Laboratory, Center for Biomedical Imaging (CIBM), University of Lausanne, Lausanne, Switzerland, ³Radiology Department, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ⁴Translational Neuroradiology Section, National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, MD, United States, ⁵Advanced Clinical Imaging Technology, Siemens Healthcare AG, Lausanne, Switzerland, ⁶Department of Neurology, Cedars-Sinai Medical Center, Los Angeles, CA, United States, ⁷Department of Neurology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ⁸Department of Neurology, Cliniques Universitaires Saint-Luc, Université Catholique de Louvain, Brussels, Belgium, ⁹Department of Neurology, Johns Hopkins University, Baltimore, MD, United States, ¹⁰Neurologic Clinic and Policlinic, Departments of Medicine, Clinical Research and Biomedical Engineering, University Hospital Basel and University of Basel, Basel, Switzerland, ¹¹Translational Imaging in Neurology (ThINk) Basel, Department of Biomedical Engineering, University Hospital Basel and University of Basel, Basel, Switzerland, ¹²Research Center for Clinical Neuroimmunology and Neuroscience (RC2NB) Basel, University Hospital Basel and University of Basel, Basel, Switzerland

Paramagnetic rim lesions (PRL) in multiple sclerosis are chronic inflammatory lesions depicted in susceptibility-based MRI where high PRL lesion burden has been associated with a more aggressive disease course. As their visual detection is subjective and time-consuming, a convolutional neural network (RimNet) has recently been developed and applied to 3T susceptibility contrast MR images. In this work, we evaluate a unimodal RimNet architecture based on either the MP2RAGE uniform contrast or the concurrently obtained T1 map at both 3T and 7T. Results show that prediction improves considerably at 7T, suggesting that 7T MP2RAGE might be helpful for automatically identifying PRL.

Seongjin Choi¹ and Daniel M Harrison^1,2
1Neurology, University of Maryland School of Medicine, Baltimore, MD, United States, ²Neurology, Johns Hopkins University School of Medicine, Baltimore, MD, United States

Three machine-learning algorithms were evaluated in the multiple sclerosis phenotype classification of a relatively small cohort. High accuracy of multiple-sclerosis phenotype classification was achievable by applying tree-based ensemble methods to integrated 7T MRI and clinical data features. Feature integration did not guarantee performance improvements in all machine learning algorithms evaluated. Features considered important may vary depending on the classification algorithm used.

Zhijia Jin¹, Ruiqi Yu², Chenglong Wang², Naying He¹, Yan Li¹, Zenghui Cheng¹, Yida Wang², Mark E. Haacke^1,3,4,5, Guang Yang², and Fuhua Yan^1
1Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, ²Shanghai Key Laboratory of Magnetic Resonance, East China Normal University, Shanghai, China, ³Magnetic Resonance Innovations, Inc., Bingham Farms, MI, United States, ⁴Department of Radiology, Wayne State University, Detroit, MI, United States, ⁵Department of Biomedical Engineering, Wayne State University, Detroit, MI, United States

A total of 104 Parkinson’s disease (PD) patients and 269 age- and sex-matched healthy controls (HCs) were scanned using 3D multi-echo gradient echo MTC sequence. In this work, a data-driven clustering approach based on iron deposition in the substantia nigra (SN) measured by quantitative susceptibility mapping (QSM) was performed to classify the PD patients into different subtypes. The clinical assessments were compared between subtype groups. Two subtypes were found by using this clustering approach. Furthermore, there are significant differences (p-values < 0.05) on MDS-UPDRS scores between these two subtype groups.

Igor Varga^1,2, Hae Sol Moon³, Adam Conrad⁴, Matthew Holbrook³, Andrei R Niculescu⁵, Abinaya Lakshmanan³, Robert J Anderson⁵, Christina L Williams⁶, Cristian T Badea^3,5, Po-Wah So², and Alexandra Badea^3,5,7,8
1Department of Cybernetics, Czech Technical University, Prague, Czech Republic, ²Department of Neuroimaging, King's College London, London, United Kingdom, ³Biomedical Engineering, Duke University, Durham, NC, United States, ⁴Georgia Institute of Technology, Atlanta, GA, United States, ⁵Radiology, Duke University Medical Center, Durham, NC, United States, ⁶Psychology and Neuroscience, Duke University, Durham, NC, United States, ⁷Neurology, Duke University, Durham, NC, United States, ⁸Brain Imaging and Analysis, Duke University Medical School, Durham, NC, United States

Whole mouse brain segmentation is an essential prerequisite for multiple quantitative image analysis tasks and pipelines. In this work, we compare three methods for whole brain segmentation (skull stripping) relying on active contours, graph cuts and convolutional neural networks. We applied these methods on mouse brain manganese enhanced MR images acquired at 100 micrometre isotropic resolution, in vivo. All three methods achieved Dice coefficients larger than 94%, but convolutional neural networks achieved a small but significant improvement (0.97±0.01) over our active contours implementation (0.94±0.05) and the difference approached significance relative to graph cuts (0.96±0.01). 

