A method for automatically extracting pulmonary artery and vein from chest plain CT images

Abstract

The application provides a chest plain CT image lung artery and vein automatic extraction method, and relates to the technical field of CT image processing. The method first acquires lung airway, lung blood vessel and lung external artery and vein data, and performs data preprocessing and unifies layer thickness; then, based on the reconstructed CT scan image, lung blood vessel skeletonization is performed; and a lung internal artery and vein growth algorithm is used to classify the lung blood vessel skeleton; then, the lung artery and vein are reconstructed according to the artery and vein marked by the lung blood vessel skeleton; finally, the classified lung blood vessel is evaluated according to the blood vessel skeleton of the marked lung internal artery and vein and the blood vessel voxel point after the lung internal artery and vein reconstruction. The method uses multiple features to classify the lung internal artery and vein, does not require the imaging mode of the input data, is an automatic algorithm, does not require manual parameter adjustment or seed point setting, has strong generalization ability, and proposes two new evaluation methods, so that the classification effect of the lung artery and vein is checked and evaluated from multiple angles.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of CT image processing, and particularly relates to a chest plain CT image lung artery and vein automatic extraction method. BACKGROUND

[0002] The human lung structure is very complex, and the left and right lungs contain five lobes, which can be subdivided into multiple lung segments. The blood vessel structure in the lung is similar to a tree, with branches growing from the main trunk at the hilum. Some lung diseases can cause abnormalities in the pulmonary artery and vein, or abnormalities in the pulmonary blood vessels. For example, pulmonary hypertension can cause pulmonary artery malformation rupture; pulmonary thromboembolism can cause pulmonary artery posterior blood vessel stenosis or occlusion; and pulmonary artery and vein malformation can cause abnormal blood vessels or "blood vessel pedicle sign" connected to the hilum.

[0003] In preoperative surgical planning, distinguishing the interwoven and pathologically variant pulmonary artery and vein can provide anatomical structure information in the lung to the doctor, reducing the complexity and uncertainty of the operation. Clinically, doctors mostly manually mark the pulmonary artery and vein, however, the response of the plain CT image to the pulmonary artery and vein is close, and the complex tree structure of the pulmonary blood vessels also makes it time-consuming and laborious to distinguish the artery and vein. Therefore, a fully automatic and high-precision pulmonary artery and vein separation algorithm can reduce the workload of doctors, speed up the preoperative surgical planning time, and improve the success rate of surgery.

[0004] The pulmonary artery and vein topological structure has high similarity, and due to the partial volume effect of CT imaging, the artery and vein appear to be adhered, so fully automatic separation of the pulmonary artery and vein is a challenging problem. Even though deep learning has good performance in the medical image field, however, the pulmonary artery and vein tree structure of different patients has large difference and weak common features, and after down-pooling, the model is difficult to learn the "continuous growth of the pulmonary blood vessels" feature, and the pulmonary artery and vein label is difficult to make and time-consuming, so using deep learning is not suitable for separating the pulmonary artery and vein.

[0005] In the algorithm process of using traditional algorithms to separate the pulmonary artery and vein, the difficulties are as follows:

[0006] The pulmonary artery and vein voxel features are not obvious, and the voxel value difference between the two is not large on the CT image of high-precision imaging, so it is difficult to classify the pulmonary artery and vein from the voxel value difference. Further research should be done from the topological point of view.

[0007] The pulmonary artery and vein are intertwined and complex in shape, and it is difficult to distinguish the two using simple features. Multiple related features should be combined.

[0008] In order to improve the clinical value, a fully automatic and end-to-end algorithm should be implemented to avoid human participation or labeling.

[0009] The technical solution described in the Chinese patent "CN114240851A Lung artery and vein extraction method, device and electronic equipment based on CT image" extracts the lung trunk blood vessels through synchronous growth operation, and extracts the lung artery and vein by connecting the lung trunk blood vessels with the intrapulmonary blood vessels. First, the lung CT image to be segmented is obtained, and N target objects to be segmented are manually established on the image; then, a threshold and a seed point are manually set on each target object to be segmented; next, the lung trunk blood vessels are extracted and classified through synchronous growth; finally, the lung artery and vein are extracted by connecting the trunk blood vessels with the intrapulmonary blood vessels. This method requires multiple manual operations, and the division of the segmented region, the threshold and the seed point need to be set by professionals, and it does not have generalization ability; the intrapulmonary blood vessels are mutually adhered, and during the CT scanning process, the local arteries and veins contact each other at multiple positions, and the partial volume and volume effect further aggravate this phenomenon, and directly connecting the lung trunk blood vessels with the intrapulmonary blood vessels does not solve the accuracy problem caused by blood vessel adhesion.

[0010] The technical solution described in the Chinese patent "CN113409328A Lung artery and vein segmentation method, device, medium and equipment of CT image" has the following process: first, the chest CT image is preprocessed to obtain the lung range; then, the lung blood vessel image is obtained by threshold segmentation; for the lung blood vessel image, the image is segmented by region growing method, first the pulmonary artery is segmented, then the pulmonary vein is segmented in the lung blood vessels excluding the pulmonary artery, and finally the lung artery and vein image is obtained by multiple region growing methods based on different data. This method only relies on region growing to grow the pulmonary artery from the lung blood vessels, which may not have generalization ability, because the pulmonary artery and vein are mutually adhered, and even with high-precision imaging CT imaging technology, the relatively limited scanning resolution, the high density of the lung cavity blood vessels, the partial volume and volume effect and other factors will cause the adhesion degree of the pulmonary artery and vein to be aggravated. Although the iodine contrast agent can improve the differentiation of the artery and vein, there is no reliable method to synchronize the contrast agent dynamics, so that the pulmonary artery or vein is selectively opaque. SUMMARY

[0011] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art, and to provide a chest plain CT image lung artery and vein automatic extraction method to realize automatic extraction of the lung artery and vein in the CT image.

