Carotid Artery Localization and Lesion Classification on 3D-MERGE MRI using Neural Network and Object Tracking methods
Li Chen1, Huilin Zhao1,2, Niranjan Balu1, Xihai Zhao3, Rui Li3, Jianrong Xu2, Thomas S Hatsukami1, Jenq-Neng Hwang1, and Chun Yuan1

1University of Washington, Seattle, WA, United States, 2Radiology, Renji Hospital, School of Medicine, Shanghai Jiao Tong University, Shanghai, China, 3Department of Biomedical Engineering, Tsinghua University School of Medicine, Beijing, China


Carotid vessel wall imaging (VWI) with MRI provides additional prognostic value for cerebro/cardiovascular ischemic events, beyond current clinical diagnostic imaging methods. While fast 3D carotid MRI is possible, manual review of the large 3D dataset is time consuming. Automatic identification of artery locations and lesion categories are therefore required for VWI screening protocols. With neural network and object tracking methods, we developed a fully automated analysis tool to find common/internal/external carotid arteries and flag possible high-risk lesion locations. The tool achieved 0.782 Intersection over Union (IoU) for artery localization, and 0.895 sensitivity for high-risk lesion classification.


Atherosclerosis is a major cause of cardio- and cerebrovascular disease mortality and morbidity globally 1–3. Atherosclerotic lesion progression leads to arterial stenosis and/or occlusion, which may cause downstream ischemic events due to atheroembolization or hypoperfusion 4. Several fast 3D carotid black blood MRI sequences are now available 5,6. 3D Motion Sensitized Driven Equilibrium prepared Rapid Gradient Echo (3D-MERGE) 5 black blood MRI can image the carotid vessel wall with high isotropic resolution in a short 2-minute scan thus ensuring patient compliance and diagnostic image quality (becoming available in clinical MRI scanners). However, screening for high-risk lesion is limited by manual image review time. Therefore, fully automated analysis tools on 3D-MERGE are needed.

An automated method to flag locations of atherosclerotic lesions and classify lesion types can help reduce radiologist workload 7–9 and improve screening for atherosclerotic lesions. Identification using deep learning and object tracking may provide a time-efficient solution.

Therefore, the goal of this study is to develop a fully automated analysis tool to identify the locations of the carotid vessel wall from 3D-MERGE images and then flag slices with characteristics of high-risk atherosclerotic lesions.


The workflow for this model is shown in Figure 1.

Study sample

3D-MERGE MRI of 269 subjects were collected as part of the Carotid Atherosclerosis Risk Assessment (CARE-II) study 10 with IRB approval and informed consent.

MR imaging

Scans was performed on a 3T Philips (Best, The Netherlands) MR scanner with: TR/TE=10.35/4.87, acquired resolution =0.7mm isotropic (interpolated to 0.35mm isotropic) using bilateral carotid phased array coils.

Image pre-processing

3D-MERGE images were resampled in the axial direction with isotropic resolution of 0.4mm. Each image slice was filtered using a vessel wall enhanced model 11 emphasizing on vessel wall edges, as wall thickness is a key characteristic in determining lesion types.

Human labeling

An experienced radiologist labeled the dataset using unfiltered axial slices. For each slice, bounding boxes of each carotid artery (minimum encompassing rectangle around each vessel wall region) were drawn and classified into three categories (normal artery, early lesion and advanced lesion), based on modified AHA classification for MRI 12 defined in Table 1. An online learning method 13 was used to minimize tedious human labeling by iteratively correcting prediction results. Slices with poor image quality or where part of artery was out of the field of view were excluded (87 subjects). Ultimately, 151, 17, 14 subjects were used as training, validation and testing sets. An average of 95.9±41.1 slices were labeled per subject.

Artery localization

The Yolo V2 neural network 14 model was trained to identify bounding boxes in each slice. Predicted bounding boxes along the slices were reconstructed into tracklets (trajectory fragments) and refined using a tracking by detection method 15,16, which utilizes the neighboring information of bounding boxes to maintain the robustness and continuity of predicted artery wall location.

Lesion classification

The vessel walls in consecutive five slices were cropped using the location of refined bounding boxes and the box in the center slice was classified using a convolutional neural network (8 layers of convolutional layers followed by 3 layers of fully connected layers and softmax layer).


Location sensitivity is 0.985, precision is 0.998, mean Intersection over Union (IoU) is 0.782.

The percentages of artery categories labeled by the radiologist are 27.4% for normal artery, 27.5% for early lesion, and 45.1% for advanced lesion. The classification accuracy of three categories is 0.746, and the weighted kappa for agreement between the radiologist and the tool is 0.77 (0.76-0.78). For advanced lesions, the accuracy is 0.895, the sensitivity is 0.892, and the precision is 0.877. An example showing prediction on one slice is shown in Figure 2. Confusion matrix is shown in Table 2.

Slices with lesions can be flagged onto artery centerlines acquired from centers of bounding boxes in 3D space. 3D Visualization of carotid artery centerline and lesions is shown in Figure 3.


The localization and classification models were shown to accurately identify artery locations and detect lesion types with high performance. Although the training process took days, the prediction was completed within a minute. 3D visualization that flags the area of high-risk lesions will substantially reduce review burden for radiologists in clinical practice.

The most common mistakes for the classification model occur mostly near the common carotid bifurcation where complex geometry of the carotid bifurcation and bulb is a challenge.


The proposed fully automated carotid 3D-MERGE MRI image analysis tool is capable of artery localization and lesion classification with a high level of accuracy and agreement compared to findings from expert reviewers. This tool can potentially provide a rapid technique for rapid high-risk lesion screening in large populations.