Eve L Kazarian¹, Mariam S Aboian¹, and John D Port^2
1Department of Radiology and Biomedical Imaging, Yale University, New Haven, CT, United States, ²Department of Radiology, Mayo Clinic, Rochester, MN, United States

A novel tree-based machine learning model was applied to head MR imaging scans from multiple sclerosis (MS) patients in order to detect cortical gray matter lesions. Cortical lesions are strong indicators of MS disability yet are not easily visualized in clinical practice. The model was useful for accurately detecting cortical lesion presence (AUC of 0.78 ± 0.02).

Jiaming Wu¹, Hugh G. Pemberton^1,2, Ivar Kommers³, Domenique M.J. Müller³, Sjoerd B. Vos^1,2, Ferran Prados^1,2, Yipeng Hu¹, Pierre A. Robe⁴, Hilko Ardon⁵, Lorenzo Bello⁶, Marco Rossi⁶, Tommaso Sciortino⁶, Marco Conti Nibali ⁶, Mitchel S. Berger⁷, Shawn L. Hervey-Jumper⁷, Wim Bouwknegt⁸, Wimar A. Van den Brink⁹, Julia Furtner¹⁰, Seunggu J. Han¹¹, Albert J. S. Idema¹², Barbara Kiesel¹³, Georg Widhalm¹³, Alfred Kloet¹⁴, Michiel Wagemakers¹⁵, Aeilko H. Zwinderman¹⁶, Sandro M. Krieg^17,18, Emmanuel Mandonnet¹⁹, Philip de Witt Hamer³, Roelant S. Eijgelaar³, and Frederik Barkhof^1,20
1Centre for Medical Image Computing (CMIC), University College London, London, United Kingdom, ²Neuroradiological Academic Unit, UCL Queen Square Institute of Neurology, University College London, London, United Kingdom, ³Neurosurgical Center Amsterdam, Amsterdam UMC, Vrije Universiteit, Amsterdam, Netherlands, ⁴Department of Neurology & Neurosurgery, University Medical Center Utrecht, Utrecht, Netherlands, ⁵Department of Neurosurgery, St. Elisabeth Hospital, Tilburg, Netherlands, ⁶Neurosurgical Oncology Unit, Departments of Oncology and Hemato-Oncology, Università degli Studi di Milano, Humanitas Research Hospital, IRCCS, Milan, Italy, ⁷Department of Neurological Surgery, University of California, San Francisco, San Francisco, CA, United States, ⁸Department of Neurosurgery, Medical Center Slotervaart, Amsterdam, Netherlands, ⁹Department of Neurosurgery, Isala Hospital, Zwolle, Netherlands, ¹⁰Department of Biomedical Imaging and Image-Guided Therapy, Medical University Vienna, Vienna, Austria, ¹¹Department of Neurological Surgery, Oregon Health and Science University, Portland, OR, United States, ¹²Department of Neurosurgery, Northwest Clinics, Alkmaar, Netherlands, ¹³Department of Neurosurgery, Medical University Vienna, Vienna, Austria, ¹⁴Department of Neurosurgery, Medical Center Haaglanden, The Hague, Netherlands, ¹⁵Department of Neurosurgery, University of Groningen, University Medical Center Groningen, Groningen, Netherlands, ¹⁶Department of Clinical Epidemiology and Biostatistics, Academic Medical Center, Amsterdam, Netherlands, ¹⁷TUM-Neuroimaging Center, Klinikum rechts der Isar, Technische Universität München, Munich, Germany, ¹⁸Department of Neurosurgery, Klinikum rechts der Isar, Technische Universität München, Munich, Germany, ¹⁹Department of Neurosurgery, Lariboisière Hospital, APHP, Paris, France, ²⁰Radiology & Nuclear Medicine, VU University Medical Center, Amsterdam, Netherlands

Accurate segmentation and morphological assessment of glioma can guide treatment and support follow-up. The Brain Tumour Segmentation (BraTS) challenge has been instrumental in promoting research and comparing various automated segmentation algorithms. However, models in the challenge are trained and measured on a strictly curated and high-quality dataset, which is not representative of clinically acquired MRI data. Therefore, we have tested the generalisability of three network architectures from two of the top performing BraTS challenge models. We show the utility of these models in the presence of missing sequences and different scanners in multi-centre hospital data of varying quality.