[0012] To solve the above technical problems, the technical solution adopted by the present application is: a chest plain CT image lung artery and vein automatic extraction method, comprising the following steps:

[0013] Step 1, obtaining lung trachea, lung blood vessels and lung external artery and vein data, and performing data preprocessing to unify the layer thickness;

[0014] Step 1.1, obtaining lung trachea, lung blood vessels and lung external artery and vein data;

[0015] A data set containing multiple groups of CT scan images is acquired, each group of CT scan images is reconstructed using a 512*512 matrix; the reconstruction method adopts a mediastinal window algorithm, and the layer thickness parameters of the reconstruction are different; each group of CT scan images contains different lung tissues and corresponding labels of the same lung, including lung parenchyma, lung trachea, lung lobe, pulmonary artery and pulmonary vein;

[0016] The lung parenchyma label and the pulmonary artery and vein label are used to perform an intersection and union operation to obtain a pulmonary artery and vein label outside the lung; the categories of the pulmonary artery and vein label are reset as unclassified pulmonary vessels;

[0017] Step 1.2, for the lung trachea, pulmonary vessels and pulmonary artery and vein outside the lung and the corresponding labels, the nearest neighbor method is used to unify the layer thickness;

[0018] Step 2, based on the reconstructed CT scan images, the pulmonary vessels are skeletonized;

[0019] Using numerical analysis software, a group of reconstructed two-dimensional CT scan images is converted into a three-dimensional image data, in which the voxel points not belonging to the pulmonary vessels are defined as background points, the surface voxel points of the pulmonary vessels in contact with the background points are defined as boundary points, and the voxel points belonging to the pulmonary vessel category are defined as foreground points;

[0020] In the process of skeletonizing the pulmonary vessels, all the pulmonary vessel voxel points in the three-dimensional image data are traversed, and the category of the current voxel point is judged, and different choices are made: for a certain voxel point, if it is a background point, the next voxel point is scanned; otherwise, it is judged whether it is a boundary point; if it is a boundary point, the number of foreground points in the 26-neighborhood of the boundary point is counted, if there is only one foreground point, the boundary point cannot be deleted, otherwise the Euler invariance of the boundary point is continued to be judged; if the value of the Euler characteristic is unchanged, it is continued to be judged whether the boundary point is a simple point, and the simple point is judged by judging whether removing the point will produce a cavity or a hole, if it is not a simple point and the boundary point is not marked, the boundary point is added to the queue and marked; then the boundary points in the queue are judged whether they are simple boundary points, if they are, they are deleted; if not, they are retained, and the simple boundary points are deleted until the queue is empty.

[0021] The data structure after the pulmonary vessel skeletonization is described as follows: for each pulmonary vessel skeleton bifurcation node object , its attributes include identity number , three-dimensional coordinates , identity numbers of all connected edges , whether it is a leaf node; for each pulmonary vessel skeleton edge object , its attributes include identity number j , two end points , the discrete point set constituting the side ;

[0022] Step 3, using the pulmonary artery and vein growth algorithm to classify the pulmonary vessel skeleton;

[0023] Step 3.1, determine the pulmonary artery and vein prior seed point;

[0024] An inflation algorithm is used to fill the gap between the pulmonary artery and vein label and the pulmonary vessel label, and the inflation scale is not more than 20 voxel points to maintain the connectivity between the pulmonary vessels; the determination process of the prior seed point is: superimpose the pulmonary artery and vein and the pulmonary vessel image together, obtain the joint part of the two at the hilum, and set the voxel points in the joint part as the pulmonary artery and vein prior seed point, and mark the prior seed point category according to the label of the pulmonary artery and vein;

[0025] Step 3.2, use the Euclidean space vector feature to constrain the multi-branch node caused by the adhesion point in the pulmonary vessel skeleton, so as to weight the correct pulmonary vessel skeleton growth direction;

[0026] All adjacent edges of the multi-branch node caused by the adhesion point in the pulmonary vessel skeleton are traversed, and the direction of the adjacent edge with the largest angle with the current pulmonary vessel skeleton growth direction is selected as the optimal growth direction;

[0027] In order to avoid false growth, the cosine value between the optimal growth direction and the current pulmonary vessel skeleton growth direction is less than a set threshold C; when calculating the angle, the growth direction starts from the multi-branch node, and the cosine value is used to evaluate whether the growth direction is consistent;

[0028] For the current pulmonary vessel skeleton growth direction vector , all adjacent edge vectors of the multi-branch node are , then the weight of the optional growth direction is calculated as shown in the following formula:

[0029] ;

[0030] Wherein, N is the total number of all adjacent edges of the multi-branch endpoint;

[0031] Step 3.3, pulmonary bronchus skeletonization, using pulmonary bronchus guidance and distance features to weight and constrain the correct pulmonary vessel skeleton growth direction;

[0032] Using the pulmonary vessel skeletonization method to skeletonize the pulmonary bronchus; and using the Euclidean space vector angle and space distance to realize the pulmonary bronchus guidance and distance constraint in a weighted manner; wherein, the distance feature between the pulmonary bronchus and the pulmonary vessel is a thresholdD Subtract the Euclidean distance between them The evaluation criterion of Euclidean distance is the lung airway The midpoint of the skeletal structure to the lung blood vessel The midpoint of the skeletal structure Euclidean distance; the closer the Euclidean distance between them, the greater the distance feature between them, and if the Euclidean distance between them exceeds the threshold, the distance feature is set to 0; the lung airway The vector feature of the lung blood vessel is the inverse of the sine value between the two skeletal structures; the vector is the direction vector between the two end points of the lung blood vessel edge; the direction vector of the lung airway The closer the growth direction of the lung airway and , the smaller the sine value, and the greater the feature value; the overall weight feature value of the lung airway to the artery is obtained according to the orientation feature and the distance feature of the lung airway , as shown in the following formula:

