CN104504737B - A kind of method that three-dimensional tracheae tree is obtained from lung CT image - Google Patents

Abstract

The present invention provides a kind of method that three-dimensional tracheae tree is obtained from lung CT image, comprises the following steps：Main bronchus is extracted using adaptive three-dimensional spatial area growth method；Other segmenta bronchopulmonalias in addition to main bronchus are extracted using the method for optimization image characteristics extraction；Main bronchus and the segmenta bronchopulmonalia are carried out by " suture " using fuzzy connectedness algorithm, obtain three-dimensional tracheae tree.In the processing scheme of the present invention, main bronchus and segmenta bronchopulmonalia are first extracted respectively, efficiently solves the problem of main bronchus in conventional art easily blocks, segmenta bronchopulmonalia is easily revealed.Then, the main bronchus extracted, segmenta bronchopulmonalia are combined into one, obtain complete three-dimensional tracheae tree.

Description

A Method for Obtaining 3D Tracheal Tree from CT Images of Lungs

technical field

The invention relates to the technical field of computer image processing, in particular to a method for obtaining a three-dimensional trachea tree from a lung CT image.

Background technique

Accurate pulmonary tracheal tree is the basis for automatic diagnosis of pulmonary tracheal parameters, and accurate extraction of pulmonary tracheal tree is of great significance for the computer-aided diagnosis system of pulmonary diseases. Using preoperative three-dimensional CT images of the lungs, combined with medical image processing technology, computer graphics technology and modern electronic technology to reconstruct the three-dimensional tracheal tree, to achieve real-time guidance during the operation, which can greatly reduce the probability of accidents in patients undergoing bronchoscopy.

In order to reconstruct more levels of bronchi, many researchers have proposed the method of using the structural and anatomical information of the bronchial tree and related information. Such methods can be roughly divided into five categories: 1) knowledge or rule-based methods; 2) template matching methods; 3) morphological methods; 4) shape analysis methods; 5) hybrid methods. Knowledge-based or rule-based methods try to introduce prior knowledge of the anatomical relationship between the pulmonary bronchi and blood vessels, tracheal geometry, local image features, fuzzy logic, and connectivity when reconstructing the bronchial tree. The principle of the template matching method is to pre-define a set of masks or templates of different sizes according to the anatomical structure of the bronchi to assist the extraction of bronchial structural features in the two-dimensional or three-dimensional image space. For example, Kaftan tries to use the structural template of the tree path for the reconstruction of the bronchi. Morphological methods are often used to refine 3D bronchial trees that have been initially reconstructed (e.g., by 3D region growing algorithms), and the design idea is to try to use various morphological operators to connect or mix broken 2D/3D bronchial image regions . For example, Aykac et al. proposed to use dilated morphological operators to connect the same regions in adjacent image layers to improve the accuracy of bronchial reconstruction of single-layer images. In the literature, Fetita et al. proposed a mathematical morphology method based on a connection cost function to detect bronchial regions. The local image feature method is based on the characteristics of the bronchi that usually appear as a duct structure, and uses the method of solving the eigenvalues of the Hessian matrix to enhance and reconstruct the three-dimensional tubular tracheal tree by analyzing the second derivative of the tracheal boundary.

As is well known to those skilled in the art, leaks and obstructions are two major problems in bronchial tree reconstruction today. The main cause of leakage and obstruction is the partial volume effect of CT images, which reduces the lumen contrast between the wall of the pulmonary bronchus and the air in the airway. Leakage will result in fusion of the reconstructed tracheal tree with its surrounding lung tissue (eg, lung parenchyma); whereas obstruction will result in rupture of the reconstructed tracheal tree and discontinuity of the reconstructed trachea. In addition, image noise, image artifacts, and motion artifacts or image blur caused by respiratory movement during imaging will pose great challenges to the reconstruction of the bronchial tree. And when encountering certain lung diseases, such as chronic obstructive pulmonary disease or interstitial lung disease, the difficulty of reconstruction will be more obvious. Leakage and obstruction are a pair of contradictions in the extraction of the 3D tracheal tree, and this contradiction cannot be completed under a single algorithm framework.

Therefore, there is an urgent need for a method for obtaining a three-dimensional tracheal tree from lung CT images that can reduce the effects of leakage and obstruction.

Contents of the invention

The purpose of the present invention is to provide a method for obtaining a three-dimensional tracheal tree from lung CT images that can reduce the influence of leakage and obstruction during tracheal reconstruction. The method comprises the following steps: using an adaptive three-dimensional space region growing method to extract the main bronchi; using an optimized image feature extraction method to extract other bronchial segments except the main bronchi; The segments are "stitched" to obtain a three-dimensional tracheal tree.