Yadi Li¹, Ping Cheng¹, Pu-Yeh Wu², Wenwen Shen³, Huifen Liu⁴, Jianbing Zhang⁴, Haibo Dong¹, and Wenhua Zhou ^3
1Department of Radiology, Ningbo Medical Treatment Center Lihuili Hospital, Ningbo University, Ningbo, China, ²GE Healthcare,Beijing,China, Beijing, China, ³Laboratory of Behavioral Neuroscience, Ningbo Addiction Research and Treatment Center, Ningbo, China, Ningbo, China, ⁴Ningbo Addiction Research and Treatment Center, Ningbo, China, Ningbo, China

This is a pilot study of the weighted white matter (WM) network in MA-dependent patients. By combining DTI-based probabilistic tractography and graph theory, the WM networks of MA-dependent patients presented small-worldness, and these networks tend to be random networks. The network metrics, that presented inter-group differences were used to construct a support vector machine, that achieved an excellent performance in discriminating MA-dependent patients from normal controls. Overall, the current study demonstrated that MA dependence is associated with abnormal network metrics, and these metrics can be promising features to train a  classifier which need further verification with a larger sample size.
 

Mariana Bento^1,2, Justin Park^2,3, and Richard Frayne^1
1Radiology and Clinical Neuroscience, Hotchkiss Brain Institute, University of Calgary, Calgary, AB, Canada, ²Calgary Image Processing and Analysis Centre, Foothills Medical Centre, Calgary, AB, Canada, ³Mechanical Engineering, University of Calgary, Calgary, AB, Canada

Multi-centre heterogeneous imaging datasets are frequently required to develop image-based computer-aided diagnosis and treatment monitoring tools. However, these datasets may present large underlying variability, potentially impacting the performance of the developed tools. Here, as a proof-of-concept, we propose a machine learning method to study scan variability related to the scanner vendor and magnetic field strength in brain MR images from two cohorts of healthy subjects. Our model has high accuracy rates (>92%), confirming the presence of scan variability in heterogeneous, multi-centre datasets. This model may be further incorporated into automated diagnostic tools, potentially allowing more reliable and robust results.

Ori Ben Zvi^1,2, Netanell Avisdris^1,3, Bossmat Yehuda^1,2, Daphna Link Sourani¹, Leo Joskowicz³, Elka Miller⁴, Liat Ben Sira^2,5, and Dafna Ben Bashat^1,2,5
1Sagol Brain Institute, Tel Aviv Sourasky Medical Center; Israel, Tel Aviv, Israel, ²Sagol School of Neuroscience, Tel Aviv University; Israel, Tel Aviv, Israel, ³School of Computer Science and Engineering, The Hebrew University of Jerusalem, Jerusalem, Israel, Jerusalem, Israel, ⁴Medical Imaging, Children’s Hospital of Eastern Ontario, University of Ottawa, Ottawa, Canada, Ottawa, ON, Canada, ⁵Sackler Faculty of Medicine, Tel Aviv University; Israel, Tel Aviv, Israel

Segmentation of the fetal brain into its components is important for quantitative assessment of fetal development. This study proposes a fully automatic method based on deep learning for fetal brain segmentation into six components, including a separation of right and left hemispheres. The method’s performance demonstrated high Dice scores for all brain components and robustness to different contrasts, scan resolutions, gestational age and fetal brain pathologies. Preliminary results demonstrated significant larger ventricle’s volumes and asymmetry in fetuses with ventriculomegaly compared to normal fetuses.  The method is suggested to improve fetal assessment and assist radiologists in routine clinical practice.

Junchuan Peng¹, Yashi Nan¹, Li Zhao², Huanhui Xiao¹, and Silun Wang^1
1YIWEI Medical Technology Co., Ltd, Shenzhen, China, ²College of Biomedical Engineering & Instrument Science, Zhejiang University, Hangzhou, China

Morphological changes in neurodegenerative diseases can be detected with structural MR images, but it requires detailed brain parcellation. Therefore, a task-aware V-Net was proposed to segment the brain into 40 regions. Task-aware features were achieved by three cascading branches, including brain and non-brain regions, 25 regions with the bilateral regions grouped, and 40 regions, respectively. The proposed model was developed on 8938 subjects and validated using additional196 subjects. The proposed method outperformed the typical 3D U-Net and V-Net, and achieved state-of-the-art results on both datasets with a mean Dice score of 0.886 ± 0.029 and 0.874 ± 0.013, respectively.

