CN108765385B - Double-source CT coronary artery automatic extraction method - Google Patents

Abstract

The invention relates to a double-source CT coronary artery automatic extraction method, which comprises the following steps: s1, solving eigenvalues lambda 1, lambda 2 and lambda 3 of a Hessian matrix for each pixel point in the double-source CT data, and constructing a new blood vessel model equation; the new blood vessel model is utilized to effectively strengthen the blood vessel structure; the gray level of the blood vessel area is bright, and the background area is dark; performing Gaussian modeling on the blood vessel region and the background region by using a statistical model; s2, dividing the enhanced data into 2 types, a background area and a blood vessel area by using a Gaussian mixture model; s3, in the blood vessel region, the neighborhood relationship is restricted through the growth of a spherical model, the coronary artery is grown, and the pseudo branch blood vessel in the grown coronary artery is detected and removed through branch detection based on hierarchical clustering.

Description

Double-source CT coronary artery automatic extraction method

Technical Field

The invention relates to the field of medical image processing, in particular to a double-source CT coronary artery automatic extraction method.

Background

Coronary artery segmentation is the first step in image analysis of cardiovascular disease. CT angiography is one of the currently prevalent examination modalities. The CT angiography data volume is large, the individual difference of patients is large, and the cardiovascular structure is complex. At present, the coronary artery segmentation method has the problems that manual interaction is needed, and pipeline branching errors are easy to occur during pipeline reconstruction, so that analysis of a doctor on a blood vessel is directly influenced. Therefore, most doctors still use two-dimensional images such as Multi-planar reconstruction (MPR), Curved reconstruction (CPR), and Maximum Intensity Projection (MIP) to observe blood vessels and analyze the condition of the blood vessels by a semi-manual method to make a diagnosis, which depends on the subjective experience of the doctors, takes a lot of time, and easily delays the disease condition.

The existing coronary artery extraction methods mainly include a region growing method, a statistical method, an active contour model method, a centerline method, multi-scale filtering and the like, and most of the existing coronary artery extraction methods simply use the gray level features or morphological features of blood vessels for judgment. Frangi proposes a three-dimensional blood vessel enhancement method based on a multi-scale space, and judges whether each individual pixel belongs to a tubular structure [1] by analyzing eigenvalues of hessian matrix under different scales, but the method is not enough in response to a blood vessel boundary, is sensitive to noise, and greatly enhances the noise while enhancing the blood vessel. Tim Jerman improved Frangi's method and also successfully extracted blood vessels with hemangiomas [2 ]. Yang improves the Frangi method, adds GVF force field information, can extract the whole coronary artery [3], however, the lumen boundary can not be accurately detected. In 2016, Zhaojie et al used an automatic vessel segmentation method based on multi-scale filtering and probabilistic decision [4 ]. Since this method is directly based on the multi-scale filter enhancement method of Frangi, the shortcomings of the Frangi method are also reflected in the segmentation result. Glorio et al better segment coronary arteries from dual-source CT data, but do so based on two-dimensional data for each slice, and the resulting segments do not reconstruct the vessel in three dimensions smoothly [5 ]. In previous work, Huang Chang Peng et al proposed a sphere model, treating the vessels as a binary tree, and extracting the centerline of the vessel by rolling the sphere [6-7 ]. But when multiple branches are encountered, pseudo-vessels may appear, which is time consuming in traversing the vessel.

The above scheme has the following technical defects in specific use:

1. coronary arteries cannot be accurately extracted based on gray level features or morphological features. Because, the coronary vascular network presents very complex structural and pathological features such as vessel dimensions and curvatures varying to a large extent, as well as problems of vessel stenting, calcification, aneurysms, luminal narrowing, etc. in pathological and traumatic situations.

2. In the existing heart blood vessel segmentation methods, some methods can well segment blood vessels in a two-dimensional image, but the model is complex and slow, and is not suitable for three-dimensional segmentation. Some methods require too many manual interactions, which restricts the application of the methods.

3. Due to the fact that the coronary arteries and surrounding tissues are close in gray level, and noise interference is caused by motion artifacts and external magnetic fields, the method of Frangi et al is weak in response to the boundary of blood vessels, and the method of Huang-Tec et al is difficult to detect multi-branch blood vessels.