As a preferred solution, before the step of extracting the main bronchus, the CT image is first smoothed, and the CT value of a certain voxel point in the CT image and the second-order arrangement structure of the local brightness change around it are combined and analyzed , and analyze whether the voxel points belong to the tubular structure, screen out the voxel points belonging to the bronchi, and gather all the voxel points belonging to the bronchi to obtain a first-level trachea preprocessing image.

As a preferred solution, before the step of extracting the main bronchus, the closed operation is performed on the first-level tracheal pre-processing image, and the structural elements of the first-level tracheal pre-processing image are firstly expanded, and then the expanded The final image is corroded with structural elements, which are used to fill small holes and smooth the boundaries of objects to obtain a secondary trachea preprocessed image.

As a preferred solution, in the step of extracting the main bronchus, it specifically includes the following steps: selecting a starting seed point; selecting the adaptive local adjacent threshold method as a criterion for region growth to obtain a threshold, which will be greater than or equal to the threshold The voxel points are merged into the seed area as the seed points; when all the surrounding seed points are merged into the seed area, the image of the main bronchus is obtained.

As a preferred solution, in the step of extracting other bronchial segments except the main bronchus, the following steps are specifically included: extracting several image features to construct energy items in the cost function; adopting the method of multi-kernel learning, using the The combined kernel function of the above features is embedded into the three-dimensional bronchial seed point extraction algorithm to obtain the seed points that make up the bronchial segment; according to the continuity, the seed points that make up the bronchial segment are continuously obtained to obtain each independent bronchial segment.

As a preferred solution, the several image features include 3D pipeline feature coefficients based on multi-scale Hessian matrix, local phase, SIFT features, and low-resolution version Haar-like features.

As a preferred solution, a method based on the Snake spline model is used in the step of performing continuous operation of the seed points forming the bronchial segment.

As a preferred solution, in the step of "sewing" the main bronchus and the bronchial segment, the main bronchus and the independent bronchial segment at the adjacent end are first sutured, and then iterative calculations are performed until all The bronchial segments are "sewn" together.

As a preferred solution, in the step of "sewing" the main bronchus and the bronchial segment, the central line extraction algorithm is used to extract the center line of the main bronchus, and the spatial distance, local image features, and bronchial anatomy are integrated. One or more of the criteria in the structure constructs a fuzzy connectivity function, starting from the end of the main bronchus, using a three-dimensional fuzzy connection algorithm to connect the main bronchus and the bronchial segment.

As a preferred solution, in the step of "sewing" the main bronchus and the bronchial segment, the centerline extraction algorithm is used to extract the centerline of each bronchial segment, and the spatial distance, local image features, and bronchial anatomical structure are integrated. One or more of the criteria are used to construct a fuzzy connectivity function, and the three-dimensional fuzzy connection algorithm is used to iteratively connect the bronchial segments until the entire bronchial tree is reconstructed.

By implementing the present invention, a method for obtaining a three-dimensional tracheal tree from CT images of the lungs can be obtained, which reduces the influence caused by leakage and blockage during the reconstruction of the tracheal tree.

Description of drawings

Fig. 1 is a technical roadmap of a method for obtaining a three-dimensional tracheal tree from a lung CT image provided by the present invention;

Fig. 2 is a human body tracheal tree structure simulation diagram obtained by a method for obtaining a three-dimensional tracheal tree from a lung CT image provided by the present invention;

Fig. 3 is a simulation diagram of the extraction result of the main bronchus obtained by using a method for obtaining a three-dimensional tracheal tree from a CT image of the lungs provided by the present invention;

Fig. 4 is a simulation diagram of the bronchial segment extraction process obtained by using a method for obtaining a three-dimensional tracheal tree from a lung CT image provided by the present invention;

Fig. 5 is a simulation diagram of the extraction result of the bronchial segment obtained by using a method for obtaining a three-dimensional tracheal tree from a lung CT image provided by the present invention;

Fig. 6 is a schematic diagram of the result of "stitching" using a method for obtaining a three-dimensional tracheal tree from a lung CT image provided by the present invention.

detailed description

Referring to Fig. 1 and Fig. 2, the present invention provides a method for obtaining a three-dimensional tracheal tree from a lung CT image, which mainly includes the following steps: as shown by arrow 1, the main bronchus is extracted by using an adaptive three-dimensional space region growing method, as For the convenience of description, this step is referred to as step S101 for short; as shown by arrow 2 and arrow 3, use the method of optimized image feature extraction to extract other bronchial segments except the main bronchus, and this step is referred to as step S103; as shown by arrow 4 As shown, the main bronchi and bronchial segments are "stitched" using the fuzzy connectivity algorithm to obtain a three-dimensional tracheal tree as shown in Figure 2, and this step is referred to as step S105 for short.

