3D carotid artery segmentation using shape-constrained active contours

https://doi.org/10.1016/j.compbiomed.2022.106530Get rights and content

Highlights

  • A shape-constrained active contour model (SACM) is proposed for segmenting the carotid artery from MR images.

  • The shape-constraint maps generated by 2.5D CNN are integrated into SACM as prior information to detect the vessel wall.

  • The effectiveness of the proposed method is demonstrated by clinical data (25 patients, 4500 images).

  • The proposed method is shown to be feasible after being compared against manual delineations.

Abstract

Reconstruction of the carotid artery is demanded in the detection and characterization of atherosclerosis. This study proposes a shape-constrained active contour model for segmenting the carotid artery from MR images, which embeds the output of the deep learning network into the active contour. First the centerline of the carotid artery is localized and then modified active contour initialized from the centerline is used to extract the vessel lumen, finally the probability atlas generated by the deep learning network in polar representation domain is integrated into the active contour as a prior information to detect the outer wall. The results showed that the proposed active contour model was efficient and comparable to manual segmentation.

Introduction

Atherosclerosis, a kind of progressive disease, is manifested as thickened vascular wall and narrowed lumen at the early stage and plaque lesions at the late stage [1]. In unstable plaques, thin fibrous caps may rupture, resulting in the entry of plaque contents into the vascular lumen, finally causing cerebral ischemia or stroke. Therefore, it is critical to evaluate the thickness and compositions of the vascular wall accurately for identifying high-risk patients.

With the development of imaging technology, magnetic resonance imaging (MRI) has been widely used for noninvasive imaging diagnosis of carotid arteries and characterization of atherosclerotic plaques, which makes the assessment of disease progression possible [2]. There are reports of MRI-based automatic or semi-automatic detection methods for the vascular wall, including: (i) Geometric Deformable Model-Based Segmentation, which has been widely studied and mainly categorized as region-based active contour [3,4], and edge-based active contour [[5], [6], [7], [8]]. Active contour is a kind of segmentation method based on image intensity, so the lumen of the vessel can be easily detected using active contour due to the high contrast between the blood-filled lumen and the vessel tissues. However, when detecting vessel walls, such methods may lead to over-segmentation because of the similar intensity of the vessel wall and the surrounding tissues in MRI, so it is often necessary to introduce shape constraints on the active contours, such as the ellipse constraint [7], the prior shape of the lumen [8], and others [9,10]. In the study by Y. Wang et al. [8], the inner lumen was used as the prior shape to segment the outerwall, first, the lumen of the vessel was detected in contrast-enhanced MR angiography (CE-MRA) using the Hough transform and active contour, then, the contrast-enhanced MR images were registered with the black blood MR images, and the outer wall was segmented on the registered black blood MR images based upon the similarity between the inner and outer wall. (ii) Parametric Deformable Model-Based Segmentation, including ellipse fitting [11], statistical shape modeling, circle-model [12], and shape modeling [13,14]. In this approach, the vessels are segmented by fitting the model to the image data within a 2D (two dimensional) or 3D (three dimensional) Region. Typically, the shape of the vessel is assumed, and then the predefined shape is fitted to the vessel via parametric fitting. For example, in the study by R. van T Klooster et al. [14], a coarse deformable non-uniform rational B-spline surface (NURBS) was pre-constructed, and this blood vessel model was then deformed to fit the vessel wall. The model demonstrates the high accuracy and reproducibility with limited user interaction. However, some parametric models require dynamic reparameterization when the initial model and the desired object differ in shape. This is computationally expensive [15]. Moreover, in contrast to geometric deformable models, parametric deformable models cannot handle topological changes during deformation. (iii) Pattern Recognition Segmentation method, such as graph-based segmentation [16,17], threshold segmentation, clustering, etc. For graph-based methods [16,17], the regional probability graph was constructed using edge information or Support Vector Machine classifiers, and then the vessels surface was obtained by maximizing the graph while considering smoothness and topological constraints. The segmentation was performed in 3D, and this helps maintain smoothness and topology in spatial. But, same as active contours, it may still over-segment vessels when the vessel boundaries are not clearly shown in images, so the shape constraints are still required. (iv) Recent deep learning approaches for segmentation [[18], [19], [20], [21], [22], [23], [24], [25], [26]]. Some studies proposed end-to-end networks that can directly locate and segment vessels, such as a study by C. Zhu et al. [19], 3D images were directly fed into the proposed 3D residual U-net, and the U-net was able to make decisions based on “patch-level” and “global-level”. However, most studies segmented vessels in 2D after locating them. For example, a fully automated system was designed to segment the vessel wall in Ref. [20]. The system first located the vessel using a Tracklet Refinement algorithm, and then a convolutional neural network (CNN) was employed to segment the vessel wall in the polar representation domain. F. Shi et al. [21] straightened the vessel in the 3D domain and then segmented outerwall in the 2D cross-sectional plane using the U-net-like fully convolutional networks (FCN) architecture.