Reference to the literature

[1]Frangi,A.F.,Niessen,W.J.,Vincken,K.L.,Viergever,M.A.Multiscale vessel enhancement filtering.Medical Image Computing and Computer-Assisted Intervention 1496,130-137,http://dx.doi.org/10.1007/BFb0056195(1998).

[2]Tim,J.,Franjo,P.,Bostjan,L.&Ziga,S.Beyond Frangi:an improved multiscale vesselness filter.International Conference on

Information Science3,http://dx.doi.org/10.1117/12.2081147(2015).

[3]Yang,Y.,Tannenbaum,A.,Giddens,D.&Stillman,A.Automatic segmentation of coronary arteries using Bayesian driven implicit

surfaces.IEEE International Symposium on Biomedical Imaging5,189-192,http://dx.doi.org/10.1109/ISBI.2007.356820(2007).

[4] Zhaojie, Jiang Shi faithful, Huang Zhan Peng, Eurasian and Xing, research on a double-source CT coronary artery three-dimensional segmentation method, biomedical engineering research, 2016,35(3):197-201.

[5]Ken Cai,Rongqian Yang,Lihua Li,et al.A Semi-Automatic Coronary Artery Segmentation Framework Using Mechanical Simulation[J].J Med Sys.2015,39(10):1-7.

[6]Jiang,S.,Xiao,W.,Ou,S.,Huang,Z.&Zhao,J.Three Dimensional Coronary Artery Centerline Extraction Based on Sphere Model with Branch Detection.Journal of Computational Information Systems9,9833-9840,http://dx.doi.org/10.12733/jcis8163(2013).

[7]Huang,Z,Jiang,S.,Liu,Y.,Bao,S.Interactive 3D Liver Vessels Centerline Extraction based on Moving Sphere Model.International Conference on Information Science3,2100-2103,http://dx.doi.org/10.1109/InfoSEEE.2014.6946295(2014).

Disclosure of Invention

Aiming at the problems that the existing heart coronary artery segmentation method is difficult in branch detection and needs manual operation, the gray-scale feature, morphological feature and spatial relation of blood vessels are considered, a statistical classification method is combined, a multi-scale filtering and spherical model is improved, and the three-dimensional coronary artery segmentation can be automatically realized without the interactive point selection of doctors.

In order to realize the purpose, the technical scheme is as follows:

a dual-source CT coronary artery automatic extraction method comprises the following steps:

s1, solving eigenvalues lambda 1, lambda 2 and lambda 3 of a Hessian matrix for each pixel point in the double-source CT data, and constructing a new blood vessel model equation:

wherein

The new blood vessel model is utilized to effectively strengthen the blood vessel structure; the gray level of the blood vessel area is bright, and the background area is dark; performing Gaussian modeling on the blood vessel region and the background region by using a statistical model;

s2, dividing the enhanced data into 2 types, a background area and a blood vessel area by using a Gaussian mixture model;

s3, in the blood vessel region, the neighborhood relationship is restricted through the growth of a spherical model, the coronary artery is grown, and the pseudo branch blood vessel in the grown coronary artery is detected and removed through branch detection based on hierarchical clustering.

Preferably, the step S3 includes the following steps of performing detection based on hierarchical clustering:

(1) finding a sphere meeting the condition from the neighborhood of the root node;

(2) if the radius is larger than 0, all points with R as the radius are searched, otherwise, the operation is finished;

(3) if a conforming sphere is found, constructing a new sphere with the radius of R +1, R +2 and R + 3; otherwise, R is R-1, and whether the radius is larger than 0 is recalculated;

(4) clustering rays emitted from the center of the sphere by a top-down hierarchical clustering method;

(5) if the cluster is 2 types, setting the sphere as the left child of the current node; if the cluster is 3 types, setting the sphere as the right child of the current node;

(6) and (5) repeating the steps (2) to (5) and searching for the next new node until R <0, and ending.

