Cerebrovascular segmentation of TOF-MRA based on seed point detection and multiple-feature fusion

https://doi.org/10.1016/j.compmedimag.2018.07.002Get rights and content

Highlights

  • A novel method is proposed to extract 3D vascular structures from time-of-flight magnetic resonance angiography dataset.

  • A seed point detection method is proposed based on maximum intensity projection, and it can automatically extracted sufficient seed points that located in blood vessels.

  • A fuzzy interference-based vascular multiple-feature fusion method is proposed to enhance the intensity information of vascular structures in the TOF-MRA image.

Abstract

The accurate extraction of cerebrovascular structures from time-of-flight (TOF) data is important for diagnosis of cerebrovascular diseases and planning and navigation of neurosurgery. In this study, we proposed a cerebrovascular segmentation method based on automatic seed point detection and vascular multiple-feature fusion. First, the brain mask in the T1-MR image is detected to enable the extraction of the TOF brain structure by simultaneously acquiring the TOF image and its corresponding T1-MRI. Second, local maximum points are detected on three maximum-intensity projections of TOF-MRA data and then be traced back in three-dimensional space to detect seed points for the initialization of vascular segmentation. Third, the TOF-MRA image and its corresponding vesselness image are fused to enhance vascular features on the basis of fuzzy inference for the extraction of whole cerebrovascular structures, particularly miniscule cerebral vessels. Finally, detected seed points and multiple-feature fused enhanced images are provided to the procedure of region growing, and cerebrovascular structures are segmented. Experimental results show that compared with traditional methods, the proposed method has higher accuracy for vascular segmentation and can avoid over- and under-segmentations. The proposed cerebrovascular segmentation method is not only effective but also accurate. Therefore, it has potential clinical applications.

Introduction

Cerebrovascular disease, which is mainly caused by vascular stenosis or aneurysm as a result of vascular malformations, is a major threat to human health. Three-dimensional time-of-flight magnetic resonance angiography (TOF-MRA) is a popular clinical cerebrovascular imaging modality given its non-invasiveness, rapidity, and high resolution. The accurate segmentation of the three-dimensional structure of blood vessels from TOF images is important for the diagnosis and quantitative analysis of cerebrovascular disease. In addition, accurate cerebrovascular segmentation is an important prerequisite for neurosurgical planning and navigation (Du et al., 2013).

Over the past two decades, numerous methods have been proposed for the segmentation of three-dimensional cerebrovascular structures from TOF-MRA (Kirbas and Quek, 2004; Lesage et al., 2009). The statistical model-based method is a typical method for cerebrovascular segmentation and based on the principle of Bayesian statistical classification. This method constructs two grayscale distribution functions to fit the background and blood vessels in an image. The threshold for blood vessel segmentation is obtained by optimizing grayscale distribution functions. Hassouna et al. (2006) proposed a vascular segmentation method based on a statistical model. In this method, the intensity of the background and blood vessels is fitted using a Rayleigh distribution function and two normal distribution functions. Expectation maximization algorithm is utilized to optimize initial model parameters, which would converge to an optimal solution. Wen et al. (2014) introduced normal and Rayleigh functions to simulate the image background and used a normal function to fit blood vessels. Then they utilized a particle-swarm optimization algorithm to optimize the parameters of the model. Blood vessels are mainly located in an area of high intensity in a TOF-MRA dataset. Therefore, large vessels of high intensity could be easily differentiated when using a statistical model-based approach to segment blood vessels. However, some small vessels are not easily identified using statistical models because of their low intensity.

Vascular tracking is also a commonly used vascular segmentation method and operates on the principle of vascular grayscale distribution. Unlike the statistical model-based method, which uses the entire gray distribution of blood vessels in an image, vascular tracking uses the characteristics of a blood vessel’s cross section, which has a Gaussian-like intensity distribution (Friman et al., 2010; Yang et al., 2009). Therefore, in this method, the grayscale distribution of a vascular cross-section is firstly fitted by tracking starting from some defined initial seed points, and vascular centerlines, diameters, and bifurcations are extracted to reconstruct the spatial structure of blood vessels (Lesage et al., 2009). For example, Aylward et al. (Aylward et al. (1996); Aylward and Bullitt, 2002) proposed a vascular segmentation method based on multi-scale ridges and the features of vascular sectional grayscale distribution. The method extracts whole vascular structures through ridge tracking on the basis of the gradient and Hessian matrix-defined vascular ridge. The tracking-based vascular segmentation method can effectively detect small vessel branches and ensure blood vessel integrity. However, extracting the section of vascular disease, where the grayscale distribution is different from Gaussian distribution, is difficult. Tracking would be prematurely terminated, resulting in an incomplete segmentation result.

Region growing is another commonly used vascular segmentation method (Masutani et al., 1998). In region growing, segmentation is based on initially defined seed points and expands iteratively to obtain the vascular area by analyzing the relationship between adjacent voxels and the segmented region until converging to the final segmentation result. Since blood vessels are connected in human’s body, vascular structures can be iteratively extracted by region growing algorithm, if one or a small amount of seed points are placed on each vascular branch (Yang et al., 2014; Cong et al., 2015). Therefore, compared with previous two methods, it is more simple and effective. However, grayscale distribution within the blood vessel region is not very consistent in practice; thus, a large number of seed points have to be set to guarantee completely segmented vascular structures. In addition, manual seed point setting is not only time consuming but also cannot be replicated, and different seed points provided by different operators might result in different vascular segmentation results. Therefore, if a large number of seed points could be automatically detected and precisely located within the blood vessel, the efficiency of segmentation might be greatly improved.