Related studies have shown the excellent performance of traditional active contour models and the great potential of deep learning methods. Limited by algorithms and the size of the data set, it is difficult to maintain the commonality of regions and produce continuous 3D segmentation. In this study, we combine these two methods to design a hybrid model which embeds the semantic information extracted by deep networks into the active contours. For deep networks that provide the prior information, we transform the representation domain of the input data and set continuous slices as input to generate prior information. For the active contour model, we introduce the edge term and shape constrain term to the original active contour model, and then embed the deep networks information into the active contour to produce smooth surfaces. A series of experiments were then performed on carotid MRI datasets to validate the effectiveness of the model.

Section snippets

Overview of the segmentation framework

The entire workflow of the proposed method can be summarized as follows (see Fig. 1 and Fig. 2).

  • (i)

    The original data were preprocessed to reduce noises and correct the inhomogeneity.

  • (ii)

    The centerline of a carotid artery was localized and the proposed shape-constrained active contour model (SACM) was initialized based on the centerline to detect the carotid artery lumen.

  • (iii)

    The probability map generated by the deep learning network in the polar (r, θ) representation domain was integrated into the SACM as

Image acquisition and annotation

The study was approved by the Human Research Ethics Committee at the Princess Alexandra Hospital (PAH) in Brisbane, Australia, and by the Queensland University of Technology's (QUT) Office of Research Ethics and Integrity (HREC/17/QPAH/181). The carotid datasets were acquired by Magnetom Prisma (Siemens Medical Solutions, Malvern, PA, USA) 3T MR whole body system from the Princess Alexandra Hospital (PAH) in Brisbane, Australia. Four contrast-weighted and Time of Flight (TOF) images were

Results

Image preprocessing highlighted the lumen, decreased tissue noises, and obviously corrected the inhomogeneity of images, and then the centerline was extracted on processed images. Table 1 and Table 2 demonstrate the mean metrics of our method over 5 folds. The results were evaluated on AA, BB, and CC separately [Fig. 4(a)]. The common carotid artery was included in BB, the bifurcation was included in AA, and the external and common carotid arteries were included in CC.

With the same training

Discussion

This study proposed a model for carotid artery segmentation, which embeds the output of the CNN into the active contour. The method was verified in a series of experiments subsequently. The preprocessing algorithm corrects the bias and facilitates the active contour to move in the right direction pulled by regional and local gradient forces [31].

From the experimental results, it can be seen that there is a slight improvement in the performance after improving the 2D network to 2.5D and

Conclusion

This paper proposes an active contour model for carotid segmentation. This model combines the semantic information extracted by multi-layer convolutional modules with the level set, which takes advantage in refining the results and keeping the spatial continuity. For convolutional modules that provide the prior information, the transformation of the representation domain yields a new set of geometrically constant image features and helps to deal with the inter-class imbalance problem. For the

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgement

This work is partially supported by the National Natural Science Foundation of China (11972118, 61821002, 12172089), Australia Research Council (DP200103492), and Medical Research Future Fund (2016165).

References (44)

  • A.J. Lusis

    Atherosclerosis., Nature

    (2000)
  • G. Trebuchet et al.

    Region-based Active Contours for Computer-Aided Analysis of Carotid Phase Contrast Mri

    (2013)
  • L. Saba et al.

    Semiautomated analysis of carotid artery wall thickness in mri

    J. Magn. Reson. Imag.

    (2014)
  • J.A. Rosado-Toro et al.

    Dynamic programming using polar variance for image segmentation

    Ieee T. Image Process.

    (2016)
  • D.C. Cheng et al.

    Automatic detection of the carotid artery boundary on cross-sectional mr image sequences using a circle model guided dynamic programming

    Biomed. Eng. Online

    (2011)
  • I.M. Adame et al.
    (2004)
  • J. Zhang et al.

    Automatic segmentation of mr depicted carotid arterial boundary based on local priors and constrained global optimisation

    IET Image Process.

    (2019)
  • S. Gao et al.

    Learning-based automated segmentation of the carotid artery vessel wall in dual-sequence mri using subdivision surface fitting

    Med. Phys.

    (2017)
  • R. van T Klooster et al.

    Automatic lumen and outer wall segmentation of the carotid artery using deformable three-dimensional models in mr angiography and vessel wall images

    J. Magn. Reson. Imag.

    (2012)
  • F. Bukenya et al.

    A Review of Blood Vessel Segmentation Techniques

    (2018)
  • E. Ukwatta et al.

    3-d carotid multi-region mri segmentation by globally optimal evolution of coupled surfaces

    IEEE Trans. Med. Imag.

    (2013)
  • A.M. Arias Lorza et al.

    Maximization of regional probabilities using optimal surface graphs: application to carotid artery segmentation in mri

    Med. Phys.

    (2018)
  • View full text