Compared with the prior art, the invention has the beneficial effects that:

(1) the prior art only focuses on morphological, gray-scale or neighborhood features when extracting blood vessels. The invention detects blood vessels from three aspects simultaneously. The shape of the blood vessel is tubular, the eigenvalue of a Hessian matrix is calculated for each pixel point, the eigenvalues lambda 1, lambda 2 and lambda 3 are calculated for the Hessian matrix of each pixel point, and the eigenvalue can be used for distinguishing tubular, sheet and dot-shaped structures. The gray scale of the blood vessel area of the enhanced image is brighter, the background area of the enhanced image is darker, and Gaussian modeling is carried out on the blood vessel class and the background class by utilizing a statistical model. In the blood vessel class, the neighborhood relationship is restricted by the growth of a spherical model, and coronary arteries are grown. In order to avoid false branch vessels and save time, hierarchical clustering based branch detection is established.

(2) The invention improves the tubular structure enhancement method proposed by frangi, modifies Rb factor, reduces the influence of uneven gray scale and noise by using a matrix mode for substitution, and cancels parameter beta to obtain a new blood vessel equation.

(3) Although the sphere model can traverse the tree structure to obtain the segmentation result of the blood vessel, the operation efficiency is also high. However, the coronary arteries are branched, and the wrong branching causes pseudo-blood vessels and consumes a lot of computation time. In order to solve the above problems, the present invention provides a branch detection method based on clustering. And judging the number of branches at the current position by judging the number of clusters. If the ray emitted from the center point of the current ball is gathered into 2 types, the current position is not branched. If the ray is grouped into 3 types, it indicates that there is a branch.

Drawings

FIG. 1 is a flow chart of a method provided by the present invention.

Fig. 2 is a schematic diagram of a ball searching along a blood vessel.

FIG. 3 is a schematic diagram of cluster-based branch detection.

FIG. 4 is a flow chart of a sphere model based on branch detection.

Fig. 5 is a left coronary artery segmentation result graph.

FIG. 6 is a comparison of a vessel enhancement method.

FIG. 7 is a graphical representation of the ROC curve for the present invention and Frangi coronary artery enhancement method.

Fig. 8 is a schematic diagram of coronary artery centerline and caliber measurement.

Detailed Description

The drawings are for illustrative purposes only and are not to be construed as limiting the patent;

the invention is further illustrated below with reference to the figures and examples.

Example 1

The invention simultaneously segments coronary artery from three aspects of blood vessel shape, gray level feature and neighborhood relation. The shape of the blood vessel is tubular, and the characteristic value of a Hessian matrix is calculated for each pixel point, the Rb operator of a frangi model is improved, and a new blood vessel model is established to enhance the blood vessel structure. The gray scale of the blood vessel area of the enhanced image is brighter, the background area of the enhanced image is darker, and Gaussian modeling is carried out on the blood vessel class and the background class by utilizing a statistical model. In the blood vessel class, the neighborhood relationship is restricted by the growth of a spherical model, and coronary arteries are grown. In order to avoid false branch vessels and save time, hierarchical clustering based branch detection is established. The general flow chart of the specific implementation method of the invention is shown in fig. 1.

The present invention improves the vascular enhancement model of Frangi. Since the method proposed by Frangi to enhance blood vessels is sensitive to noise and also enhances noise, some improvement is needed to suppress background noise. The invention modifies Rb factor in frangi method, uses matrix mode to replace to reduce the influence of gray scale unevenness and noise, and cancels parameter beta.

Wherein

Assuming that each tissue class in the image conforms to the same Gaussian distribution, all the classifications in the image can be regarded as the sum of a plurality of Gaussian distributions, and the distribution of the whole volume data is formed by linearly combining the Gaussian probabilities of the tissue classes, as shown in formula (2), w_kThe weight occupied by each type of organization.

For the enhanced image, the low gray scale region corresponds to the background, and the high gray scale region corresponds to the blood region. Thus, data can be classified into 2 types. The total data probability distribution is formed by overlapping gaussian distributions of class 2 tissues, as shown in formula (3).

Parameter w_k，u_k，σ_kDepending on the maximum expected method.

And E, calculating the clustering probability of each example by using the distribution parameters, wherein t represents the number of iteration steps as shown in formula (4).