Pedro Henrique Rodrigues da Silva¹, Kaio Felippe Secchinato¹, Julia Palaretti¹, and Renata Ferranti Leoni^1
1USP, Ribeirão Preto, Brazil

A challenging issue regarding the early diagnosis of the Alzheimer's disease (AD) is the selection of biomarkers. In this study, we aimed to classify cognitively normal elderly regarding the possibility to develop AD based on brain atrophy and neuropsychological scores, and using supervised machine learning algorithms. Our results suggest Naïve-Bayes (NB) classifiers with left postcentral and left middle temporal cortical thickness or right lateral ventricle, right inferior parietal and Corpus Callosum (CC) Mid Posterior volumes can be useful to identify in the early stage the subjects with higher risks to develop AD.

Yasutaka Fushimi¹, Satoshi Nakajima¹, Akihiko Sakata¹, Takuya Hinoda¹, Sonoko Oshima¹, Sayo Otani¹, Krishna Pandu Wicaksono¹, Hiroshi Tagawa¹, Yang Wang¹, Masahiro Nambu², Rimika Imai², Koji Fujimoto³, Hitomi Numamoto⁴, Kanae Miyake⁴, Tsuneo Saga⁴, and Yuji Nakamoto^1
1Kyoto University Hospital, Kyoto, Japan, ²MRI Systems Division, Canon Medical Systems Corporation, Otawara, Japan, ³3. Department of Real World Data Research and Development, Kyoto University Graduate School of Medicine, Kyoto, Japan, ⁴Department of Advanced Medical Imaging Research, Kyoto University Graduate School of Medicine, Kyoto, Japan

Twelve-fold accelerated submillimeter 3D-T2 weighted imaging with DLR and imaging quality and sharpness of cranial nerves were examined.

Yu Zhang¹, Tongtong Li¹, Xiuwei Fu², Xianchang Zhang³, Yuan Luo⁴, and Hongyan Ni^5
1First Central Clinical College, Tianjin Medical University, Tianjin, China, ²Tianjin Medical University General Hospital, Tianjin, China, ³MR Collaboration, Siemens Healthcare Ltd., Beijing, China, ⁴Department of Radiology, China-Japan Friendship Hospital, Beijing, China, ⁵Department of Radiology, Tianjin First Central Hospital, Tianjin, China

The early diagnoses of Alzheimer’s disease (AD) and amnestic mild cognitive impairment (aMCI) are crucial. This study aimed to acquire new imaging markers to assess the severity of AD and provide early diagnoses using machine learning algorithm. Diffusion kurtosis imaging (DKI) parameters were acquired on 58 AD patients, 64 aMCI patients, and 60 healthy volunteers. It’s found that radial diffusivity value of right uncinate fasciculus was the most important feature for assessing severity. The random forest classifier showed the highest diagnostic efficacy for AD. The RF classifier can provide an early diagnosis of disease based on the quantitative DKI features.

Chao Chai¹, Pengchong Qiao², Bin Zhao², Huiying Wang³, Guohua Liu², Hong Wu², E.Mark Haacke⁴, Wen Shen¹, Xinchen Ye⁵, Zhiyang Liu², and Shuang Xia^1
1Department of Radiology, Tianjin First Central Hospital, Tianjin Medical Imaging Institute, School of Medicine, Nankai University, Tianjin, China, ²Tianjin Key Laboratory of Optoelectronic Sensor and Sensing Network Technology, College of Electronic Information and Optical Engineering, Nankai University, Tianjin, China, ³Department of Radiology, Tianjin First Central Hospital, School of Medicine, Nankai University, Tianjin, China, ⁴Department of Radiology, Wayne State University, Detroit, MI, United States, ⁵DUT-RU International School of Information Science and Engineering, Dalian University of Technology, Dalian, China

This study focused on developing an automatic gray matter nuclei segmentation method. A 3D convolutional neural network based method was proposed, which adopted patches with different resolutions as input for segmentation. Experimental results showed much higher segmentation accuracy over the atlas-based method and other deep-learning-based methods in terms of both the similarity and the surface distance metrics. The segmentation results of the proposed method is also evaluated in terms of measurement accuracy, where the proposed method achieves the highest consistency with the manual delineations.