[0033] ;

[0034] Wherein, L is the threshold of the orientation feature;

[0035] Step 3.4, combined with the lung blood vessel skeleton, the sphere marching algorithm is used to grow the lung blood vessel;

[0036] Combined with the lung blood vessel skeleton as the growth guide, a sphere inscribed in the lung blood vessel wall marches continuously and marks the lung blood vessel skeleton passing through part; the sphere starts from the prior seed point, and in the marching process, the sphere adjusts its radius according to the lung blood vessel thickness to maintain the inscribed lung blood vessel wall; once the sphere grows to the adhesion point position, i.e. the sphere encounters a multi-branch node in the marching process, the sphere will preferentially select the road that keeps the change amplitude of the sphere volume minimum and still can inscribe the lung blood vessel wall; the lung airway guide and distance constraint and the Euclidean space vector constraint are used to guide the marching direction of the sphere in the sphere marching process;

[0037] Specifically, when the sphere is at a multi-branch node, the lung blood vessel radius of the sphere at the current position is , the remaining blood vessel radius of the multi-branch node is , the lung blood vessel radius is used as an index to judge the change amplitude of the sphere, and the Euclidean space vector constraint feature and the lung airway guide and distance constraint feature are used to finally obtain the weight between the current edge of the sphere and the other adjacent edges of the multi-branch node, as shown in the following formula:

[0038] ;

[0039] wherein, , , are weight coefficients; is the gradient difference of the lung vessel thickness between the current lung vessel thickness and the lung vessel thickness to be grown, is the gradient difference of the sphere travel distance, represents the change gradient of the lung vessel wall thickness during the sphere travel process;

[0040] After combining the Euclidean space vector features, the lung airway guidance and distance constraint features, the sphere selects a growth direction and continues to advance until the lung vessel grows to the end node;

[0041] Step 3.5, using a voting decision method to classify the unclassified lung vessel skeleton;

[0042] For the lung vessel skeleton without a determined class, the adjacent edges of the skeleton are obtained according to its data structure, and all the classified adjacent edges of the lung vessel skeleton will evaluate the lung vessel skeleton, and the evaluation indexes include the Euclidean space distance, the Euclidean space vector, the lung vessel thickness deviation and the lung vessel thickness gradient change difference. Each classified adjacent edge will calculate the sum of the four evaluation index results between the current unclassified lung vessel skeleton , When the value of the sum is greater than 0, it means that the features of the two are the same class; after the evaluation of all the adjacent edges, the sum is obtained If the lung vessel skeleton belongs to an artery, the sum is positive, otherwise it is negative; according to the positive and negative of the sum, the final class of the lung vessel skeleton belongs to an artery or a vein, as shown in the following formula:

[0043] ;

[0044] ;

[0045] Finally, all the lung vessel skeletons are marked as two classes of arteries and veins;

[0046] Step 4, reconstructing the pulmonary artery and vein according to the marked arteries and veins of the lung vessel skeleton;

[0047] Using a preemptive breadth-first search algorithm, all the nodes and their corresponding classes that make up the lung vessel skeleton are put into a queue and marked, then the nodes are constantly removed from the queue, and when a node is removed, its corresponding voxel point in the three-dimensional image data is assigned to the corresponding class, and all the unmarked adjacent points of the voxel point are put into the queue with the corresponding class and marked; until all the voxel points in the queue are removed from the queue, the pulmonary artery and vein reconstruction is completed.

[0048] Step 5, evaluate the classified lung vessels according to the labeled pulmonary artery and vein vessel skeleton and the vessel voxel points after pulmonary artery and vein reconstruction;

[0049] (1) Distance weighted evaluation based on the distance of extrapulmonary vessels;

[0050] First, traverse the entire lung vessel data voxel points, and obtain the center voxel point and the farthest voxel point of the entire lung vessel data, thereby obtaining the overall radius r of the lung vessel data. For each lung vessel skeleton, select the point in the middle of the discrete point set index, that is, the middle point of the skeleton, obtain its three-dimensional coordinates, and obtain the distance d between the point and the center point. For the middle point of the lung vessel skeleton, verify whether the category of the point is consistent with the predicted category from the gold standard data. If consistent, it is considered that the lung vessel skeleton is correctly predicted, otherwise it is considered that the lung vessel skeleton is incorrectly predicted, whether correctly predicted is denoted as w i , and further obtain the distance weighted evaluation result based on the distance of extrapulmonary vessels , as shown in the following formula:

[0051] ;

[0052] ;

[0053] The result value closer to 1, the better the lung vessel classification effect; the result value closer to 0, the worse the classification effect;

[0054] (2) Growth level grouping evaluation:

[0055] Set the lung vessel skeleton close to the hilum as the first level, and each time the new edge is grown, the level of the edge will be increased by one level, and the level of the lung vessel skeleton is denoted as a i The set composed of lung vessel skeletons with level a is denoted as Every 4 levels of growth will group these edges, which is specifically from level a to a+3 levels, and the classification accuracy of each group is evaluated, as shown in the following formula:

[0056] ;

[0057] wherein, denotes whether a lung vessel skeleton is correctly predicted;

[0058] The result value closer to 1, the better the classification effect of the group of lung vessels; the result value closer to 0, the worse the classification effect of the group of lung vessels.