And M, re-estimating the distribution parameters by using maximum likelihood estimation to obtain:

all parameters of the Gaussian mixture model are obtained through an EM (effective velocity) method, and then, the tissues belong to the kth class when the distinguishing condition is that k is not equal to j by distinguishing which class of tissues each voxel belongs to.

When the centerline of a coronary artery is extracted, starting from a certain designated position of the coronary artery, assuming that an ellipsoid with variable size advances along the blood flow direction of the coronary artery according to a pre-designated condition, and continuously adjusting the size of the sphere according to the size of the inner diameter of a blood vessel at the position to which the ellipsoid arrives, so that the current ellipsoid is an inscribed ellipsoid with the largest radius at the position, when the ellipsoid meets a bifurcation point of the coronary artery in the moving process, the current ellipsoid is split into two ellipsoids, the ellipsoid continues to advance along each branch of the coronary artery respectively according to the method, the method is stopped until the ellipsoid which meets the condition cannot be searched, and finally the sphere centers of the ellipsoids at the obtained positions are connected, and the connecting line forms the centerline of the whole coronary artery, as shown in fig. 2.

As the search progresses, the ellipsoid needs to advance along the direction of the blood in the blood vessel to complete the extraction of the whole blood vessel centerline. As can be seen from fig. 2, after the maximum inscribed ellipsoid at the current position, the sphere center of the next ellipsoid can be selected from the semi-ellipsoid of the current maximum inscribed ellipsoid, so that the moving step length of the ellipsoid is the distance from the sphere center of the current maximum inscribed ellipsoid to each point on the semi-ellipsoid. And (3) regarding each point on the semi-ellipsoid as the sphere center of the next inscribed ellipsoid, calculating the maximum inscribed ellipsoid taking each point as the sphere center according to the threshold condition of the formula (8), and finally selecting the ellipsoid with the largest volume from the maximum inscribed ellipsoids corresponding to each point to obtain the next maximum inscribed ellipsoid.

Where p (x, y, z) is the coordinate of any point and g (x, y, z) is the grayscale of point p (x, y, z). Through multiple experiments, the empirical threshold h of the coronary artery blood vessel is between 0.9 and 0.95.

Although the sphere model can traverse the tree structure to obtain the segmentation result of the blood vessel, the operation efficiency is also high. However, the coronary arteries are branched, and the wrong branching causes pseudo-blood vessels and consumes a lot of computation time. When the sphere is rolled to a vessel branch, it is a question how to determine the number of branches and to search for the correct branch.

In order to solve the above problems, the present invention provides a branch detection method based on clustering. As shown in fig. 3, rays are emitted from the center point of a sphere, with radius R, and the test is performed with spheres constructed from R +1, R +2, and R +3, respectively. Rays reaching the outermost sphere are clustered into several types, and the number of branches at the current position is judged by judging the number of clusters. If the ray emitted from the center point of the current ball is gathered into 2 types, the current position is not branched. Rays 1 and 2 are grouped into 3 classes as in fig. 3, indicating a branch.

In a general clustering method, k values are selected, namely, before classification, the k values are determined to be classified into several classes. However, because the vascular structure is complex, the individual difference is large, and the vascular malformation of patients exists, the classification into several categories before the clustering is difficult to determine, and the hierarchical clustering method can well solve the problem.

Assuming that there are N rays to be clustered, for top-down hierarchical clustering (iterative clustering), the basic steps are:

step 1, initialization. Classifying each ray sample into one class, and calculating the distance between every two classes, namely the similarity between the samples;

and 2, calculating the distance between each class, setting a threshold value, and considering that the iteration can be terminated when the distance between the two closest classes is greater than the threshold value.

Step 3, finding out the two nearest classes, and classifying the two classes into one class;

and 4, recalculating the distance between the newly generated classes and categories, and skipping to the step 2.

And calculating Euclidean distances between the ray and each point intersected with the outer surface of the sphere, and calculating the distance between the two classes by adopting a SingleLinkage method, namely taking the distance between two samples with the shortest distance in the two classes as the distance between the two sets. The single linking method is shown in equation 9.

D_min＝min|p₁-p₂I formula (9)

Wherein p1 and p2 are two points in different classes.