Li Chen and Huilin Zhao contributed equally to this work. This research is supported by grants from Natural Science Foundation of China (81271536, 61271132 and 81361120402), National Institute of Health (R01-HL103609) and Philips Healthcare. We gratefully acknowledge CARE-II investigators and support of NVIDIA Corporation for donating the Titan Xp GPU.


1. Caplan LR, Gorelick PB, Hier DB. Race, sex and occlusive cerebrovascular disease: a review. Stroke. 1986;17:648-655. doi:10.1161/01.STR.17.4.648.

2. Murray CJL, Vos T, Lozano R, et al. Disability-adjusted life years (DALYs) for 291 diseases and injuries in 21 regions, 1990-2010: a systematic analysis for the Global Burden of Disease Study 2010. Lancet. 2012;380(9859):2197-2223. doi:10.1016/S0140-6736(12)61689-4.

3. Lozano R, Naghavi M, Foreman K, et al. Global and regional mortality from 235 causes of death for 20 age groups in 1990 and 2010: a systematic analysis for the Global Burden of Disease Study 2010. Lancet. 2012;380(9859):2095-2128. doi:10.1016/S0140-6736(12)61728-0.

4. Zhang X, Zhou Y, Zhang S, Ding W, Lou M. Cerebral blood flow evaluation of intensive rosuvastatin therapy in stroke/transient ischemic attack patients with intracranial arterial atherosclerotic stenosis study: Rationale and design. Brain Behav. 2017;7(6):3-7. doi:10.1002/brb3.689.

5. Balu N, Yarnykh VL, Chu B, Wang J, Hatsukami T, Yuan C. Carotid plaque assessment using fast 3D isotropic resolution black-blood MRI. Magn Reson Med. 2011;65(3):627-637. doi:10.1002/mrm.22642.

6. Xie G, Zhang N, Xie Y, et al. DANTE-prepared three-dimensional FLASH: A fast isotropic-resolution MR approach to morphological evaluation of the peripheral arterial wall at 3 Tesla. J Magn Reson Imaging. 2016;43(2):343-351. doi:10.1002/jmri.24986.

7. Saam T, Hetterich H, Hoffmann V, et al. Meta-Analysis and Systematic Review of the Predictive Value of Carotid Plaque Hemorrhage on Cerebrovascular Events by Magnetic Resonance Imaging. J Am Coll Cardiol. 2013;62(12):1081-1091. doi:10.1016/j.jacc.2013.06.015.

8. Gupta A, Baradaran H, Schweitzer AD, et al. Carotid Plaque MRI and Stroke Risk. Stroke. 2013;44(11):3071-3077. doi:10.1161/STROKEAHA.113.002551.

9. Sun J, Zhao X-Q, Balu N, et al. Carotid Plaque Lipid Content and Fibrous Cap Status Predict Systemic CV Outcomes: The MRI Substudy in AIM-HIGH. JACC Cardiovasc Imaging. 2017;10(3):241-249. doi:10.1016/J.JCMG.2016.06.017.

10. Zhao X, Li R, Hippe DS, Hatsukami TS, Yuan C. Chinese Atherosclerosis Risk Evaluation (CARE II) study: a novel cross-sectional, multicentre study of the prevalence of high-risk atherosclerotic carotid plaque in Chinese patients with ischaemic cerebrovascular events—design and rationale. Bmj. 2017;2(1):15-20. doi:10.1136/svn-2016-000053.

11. Chen L, Sun J, Zhang W, et al. Automatic Segmentation of Carotid Vessel Wall Using Convolutional Neural Network. ISMRM2018. 2018.

12. Cai JM, Hatsukami TS, Ferguson MS, Small R, Polissar NL, Yuan C. Classification of human carotid atherosclerotic lesions with in vivo multicontrast magnetic resonance imaging. Circulation. 2002;106(11):1368-1373. doi:10.1161/01.CIR.0000028591.44554.F9.

13. Chen L, Zhao H, Balu N, et al. Carotid Artery Localization and Lesion Detection on 3D-MERGE MRI through Online Learning. SMRA 2018. 2018.

14. Redmon J, Farhadi A. YOLO9000: Better, faster, stronger. Proc - 30th IEEE Conf Comput Vis Pattern Recognition, CVPR 2017. 2017;2017-Janua:6517-6525. doi:10.1109/CVPR.2017.690.

15. Andriluka M, Roth S, Schiele B. People-tracking-by-detection and people-detection-by-tracking. In: 2008 IEEE Conference on Computer Vision and Pattern Recognition. IEEE; 2008:1-8. doi:10.1109/CVPR.2008.4587583.

16. Zheng Tang, Gaoang Wang, Hao Xiao, Aotian Zheng J-NH. Single-camera and inter-camera vehicle tracking and 3D speed estimation based on fusion of visual and semantic features. 2018:108-115. doi:10.1109/CVPRW.2018.00022.


Figure 1 Workflow for fully automated artery localization and lesion classification model in this study

Table 1 Carotid vessel wall categories with definition and example image slices

Figure 2 One slice of carotid artery from testing set with green bounding boxes showing prediction locations and artery type (text right to the bounding boxes), classification confidence follows the artery type. Reference locations by the radiologist shown in blue, yellow and red bounding boxes representing normal arteries, early lesions and advanced lesions, respectively.

Table 2 Heat-map of confusion matrix for prediction results of all slices in test sets. Rows are Reference and columns are predictions. Color for cells indicates level of counts (incremental from red to green)

Figure 3 (a) 3D visualization of carotid arteries (blue for normal artery) with predicted probable lesion locations marked by red or yellow colors for a subject in test set. Three slice views of a certain lesion location pointed by the red arrow are displayed in (b, c, d). Red circles show estimated radius of vessel wall. Green and blue (current selected) lines show centerline of arteries.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)