[0059] The beneficial effects produced by the above technical scheme are that the chest plain CT image pulmonary artery and vein automatic extraction method provided by the application has the following advantages: (1) a plurality of features are used for classifying the pulmonary artery and vein; the pulmonary artery and vein classification is realized by using the pulmonary airway guidance and distance feature, the Euclidean space vector feature and the pulmonary artery and vein feature outside the lung; the method is tested on a public data set, and good results are obtained, which proves that the method realizes the pulmonary artery and vein classification function and has clinical value; (2) the imaging mode of the input data is not required; the method mainly studies the separation of the pulmonary artery and vein, and only the pulmonary blood vessel, the pulmonary airway and the pulmonary artery and vein data outside the lung are required to separate the pulmonary artery and vein; (3) it is an automatic algorithm, and no manual parameter adjustment or seed point setting is required; the method starts from the anatomical structure characteristics of the pulmonary blood vessel, uses the pulmonary artery and vein outside the lung as the prior seed point of the pulmonary artery and vein inside the lung, uses the Euclidean space vector constraint feature, the spherical ball advancing algorithm combining the pulmonary airway guidance and distance constraint feature, ensures that the artery and vein growth algorithm can grow correctly at the adhesion position, finally uses the voting decision method to further constrain the classification result, and then separates the pulmonary artery and vein, and realizes the end-to-end output; (4) the generalization ability is strong; the method uses obvious features, utilizes the pulmonary airway guidance constraint, the pulmonary blood vessel thickness change and other features, although the pulmonary blood vessel structure is different for different people, but all conforms to these features; the method proposes a plurality of complex feature extraction formulas, and describes the pulmonary blood vessel structure features which are difficult to describe; (5) two new evaluation methods are proposed, so that the classification effect of the pulmonary artery and vein is checked and evaluated from multiple angles; the distance weighted evaluation method considers that the pulmonary blood vessel near the hilum has a larger thickness and more small blood vessel branches, so the weight should account for a higher part in the evaluation. The growth level grouping evaluation method classifies and evaluates the pulmonary blood vessel, sets the blood vessel skeleton near the hilum as the first level, the level of the new edge to which the growth is performed each time is improved by one level, and the growth accuracy of the algorithm can be effectively tested. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 The flowchart of the chest plain CT image pulmonary artery and vein automatic extraction method provided for the embodiment of the application;

[0061] Figure 2 The CT scan image containing different lung tissues provided for the embodiment of the application, wherein (a) is the pulmonary airway data, (b) is the pulmonary blood vessel data of a unified category, and (c) is the pulmonary artery and vein data outside the lung;

[0062] Figure 3 The pulmonary blood vessel skeletonization result provided for the embodiment of the application, wherein (a) is the pulmonary blood vessel skeleton display, (b) is the data structure of the pulmonary blood vessel skeleton bifurcation node, and (c) is the data structure of the pulmonary blood vessel skeleton edge;

[0063] Figure 4 Pulmonary vessel adhesion part and corresponding skeletonization results provided by embodiments of the present application, wherein (a) is a pulmonary vessel with a heavy adhesion degree, (b) is a pulmonary vessel skeleton corresponding to the pulmonary vessel part with a heavy adhesion degree, (c) is a pulmonary vessel with a light adhesion degree, and (d) is a pulmonary vessel skeleton corresponding to the pulmonary vessel part with a light adhesion degree;

[0064] Figure 5 Flowchart of a pulmonary artery and vein growth algorithm provided by embodiments of the present application

[0065] Figure 6 Processing method of a pulmonary vessel skeleton provided by embodiments of the present application, wherein (a) is a pulmonary artery and vein after gap filling, (b) is a pulmonary artery and vein superimposed with a vessel skeleton, (c) is a pulmonary vessel skeleton after a given artery and vein growth point, and (d) is a pulmonary vessel skeleton after straightening;

[0066] Figure 7 Information provided by a pulmonary bronchus in pulmonary vessel classification provided by embodiments of the present application, wherein (a) is a pulmonary artery and pulmonary vessel structure, (b) is a pulmonary artery and pulmonary vessel skeleton, (c) is pulmonary vessel topological similarity, and (d) is pulmonary skeleton topological similarity;

[0067] Figure 8 Pulmonary vessel skeleton classification method and classification result provided by embodiments of the present application, wherein (a) is a sphere traveling in a pulmonary vessel, (b) is a pulmonary vessel skeleton after classification, and (c) is a reconstructed pulmonary artery and vein;

[0068] Figure 9 Vessel hierarchical diagram provided by embodiments of the present application;

[0069] Figure 10 Visualization result of pulmonary vessel classification provided by embodiments of the present application, wherein (a) is a pulmonary vessel skeleton classification result, (b) is a pulmonary vessel skeleton gold standard, (c) is a pulmonary vessel reconstruction result, and (d) is a pulmonary vessel gold standard;

[0070] Figure 11 Comparison result diagram of pulmonary artery and vein classification of different algorithms provided by embodiments of the present application. DETAILED DESCRIPTION

[0071] The specific embodiments of the present application are described in further detail below in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0072] In this embodiment, a chest plain CT image pulmonary artery and vein automatic extraction method is provided, as shown in FIG. 1. Figure 1As shown, comprising the following steps:

[0073] Step 1, acquire lung airway, lung blood vessel and pulmonary extraparenchymal vein data, and perform data preprocessing, unify the layer thickness;

[0074] Step 1.1, acquire lung airway, lung blood vessel and pulmonary extraparenchymal vein data;

[0075] Obtain a data set containing multiple groups of CT scan images, each group of CT scan image is reconstructed using a 512*512 matrix; the reconstruction method uses a mediastinal window algorithm, and the layer thickness parameters of the reconstruction are different; each group of CT scan images contains different lung tissues and corresponding labels of the same lung, including lung parenchyma, lung airway, lung lobe, pulmonary artery and pulmonary vein;

[0076] Use the lung parenchyma label and the lung artery and vein label to do intersection and union operation to get the pulmonary extraparenchymal vein label; reset the category of the lung artery and vein label as unclassified lung blood vessel; at this time, the three kinds of known data required for CT image lung artery and vein have been prepared, as shown in Figure 2 .