If the distance between the classes is greater than the distance calculated by the following equation (10), the clustering operation is terminated.

Where α is the angle between ray 1 and ray 2, as shown in FIG. 3.

A flow chart of the 3D vessel branch detection method designed by the present invention is shown in fig. 4. The method comprises the following specific steps:

(1) finding a sphere meeting the condition from the neighborhood of the root node;

(2) if the radius is larger than 0, all points with R as the radius are searched, otherwise, the operation is finished;

(3) if a conforming sphere is found, constructing a new sphere with the radius of R +1, R +2 and R + 3; otherwise, R is R-1, and whether the radius is larger than 0 is recalculated;

(4) clustering rays emitted from the center of the sphere by a top-down hierarchical clustering method;

(5) if the cluster is 2 types, setting the sphere as the left child of the current node; if the cluster is 3 types, setting the sphere as the right child of the current node;

(6) and (5) repeating the steps (2) to (5) and searching for the next new node until R <0, and ending the method.

The invention combines statistics, multi-scale space and a sphere model method based on gray scale, considers the gray scale characteristics, morphological characteristics and spatial position relation of coronary artery, and can accurately and automatically segment the coronary artery. As shown in fig. 5, the segmented left coronary artery has better reconstruction effect, avoids the problem of blood vessel leakage, has no pseudo blood vessel and makes the artery stenosis clearly visible.

The new method for enhancing the blood vessel provided by the invention firstly carries out Hessian operation on each pixel of volume data, and then detects the blood vessel by using a formula (4) to obtain a figure 6(b), so that the tubular structure is enhanced. Fig. 6(c) is an image obtained by the vascular enhancement method proposed by Frangi. By comparing the results of fig. 6(b) and 6(c), there were no significant holes and fractures in (b) and the enhancement effect was superior to (c).

The two methods were evaluated using Receiver Operating characteristics Curve (ROC) to obtain FIG. 7. Blue is the evaluation result of the present invention, red is the evaluation result of Frangi, and it can be seen that the improved effect of the present invention is obtained, the Area Under the Curve (AUC) becomes large, and the evaluation effect is superior to that of the Frangi method.

Aiming at the difficulty of extracting multiple coronary branches by a spherical model, the invention provides a branch detection method based on clustering. The coronary artery extracted in fig. 8 is obviously narrow, and the narrow part is only 1-2 pixels wide, but the method provided by the invention can still be correctly divided without fracture.

It should be understood that the above-described embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. And are neither required nor exhaustive of all embodiments. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.

Claims (2)

1. A double-source CT coronary artery automatic extraction method is characterized in that: the method comprises the following steps:

s1, solving characteristic value lambda of Hessian matrix for each pixel point in double-source CT data₁,λ₂And λ₃And constructing a new blood vessel model equation:

wherein

The new blood vessel model is utilized to effectively strengthen the blood vessel structure; the gray level of the blood vessel area is bright, and the background area is dark; performing Gaussian modeling on the blood vessel region and the background region by using a statistical model;

s2, dividing the enhanced data into 2 types, a background area and a blood vessel area by using a Gaussian mixture model;

s3, in the blood vessel region, the neighborhood relationship is restricted through the growth of a spherical model, the coronary artery is grown, and the pseudo branch blood vessel in the grown coronary artery is detected and removed through branch detection based on hierarchical clustering.

2. The dual-source CT coronary artery automatic extraction method according to claim 1, characterized in that: the step S3 uses hierarchical clustering to perform detection, which includes the following steps:

(1) finding a sphere meeting the condition from the neighborhood of the root node;

(2) if the radius is larger than 0, all points with R as the radius are searched, otherwise, the operation is finished;

(3) if a conforming sphere is found, constructing a new sphere with the radius of R +1, R +2 and R + 3; otherwise, R is R-1, and whether the radius is larger than 0 is recalculated;

(4) clustering rays emitted from the center of the sphere by a top-down hierarchical clustering method;

(5) if the cluster is 2 types, setting the sphere as the left child of the current node; if the cluster is 3 types, setting the sphere as the right child of the current node;

(6) and (5) repeating the steps (2) to (5) and searching for the next new node until R <0, and ending.

Images