Currently, only a few papers have reported the detection of seed points in three-dimensional space. The majority of related studies have focused on the detection of seed points in a two-dimensional angiographic image. For example, Fritzsche et al. (2003) proposed a seed-point detection method by extracting local maximum points on a retinal image. The image is sampled and compared with four adjacent pixels to filter out seed points. We previously (Xiao et al., 2015) proposed a seed-point detection method based on ridge-point extraction from coronary angiograms. A ridge-point detection discriminant is constructed by the gradient and Hessian matrix of pixels to extract the amount of seed points located on vascular centerlines.

If seed points are local maximum points in a three-dimensional space, then they would be also the local maximum points in a two-dimensional projection plane. In accordance with this principle, an automatic seed-point detection method based on maximum intensity projection (MIP) is proposed. First, a TOF-MRA image is projected onto a two-dimensional plane, and local maximum intensity points are extracted in the projection plane. Then, these points are traced back to the original three-dimensional space to obtain their three-dimensional positions. Compared with the method for the direct detection of seed points in a three-dimensional space, the proposed method extracts three-dimensional seed points from a two-dimensional MIP image.

In this paper, the brain region in TOF-MRA will be extracted to remove the interference of the skull during cerebrovascular segmentation. Then, seed points in the TOF-MRA image are detected using the proposed algorithm. An enhanced vesselness image can be calculated according to the Hessian matrix of the TOF-MRA image in multi-scale space because the intensity of small vascular braches is weak and difficult to detect in the original TOF image. A fuzzy interference-based vascular multiple-feature fusion method is proposed to combine the features of the TOF-MRA image and its corresponding vesselness image. Finally, all detected seed points are utilized for region growing to segment vascular structures on the enhanced vascular multiple-feature fusion image.

Section snippets

Method

The main steps of the proposed method for vascular segmentation can be divided as follows: First, the brain mask in the T1-MR image is detected to enable the extraction of the TOF brain structure by simultaneously acquiring the TOF image and its corresponding T1-MRI. Second, seed points are automatically detected in the extracted TOF brain structure. Third, the vascular multiple-feature fusion image is constructed. Finally, three-dimensional vascular structures are segmented by region growing

Experimental result and discussion

Data from 12 clinical cases were randomly selected and tested in this study to evaluate the feasibility and effectiveness of the proposed algorithm. The clinical dataset comprises seven men and five women aged 14 to 50 years old (average age = 30.6 years). The dataset was provided by Tsinghua University Imaging Center. The TOF datasets were evaluated on a 3 T Philips Medical MRI Systems (Achieva) using a 32-channel head coil. The main angiography parameters were as follows: repetition

Conclusions

The accurate segmentation of cerebrovascular structures from TOF-MRA datasets is important for diagnosis of cerebrovascular diseases and planning and navigation of neurosurgery. A novel cerebrovascular segmentation method is proposed in this study. The proposed method can be divided into four steps: brain mask detection, seed-point detection, vascular multiple-feature fusion, and region growing segmentation. In seed-point detection, seed points are extracted on the basis of MIP. Numerous seed

Conflict of interest

The authors declared that they have no conflicts of interest to this work.

Acknowledgments

This work was supported by National Key R&D Program of China (2017YFA0205904, 2016YFC 0105803), National Natural Science Foundation of China (81471759, 61701022), and the Fundamental Research Funds for the Central Universities (FRF-TP-16-045A1).

References (24)

  • R. Wang et al.

    Threshold segmentation algorithm for automatic extraction of cerebral vessels from brain magnetic resonance angiography images

    J. Neurosci. Methods

    (2015)
  • S.R. Aylward et al.

    Initialization, noise, singularities, and scale in height ridge traversal for tubular object centerline extraction

    IEEE Trans. Med. Imaging

    (2002)
  • Cited by (10)

    • Cerebrovascular segmentation in phase-contrast magnetic resonance angiography by multi-feature fusion and vessel completion

      2022, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      For this reason, Xiao et al. Xiao et al. (2018) adopted a fuzzy-based method of grayscale features and vascular structure features, which can balance the two features well and effectively extract cerebrovascular from TOF-MRA images. However, this method is cumbersome and has weak generalization, so these traditional methods will not be the first choice for current research.

    • Generative adversarial network based cerebrovascular segmentation for time-of-flight magnetic resonance angiography image

      2022, Neurocomputing
      Citation Excerpt :

      The author in [21] used a statistical model for cerebrovascular segmentation, which describes the intensity histogram of the TOF-MRA data by mixing one Rayleigh and two normal distributions. The author in [22] proposed a region growing model, which uses local maximum points for the initialization seeds of vascular segmentation. Summarily, these model-based methods are based on some assumptions, and their performance depends on some manually determined parameters tuned on a local dataset, which limits the generalization in a large dataset [23].

    • A comparative analysis framework of 3T and 7T TOF-MRA based on automated cerebrovascular segmentation

      2021, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      Xiao et al. Xiao et al. (2018b) also proposed cerebrovascular segmentation based on seed point detection, but this method used the traditional Frangi‘s Hessian enhancement. In recent years, deep learning has achieved remarkable results in medical image processing.

    • Carotid artery segmentation using level set method with double adaptive threshold (DATLS) on TOF-MRA images

      2019, Magnetic Resonance Imaging
      Citation Excerpt :

      Jodas et al. [12] combined K-means algorithm and active contour in coronary artery segmentation. Xiao et al. [13] applied region growing with detected seed points to segment cerebrovascular on multiple-feature fused enhanced images. Wang et al. [14] extracted the cerebral vessels with the threshold, which was calculated by comparing Gumbel distribution and unilateral normal distributions.

    • Implementation of deep learning neural network for retinal images

      2022, Research Anthology on Improving Medical Imaging Techniques for Analysis and Intervention
    View all citing articles on Scopus
    View full text