When performing step S101, it is preferable to smooth the three-dimensional CT image first, and analyze the CT value of a certain voxel point in the CT image and the second-order arrangement structure of local brightness changes around it, and analyze the voxel Whether the point belongs to the tubular structure, the voxel points belonging to the bronchi are screened out, and all the voxel points belonging to the bronchi are collected to obtain a first-level tracheal preprocessing image. The specific method is to eliminate the influence of noise on the extracted trachea by analyzing the Hessian matrix.

Table 1 The relationship between various possible structures and the eigenvalues of the Hessian matrix in the three-dimensional case

The above table shows the relationship between various possible structures and the eigenvalues of the Hessian matrix in the three-dimensional case. λk represents the kth eigenvalue with the smallest amplitude, and the three eigenvalues λ1, λ2, and λ3 of the Hessian matrix (|λ1|≤|λ2|≤ In |λ3|), the eigenvector corresponding to the eigenvalue with the largest amplitude represents the direction of the largest curvature of a voxel point, while the eigenvector corresponding to the eigenvalue with the smallest amplitude represents the direction of the smallest curvature of a voxel point. In the CT image, the trachea is always dark, so the eigenvalue of the 3D voxel point where the trachea is located in the lung CT image should be λ1 is small, almost 0, and λ2 and λ3 are both positive numbers. Calculate the Hessian matrix of each voxel point on the CT image, and calculate its eigenvalues to determine whether it is a tracheal voxel.

After implementing the above steps, the enhanced tracheal structure has been clearly shown. However, the extracted main bronchus may still be discontinuous due to inconsistencies in the gray scale inside the trachea that may be caused by sputum or other body fluids, missed detection of the trachea caused by tracheal branches, or other local noises. Therefore, the closed operation is performed on the first-level tracheal pre-processing image, that is, the structural elements of the first-level tracheal pre-processing image are first expanded, and all background points in contact with the object are merged into the object to expand the boundary to the outside. The final The result is that the image is enlarged as a whole, and then the expanded image is corroded with structural elements to eliminate boundary points and shrink the boundary to the inside. The entire closed operation process smoothes the trachea boundary of the CT image, and obtains the second-level trachea preprocessing image.

Step S101 is performed on the basis of obtaining the secondary trachea preprocessing image, which specifically includes the following steps:

Select the initial seed point, the preferred method is manual, which can more accurately and quickly find the exact 100 voxels of the main bronchi as the initial seed point;

Select the adaptive local neighbor threshold method as the criterion for region growth to obtain the threshold value, and merge the voxel points greater than or equal to the threshold value as seed points into the seed region;

After all the surrounding seed points are merged into the seed region, the image of the main bronchi 100 as shown in FIG. 3 is obtained.

Step S103 specifically includes the following steps:

Extract several image features such as 3D pipeline feature coefficients based on multi-scale Hessian matrix, local phase, SIFT features, and low-resolution version Haar-like features. When the form of the energy item of the cost function is determined, the multi-core learning method is used. The combined kernel function of the above features is embedded into the three-dimensional bronchus seed point extraction algorithm to obtain the appropriate weights of different features, and construct the energy term in the cost function to drive the shape deformation;

Obtain the seed points that make up the bronchial segment 200;

The seed points composing the bronchial segment 200 are continuously obtained by adopting the Snake spline model method according to the continuity to obtain each independent bronchial segment 200 , as shown in FIG. 4 and FIG. 5 .

Using the image feature extraction algorithm not only avoids the "break" of the bronchi during the growth and construction process - the growth is stopped, but also avoids the "leakage" - the growth of tissues outside the lung trachea, such as alveoli and other areas. According to clinical experience, the subtle bronchial segment 200 has significant multi-scale image features in the local structure, but it cannot be guaranteed to be connected to each other, so it is necessary to perform step S105 for "stitching".

Step S105 specifically includes the following steps:

Firstly, the main bronchus 100 is "sutured" with the independent bronchial segment 200 at the adjacent end, and the specific method is to "fuzzy" connect the adjacent bronchial segments 200 based on the central line direction of the local main bronchus 100 . After the main bronchus 100 is sutured with the independent bronchial segment 200 at the adjacent end, the end of the sutured bronchial segment 200 is used to defuzzify and connect the adjacent bronchial segment 200, and then iterative calculation is used until all the bronchial segment 200 "stitched" together to obtain the three-dimensional tracheal tree shown in Figure 6.

In the above suturing steps, the centerline extraction algorithm is used to extract the centerline of the main bronchi 100, and the same method is used to extract the centerlines of each bronchial segment 200, and the spatial distance, local image features, bronchial anatomical structure and other judgments are integrated, and the structure is blurred. The connectivity function starts from the end of the main bronchus 100 and uses a three-dimensional fuzzy connection algorithm to iteratively connect the bronchial segments 200 until the entire bronchial tree is reconstructed.