[0077] Step 1.2, in the reconstruction process of each group of CT image data set, the layer thickness parameters are different, containing three different thicknesses: 0.75mm, 1.00mm, 1.50mm, since the algorithm uses the lung blood vessel thickness feature, and the data set format is a three-dimensional matrix, considering the actual imaging real size, and the layer thickness and physical size of different patients are different, therefore, it needs to be unified to the same standard layer thickness. For lung airway, lung blood vessel and pulmonary extraparenchymal vein and corresponding label, the nearest neighbor method is used to unify the layer thickness;

[0078] Step 2, based on the reconstructed CT scan image, the lung blood vessel is skeletonized;

[0079] The main purpose of extracting the lung blood vessel skeleton is to reduce the number of voxel points as much as possible under the premise of ensuring that the blood vessel topological features are not lost, so as to reduce the difficulty of distinguishing lung artery and vein. In three-dimensional Euclidean space, the skeleton of a geometric body is the motion trajectory of all its maximum inscribed sphere centers. These spheres are tangent to the geometric body boundary at multiple points. It can be regarded as a simplified feature of a geometric body, which is used to reduce the feature search space of the geometric model.

[0080] Using numerical analysis software, a group of reconstructed two-dimensional CT scan images is converted into a three-dimensional image data, in the three-dimensional image data, the voxel points not belonging to the lung blood vessel are defined as background points, the lung blood vessel surface voxel points in contact with the background points are defined as boundary points, and the voxel points belonging to the lung blood vessel category are defined as foreground points;

[0081] In the process of lung vessel skeletonization, all lung vessel voxels in the three-dimensional image data are traversed, and the category of the current voxel is determined, and then different choices are made: for a voxel, if it is a background point, the next voxel is scanned; otherwise, it is determined whether it is a boundary point; if it is a boundary point, the number of foreground points in the 26-neighborhood of the boundary point is counted, if the foreground point is only one, the boundary point cannot be deleted, otherwise the Euler invariance of the boundary point is continued to be determined; if the value of the Euler characteristic is unchanged, it is continued to be determined whether the boundary point is a simple point, and the determination of the simple point is to determine whether the removal of the point will not produce a cavity or a hole, if it is not a simple point and the boundary point is not marked, the boundary point is added to the queue and marked; then the boundary point in the queue is determined whether it is a simple boundary point, if it is, it is deleted; if not, it is retained, and the simple boundary point is deleted until the queue is empty. The pseudo code of the above lung vessel skeletonization process is shown in Table 1. The structure of the data labeled skeletonization is shown in Figure 3 (a).

[0082] Table 1 lung vessel skeletonization algorithm

[0083]

[0084] In this embodiment, the data structure after lung vessel skeletonization is described as follows: for each bifurcation node object of the lung vessel skeleton , its attributes include identity number , three-dimensional coordinates , identity numbers of all connected edges , and whether it is a leaf node; for each edge object of the lung vessel skeleton , its attributes include identity number j , two end points it has , and discrete point set constituting the edge . Figure 3 The node structure and edge structure in the lung vessel skeletonization structure are respectively shown in Figure 3 (b) and (c), which are components in (a).

[0085] Step 3, using the pulmonary artery and vein growth algorithm to classify the lung vessel skeleton;

[0086] The adhesion problem of the pulmonary artery and vein is the biggest difficulty to separate them, and the adhesion point is also a big trouble for the method of the present application. The comparison between different adhesion degrees is shown in Figure 4 , the node with complex adhesion degree is shown in Figure 4 (a), Figure 4 (b) is the skeleton structure corresponding to the node, Figure 4 (c) shows the node with simple adhesion degree, Figure 4The middle (d) is the skeleton structure corresponding to the node. In combination Figure 4 It can be seen that even if the structure after extracting the lung vessel skeleton, the adhesion problem still exists. For the challenging lung vessel adhesion part growth problem, the pulmonary artery and vein growth algorithm uses multiple weights to determine the growth direction. The pulmonary artery and vein growth algorithm flow chart is shown in Figure 5 .

[0087] Step 3.1, determine the pulmonary artery and vein prior seed point;

[0088] In fact, the pulmonary artery and vein label and the pulmonary vessel label may not be completely connected, and there are some breaks and gaps, resulting in the absence of some seed points in the process of determining the prior growth seed point. For this problem, the invention uses an inflation algorithm to fill the gap between the pulmonary artery and vein label and the pulmonary vessel label, and the inflation scale is not more than 20 voxel points, so as to maintain the connectivity between the pulmonary vessels, as shown in Figure 6 (a) shows the pulmonary artery and vein label without inflation, and Figure 2 (c) shows the pulmonary artery and vein label without inflation; the determination process of the prior seed point is: superimpose the pulmonary artery and vein and the pulmonary vessel image together, obtain the connected part of the two at the pulmonary hilum, the voxel points in the connected part are set as the pulmonary artery and vein prior seed point, and the prior seed point category is marked according to the label of the pulmonary artery and vein, as shown in Figure 6 (b) and (c);

[0089] Step 3.2, use the Euclidean space vector feature to constrain the multiple branch nodes caused by the adhesion points in the pulmonary vessel skeleton, so as to weight the correct pulmonary vessel skeleton growth direction;

[0090] From the physiological structure, the pulmonary artery and vein belong to the nonlinear tubular structure, which determines that the gradient trend of the pulmonary vessel growth will not appear greatly. For the skeleton structure of the adhesion part of the pulmonary vessel, an adhesion point has at least 4 edges, and the Euclidean space vector can effectively provide the correct growth direction. From Figure 5 , it can be seen that the pulmonary vessel skeleton is not a straight line, but a zigzag and continuous one, so it is difficult to obtain its vector information. For this purpose, the two end points of each pulmonary vessel skeleton are retained, and other discrete points are discarded, and the pulmonary vessel skeleton is straightened with the two end points, so as to obtain the Euclidean space vector of the pulmonary vessel skeleton, as shown in Figure 6 (d).

