Abstract
Quantitative analysis of vessel wall structures by automated vessel wall segmentation provides useful imaging biomarkers in evaluating atherosclerotic lesions and plaque progression time-efficiently. To quantify vessel wall features, drawing lumen and outer wall contours of the artery of interest is required. To alleviate manual labor in contour drawing, some computer-assisted tools exist, but manual preprocessing steps, such as region of interest identification and boundary initialization are needed. In addition, the prior knowledge of the ring shape of vessel wall is not taken into consideration in designing the segmentation method. In this work, trained on manual vessel wall contours, a fully automated artery localization and vessel wall segmentation system is proposed. A tracklet refinement algorithm is used to robustly identify the centerlines of arteries of interest from a neural network localization architecture. Image patches are extracted from the centerlines and segmented in a polar coordinate system to use 3D information and to overcome problems such as contour discontinuity and interference from neighboring vessels. From a carotid artery dataset with 116 subjects (3406 slices) and a popliteal artery dataset with 5 subjects (289 slices), the proposed system is shown to robustly identify the artery of interest and segment the vessel wall. The proposed system demonstrates better performance on the carotid dataset with a Dice similarity coefficient of 0.824, compared with traditional vessel wall segmentation methods, Dice of 0.576, and traditional convolutional neural network approaches, Dice of 0.747. This vessel wall segmentation system will facilitate research on atherosclerosis and assist radiologists in image review.
Introduction
Atherosclerosis—the buildup of plaque in arterial walls—is a leading cause of death globally. To monitor the disease, clinicians use "black-blood" MRI to visualize the vessel wall. However, quantifying the thickness and area of these walls usually requires a radiologist to manually trace the inner (lumen) and outer boundaries across hundreds of image slices. This manual process is tedious, prone to human error, and slow. While some computer-assisted tools exist, they typically require a human to first identify the "Region of Interest" (ROI) or provide a starting point. Furthermore, standard deep learning models treat these images as flat grids (Cartesian coordinates), which fails to exploit the natural, ring-like geometry of blood vessels.
Methods
The proposed system utilizes a two-stage workflow: Localization and Segmentation.
Lumen Center Localization
The system first identifies arteries in individual slices using a YOLO V2 detector. Because detection can be noisy—potentially missing an artery in one slice or identifying a false positive in another—the researchers implemented a Tracklet Refinement algorithm. This algorithm treats the stack of MRI slices like a video, linking detections across slices to ensure a smooth, continuous centerline and filtering out noise detections that do not form a logical 3D path.
Polar Segmentation CNN
Once the artery center is localized, the image patch is converted from Cartesian $(x, y)$ to Polar $(\rho, \theta)$ coordinates. In polar space, the circular vessel wall is "unwrapped" into a rectangular strip. This geometric transformation makes it significantly easier for a convolutional neural network (CNN) to learn boundaries, as the lumen and outer wall appear as nearly horizontal lines. The system uses a CNN architecture similar to a convolutional auto-encoder to predict wall boundaries, followed by an Active Contour Model (Snake) to smooth the predicted boundaries before converting them back to the standard Cartesian view.
Datasets
The study utilized three major datasets for training and validation:
- CARE-II: 954 subjects with carotid artery images (T1-weighted).
- KOWA: 203 asymptomatic subjects (T1-weighted).
- OAI (Osteoarthritis Initiative): 30 subjects used to test the model on popliteal (behind the knee) arteries.
In total, the model was trained on over 32,000 manually labeled image slices, representing a wide spectrum of plaque types and image qualities.
Key Results and Insights
The proposed system demonstrated a clear leap in performance over existing methods:
- Accuracy: On the carotid dataset, the system achieved a Dice Similarity Coefficient (DSC) of 0.824. For context, traditional B-spline snake methods scored 0.576, and standard CNNs operating in Cartesian space scored 0.747.
- Robustness: The system maintained high performance (0.821 DSC) even when transferred to a different vascular bed (popliteal arteries) with minimal fine-tuning.
- Correlation: Automated measurements showed a high correlation (0.885) with manual measurements performed by experienced human reviewers.
By changing the coordinate system, the authors effectively bake "prior knowledge" of the vessel's ring shape into the data itself. Furthermore, by using tracklets, the model recognizes that an artery is a continuous tube, allowing the system to "fill in the blanks" if a specific slice has poor contrast.
Limitations and Future Work
The primary drawback is computational overhead; the polar conversion and restoration process currently takes roughly 40.3 seconds per slice. Future versions could utilize GPU acceleration to mitigate this. Additionally, while tested on relatively straight segments of the carotid and popliteal arteries, future work will need to adapt the tracklet approach to handle more tortuous vessels, such as intracranial arteries.
Summary updated