The biggest advantage of "suture" is that it solves the problem that the image features of local bronchioles disappear and cannot be connected due to clinical lung lesions such as bronchial wall thinning, fluid filling, bronchial collapse, and pathological changes. Although the final local 3D tracheal results have a certain geometric ambiguity, within the clinically acceptable range, it does not affect the overall effect of bronchoscopic surgical evaluation and surgical simulation training.

In the treatment scheme of the present invention, the main bronchus and the bronchial segment 200 are firstly extracted, which effectively solves the problems that the main bronchus is easily blocked and the bronchial segment 200 is easy to leak in the traditional technology. Subsequently, the extracted main bronchi and bronchial segments 200 are combined to obtain a complete three-dimensional tracheal tree.

For those skilled in the art, without departing from the concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be based on the appended claims.

Claims (8)

1. A method for obtaining a three-dimensional trachea tree from a lung CT image, comprising the following steps: The main bronchi were extracted by an adaptive three-dimensional space region growing method; Using the method of optimized image feature extraction to extract other bronchial segments except the main bronchi; The main bronchus and the bronchial segment are "sutured" using the fuzzy connectivity algorithm to obtain a three-dimensional tracheal tree; Before the step of extracting the main bronchi, the CT image is first smoothed, and the CT value of a certain voxel point in the CT image is combined with the second-order arrangement structure of local brightness changes around it, and the voxel is analyzed. Whether the voxel points belong to the tubular structure, screen out the voxel points belonging to the bronchi, and gather all the voxel points belonging to the bronchi to obtain a first-level tracheal preprocessing image; In the step of extracting other bronchial segments except the main bronchi, specifically include the following steps: Extract several image features to construct the energy term in the cost function; Using a multi-kernel learning method, using the combined kernel function of the feature to embed into the three-dimensional bronchial seed point extraction algorithm, to obtain the seed points that make up the bronchial segment; According to the continuity, the seed points composing the bronchial segment are continuous to obtain each independent bronchial segment. 2. The method for obtaining a three-dimensional tracheal tree from a lung CT image as claimed in claim 1, wherein, before the step of extracting the main bronchi, the first-level tracheal preprocessing image is subjected to a closed operation, first Structural elements of the first-level trachea pre-processing image are expanded, and then the expanded image is corroded with structural elements to fill small holes and smooth the boundary of objects to obtain a second-level tracheal pre-processing image. 3. the method for obtaining three-dimensional tracheal tree from lung CT image as claimed in claim 1, is characterized in that, in the step of extracting main bronchus, specifically comprises the following steps: Select the starting seed point; Select the adaptive local adjacent threshold method as the criterion of region growth to obtain the threshold, and merge the voxel points greater than or equal to the threshold as seed points into the seed region; When all the surrounding seed points are merged into the seed area, the image of the main bronchi is obtained. 4. The method for obtaining a three-dimensional tracheal tree from a lung CT image as claimed in claim 1, wherein said several image features include 3D pipeline characteristic coefficients, local phases, SIFT features, and low-resolution features based on multiscale Hessian matrices. rate version of Haar-like features. 5. the method for obtaining three-dimensional trachea tree from lung CT image as claimed in claim 1, it is characterized in that, in the step that the described seed point that forms described bronchi section is carried out continuous, has adopted based on Snake spline model Methods. 6. The method for obtaining a three-dimensional tracheal tree from a lung CT image as claimed in claim 1, wherein, in the step of "sewing" the main bronchus and the bronchial segment, first the main bronchus and the corresponding bronchus are The independent bronchial segments at the adjacent end are sutured, and then iterative calculations are performed until all the bronchial segments are "sewn" together. 7. The method for obtaining a three-dimensional tracheal tree from lung CT images as claimed in claim 1, characterized in that, in the step of "stitching" the main bronchus and the bronchial segment, using a central line extraction algorithm, extracting the centerline of the main bronchus, and constructing a fuzzy connectivity function based on one or more criteria of spatial distance, local image features, and bronchial anatomical structure; Algorithm, connects the main bronchus with the bronchial segment. 8. The method for obtaining a three-dimensional tracheal tree from a lung CT image according to claim 1, characterized in that, in the step of "stitching" the main bronchus and the bronchial segment, using a central line extraction algorithm, extracting the center line of each bronchial segment, and constructing a fuzzy connectivity function based on one or more judgments of spatial distance, local image features, and bronchial anatomical structure, and using a three-dimensional fuzzy connection algorithm to iteratively connect each bronchial segment, Until the entire bronchial tree is reconstructed.