[0091] All adjacent edges of the multiple branch nodes caused by the adhesion points in the pulmonary vessel skeleton are traversed, and the direction of the adjacent edge with the largest angle with the current pulmonary vessel skeleton growth direction is selected as the optimal growth direction;

[0092] To avoid erroneous growth, the cosine value between the highest priority growth direction and the current pulmonary vascular bone growth direction must be less than a set threshold C. When calculating the angle, the starting point of each growth direction is a multi-branch node, and the cosine value is used to evaluate whether the growth direction is consistent. This is because the larger the angle formed by the current pulmonary vascular bone growth direction and other adjacent edge growth directions, the smaller the cosine value. Furthermore, because a threshold is set, the cosine value of the angle between a set of valid growth directions must be negative.

[0093] For the current pulmonary vascular bone growth direction vector The vectors of all adjacent edges of its multi-way nodes are Then the weights of the selectable growth directions The calculation is shown in the following formula:

[0094] ;

[0095] Where N is the total number of all adjacent edges of the multi-branch endpoint;

[0096] Step 3.3: Skeletalize the pulmonary trachea and use pulmonary trachea guidance and distance features to apply weighted constraints on the correct pulmonary vascular skeleton growth direction;

[0097] The pulmonary artery carries venous blood containing carbon dioxide. To facilitate gas exchange, the pulmonary artery shares certain topological similarities with the trachea, such as... Figure 7 As shown in (a) and (b), this feature becomes more apparent after bone extraction, such as Figure 7 As shown in (c) and (d), the distance between the two is relatively short. However, for complex and intertwined pulmonary arteries and veins, distance alone is not sufficient as a strong criterion for distinguishing them. Experiments have shown that the trachea alone is insufficient to distinguish pulmonary arteries and veins, but the topological features of the trachea can be helpful. For pulmonary arteries and veins that are difficult to distinguish, their classification can be determined or corrected based on whether the vessel has a certain topological similarity to the trachea.

[0098] The pulmonary trachea was skeletonized using a pulmonary vascular skeletonization method; and pulmonary trachea guidance and distance constraints were achieved using a weighted approach using Euclidean space vector angles and spatial distances; among which, the pulmonary trachea With pulmonary vessels Distance features Threshold D Subtract the Euclidean distance between the two The criterion for judging Euclidean distance is the trachea. Midpoint of skeletal structure to pulmonary vessels Midpoint of skeletal structure The Euclidean distance between two points; the closer the Euclidean distance between them, the larger their distance feature. If the Euclidean distance between them exceeds a threshold, the distance feature is set to 0; trachea With pulmonary vessels vector features The reciprocal of the sine of the skeletal structure between the two; vector The direction vector between the two endpoints of the pulmonary vessels; the direction vector of the pulmonary trachea. and The closer the growth directions, the smaller the sine value and the larger the eigenvalue; the overall weight eigenvalue of the pulmonary trachea to the artery is obtained based on the guidance and distance characteristics of the pulmonary trachea. As shown in the formula below:

[0099]

[0100] Where L is the threshold of the guiding feature, used to prevent the sine value in the denominator from being too small;

[0101] Step 3.4: Combine the pulmonary vascular skeleton with the spherical movement algorithm to grow pulmonary blood vessels;

[0102] Using the pulmonary vascular skeleton as a growth guide, a sphere tangential to the pulmonary vascular wall continuously travels, marking the pulmonary vascular skeleton it passes through, with the following effect: Figure 8 As shown, the sphere starts from the prior seed point and, during its movement, adjusts its radius according to the thickness of the pulmonary vessels, maintaining tangency within the pulmonary vessel walls. Once the sphere grows to the adhesion point (i.e., encountering a multi-branch node during its movement), it prioritizes selecting the path that minimizes its volume change while still tangent to the pulmonary vessel walls. However, for the terminal branches of the pulmonary vascular skeleton, the difference in vessel thickness between the pulmonary arteries and veins is not significant. The sphere's movement is guided by pulmonary tracheal guidance and distance constraints, as well as Euclidean space vector constraints, achieving a highly accurate growth algorithm.

[0103] Specifically, when the sphere is at a multi-branch node, the radius of the pulmonary vessels at the current location of the sphere is... The radius of the remaining pulmonary vessels at the multi-branched node is Using the radius of lung vessels as an indicator of the sphere's variation, the algorithm aims to guide the sphere to grow in directions where the gradient change with lung vessel growth is minimal. Based on this objective, this invention proposes a sphere movement algorithm and uses Euclidean space vector constraints on features. and tracheal guidance and distance constraint characteristics Finally, the weights between the edge where the sphere is currently located and the other adjacent edges of the multi-way node are obtained. The formula is shown below:

[0104]

[0105] wherein, , , are weight coefficients, which need to be selected in combination with the selected uniformized layer thickness parameter; is the gradient difference of the thickness of the lung vessel to be grown and the current lung vessel thickness, is the gradient difference of the distance traveled by the sphere, represents the change gradient of the lung vessel wall thickness during the sphere travel.

[0106] After combining the Euclidean space vector features, the lung airway guidance and distance constraint features, the sphere selects a growth direction and continues to travel until the lung vessel grows to the end node;

[0107] Step 3.5, using a voting decision method to classify the unclassified lung vessel skeleton;

[0108] After the sphere travel algorithm, there may still be a part of the lung vessel skeleton that has not been grown, because their topological features have not been selected by the growth algorithm of the artery or vein. For this, the present application proposes a voting decision method. For the lung vessel skeleton without a determined class, the adjacent edges of the skeleton are obtained according to the data structure, and all the classified adjacent edges of the lung vessel skeleton will evaluate the lung vessel skeleton. The evaluation indicators include the Euclidean distance, the Euclidean vector, the lung vessel thickness deviation and the vessel thickness change gradient difference; each classified adjacent edge calculates the sum of the four evaluation indicators between the current unclassified lung vessel skeleton , When the value of the sum is greater than 0, it means that the features of the two are the same class; after the evaluation of all the adjacent edges, the sum is calculated , if the lung vessel skeleton belongs to an artery, the sum is positive, otherwise it is negative; according to the positive and negative of the sum, the final class of the lung vessel skeleton is determined to be an artery or a vein, as shown in the following formula:

[0109]

[0110]

[0111] Finally, all the lung vessel skeletons are marked as two classes of arteries and veins, as shown in Figure 8 b.

[0112] Step 4, reconstructing the pulmonary artery and vein according to the marked arteries and veins of the lung vessel skeleton;

[0113] Since each edge object in the data structure of the skeleton is used to compose a discrete point set, the pulmonary vessel skeleton can be regarded as a line composed of a large number of point sets. The present application adopts a preemptive breadth-first search algorithm, puts all the nodes and their corresponding categories that constitute the pulmonary vessel skeleton into a queue, and marks them, then constantly removes the nodes from the queue, assigns the corresponding category to the voxel point corresponding to each removed node in the three-dimensional image data, and puts all the unmarked neighborhood point sets of the voxel point and the corresponding category into the queue and marks them; until all the voxel points in the queue are removed from the queue, the pulmonary artery and vein reconstruction is completed. Using this method, the present application grows the three-dimensional image data of the pulmonary vessels, the growth step is 1, and the growth direction of each voxel is 8-neighborhood. The reason why 26-neighborhood is not selected is that the pulmonary vessels are continuous and dense, the growth direction effect of 8-neighborhood is consistent with that of 26-neighborhood, and less time and space are consumed, and the pseudocode is shown in Table 2. The reconstructed pulmonary vessel result is shown in Fig. (c) of the following Fig. 1. Figure 8

[0114] Table 2 Reconstruction of pulmonary artery and vein process

[0115]

[0116] The present application is based on the following software and hardware environment to run the method of the present application: the hardware environment includes Intel tenth generation i7 CPU, Nvidia GTX 2060Ti GPU, memory capacity of 16.00 GB, and hard disk capacity of 1 TB. The software environment includes operating system Windows 10, and algorithm development languages Matlab and Python. In the above environment, the time consumption of the method of the present application is averagely 5 to 8 minutes, which meets the requirements of preoperative planning.

[0117] The present embodiment creates and uses a high-quality labeled lung medical image dataset, which is annotated by the software DeepInsight of the Ministry of Education Key Laboratory of Intelligent Computing for Medical Imaging and is corrected manually. A number of researchers in the field of medical imaging are responsible for annotation verification. When the annotation results are inconsistent, the results are uniformly annotated through joint discussion, re-judgment and professional knowledge such as anatomical structure and imaging characteristics. The dataset contains 24 groups of CT scan images, all of which are reconstructed using a 512*512 matrix. The reconstruction method adopts a mediastinal window algorithm, and the reconstruction parameters are different, as shown in Table 3. Each group of data set contains different lung tissues of the same lung, including lung parenchyma, pulmonary artery, pulmonary vein, and pulmonary vein.

[0118] Table 3 Information of lung medical image dataset

[0119] Dataset Number of slices Layer thickness / mm 1 576 0.75 2 541 1 3 371 1 4 466 1 5 378 1 6 427 1 7 216 1.5 8 224 1.5 9 352 1.5 10 234 1.5 11 208 1.5 12 212 1.5 13 219 1.5 14 210 1.5 15 195 1.5 16 214 1.5 17 192 1.5 18 200 1.5 19 211 1.5 20 212 1.5 21 256 1.5 22 226 1.5 23 204 1.5 24 212 1.5

[0120] ​Step 5, according to the labeled pulmonary artery and vein vascular skeleton and the blood vessel voxel point of the pulmonary artery and vein reconstruction, the classified pulmonary blood vessels are evaluated;

[0121] (1) Distance weighted evaluation based on the distance of the blood vessels outside the lung; the embodiment considers that the lung blood vessels near the hilum have a larger thickness and more small blood vessel branches, so their weight should account for a higher proportion in the evaluation;

[0122] First, the entire lung blood vessel data voxel points are traversed, and the center voxel point and the farthest voxel point of the entire lung blood vessel data are obtained, thereby obtaining the overall radius r of the lung blood vessel data. For each lung blood vessel skeleton, the point in the middle of the discrete point set index, i.e., the middle point of the skeleton, is selected, and its three-dimensional coordinates are obtained to obtain the distance d between the point and the center point. For the middle point of the lung blood vessel skeleton, it is verified from the gold standard data whether the category of the point is consistent with the predicted category. If consistent, the lung blood vessel skeleton is considered to be correctly predicted, otherwise the lung blood vessel skeleton is considered to be incorrectly predicted, whether correctly predicted is represented by , and the distance weighted evaluation result based on the distance of the blood vessels outside the lung is obtained , as shown in the following formula:

[0123] ;

[0124] ;

[0125] The closer the result value is to 1, the better the lung blood vessel classification effect is; the closer the result value is to 0, the worse the classification effect is.

[0126] (2) Growth level grouping evaluation:

[0127] The embodiment makes a hierarchical evaluation of the lung blood vessels, and sets the lung blood vessel skeleton near the hilum as the first level. Each time the new edge is grown, the level of the edge will be increased by one level, and the lung blood vessel skeleton with a level of i forms a set , and the hierarchical schematic diagram is shown in Figure 9 . Every 4 levels are grown, and these edges are grouped, which is specifically shown as follows: starting from level a, growing for 4 levels, and ending at level a+3, and the classification accuracy of each group is evaluated to observe at which level the growth algorithm is more likely to grow errors, as shown in the following formula:

[0128] ;

[0129] wherein, represents whether a lung blood vessel skeleton is correctly predicted;

[0130] The result value The closer to 1, the better the classification effect of the group of pulmonary vessels; the closer to 0, the worse the classification effect of the group of pulmonary vessels.

[0131] In addition to the evaluation method proposed in this embodiment, the existing evaluation index of pulmonary artery-vein separation is also used in this embodiment. As for the evaluation method of the pulmonary vessel skeleton, the skeleton-recall is used in this embodiment, the pulmonary vessel skeleton is extracted from the gold standard, and the proportion of the correctly predicted pulmonary vessel skeleton is judged. The number of the classified pulmonary artery-vein vessel skeleton is counted in this embodiment, the accuracy of the prediction result is evaluated, and the evaluation criterion is defined as shown in the following formula.

[0132] ;

[0133] As for the evaluation method of the pulmonary vessel voxel, the voxel-recall, the voxel-precision and the voxel-accuracy are used in this embodiment, as shown in the following formula, wherein the positive class sample is considered as the artery, and the negative class sample is considered as the vein. Only the pulmonary vessel voxel is used in the evaluation process, and the zero value voxel which is not the pulmonary vessel is not included. The pulmonary vessel reconstruction may cause the voxel of the joint part of the artery and the vein to be grown incorrectly, so the evaluation method based on the pulmonary vessel voxel may have noise interference, and the result is not as accurate as the evaluation method based on the pulmonary vessel skeleton, but this evaluation method is the most strict and objective evaluation method for the final result.

[0134] ;

[0135] ;

[0136] ;

[0137] In this embodiment, the evaluation results based on the lung vessel skeleton are shown in Tables 4 and 5. Table 4 records the distance-weighted evaluation results based on the distance of the blood vessels outside the lung, and the average value is 94.04±5.20%. This evaluation index has a large variance, but it shows good results on most data. Table 4 records the edge-recall coefficient between the pulmonary artery and vein skeleton and the gold standard, and the average value of the evaluation results is 97.27±2.53%, which is the highest among all evaluation indexes. At the same time, the number of artery and vein vessel skeleton edges in the prediction process is more than 200 skeleton edges of the gold standard on average. This shows that the adhesion of arteries and veins leads to more skeleton structures extracted, and the prediction ability is also enhanced, which is the reason for this index being high. Table 5 records the growth level grouping evaluation results, and the results show that the average value of the first group (1-4 level) is 92.23±4.22%, the average value of the second group (5-8 level) is 87.86±9.16%, and the average value of the third group (9-12 level) is 83.29±15.32%. The number of samples in the remaining groups is too small to be recorded in the table. From the data, it can be seen that as the number of groups increases, the skeleton separation accuracy gradually decreases, and the variance gradually increases, which conforms to the rule that the proportion of errors in the growth process will become larger and larger. In this embodiment, the correl formula shown in the following formula is used to perform correlation analysis on the two evaluation results, and the correlation coefficient is 0.86, indicating that they are positively correlated.

[0138] ;

[0139] Table 4 Evaluation results based on distance-weighted evaluation of blood vessels outside the lung and evaluation results based on the gold standard being correctly predicted

[0140] Dataset Distance-weighted assessment (%) Predicted number of arteries Predicted number of veins Gold standard number of arteries Gold standard number of veins Skeletal-recall (%) 1 96.75 727 841 595 834 98.14 2 92.44 857 1185 1101 894 95.82 3 89.94 1496 1370 1186 1113 93.71 4 94.16 892 1062 815 865 97.96 5 95.24 859 707 643 537 98.18 6 95.87 1932 1950 1751 2014 98.34 7 97.95 1600 1622 1286 1287 99.40 8 97.12 673 844 717 789 98.96 9 97.95 672 752 570 589 98.45 10 97.72 1081 897 778 697 97.49 11 90.50 1275 1163 982 860 96.12 12 95.40 704 684 571 490 98.28 13 95.00 799 1071 788 992 96.39 14 94.97 826 813 626 632 97.95 15 96.25 958 645 736 606 98.11 16 90.16 1542 1301 1548 1145 95.93 17 91.19 1046 866 740 761 93.25 18 99.24 694 765 486 558 99.64 19 93.09 810 961 654 684 98.41 20 94.60 806 1152 738 880 97.98 21 90.38 1639 1995 1535 1300 97.37 22 85.12 2400 2171 2494 1756 91.23 23 95.62 541 530 488 400 98.91 24 90.39 1642 2166 2052 1711 94.69 mean 94.04±5.20 1103 1146 995 933 97.11±2.53

[0141] Table 5 Evaluation results based on lung vessel classification

[0142] Dataset First group accuracy (%) Second group accuracy (%) Third group accuracy (%) 1 91.56 87.75 89.59 2 90.43 83.73 77.00 3 87.18 84.03 75.66 4 91.47 85.64 79.39 5 93.50 89.80 83.95 6 98.11 87.62 88.36 7 89.77 95.87 97.00 8 89.47 96.42 95.87 9 91.67 97.2 94.61 10 96.46 93.67 93.56 11 90.52 89.15 75.86 12 88.98 85.71 84.16 13 87.43 87.07 82.04 14 94.74 92.40 83.79 15 96.45 92.73 83.90 16 96.26 87.83 72.60 17 88.98 80.15 72.90 18 94.19 96.26 98.61 19 94.64 78.68 80.25 20 90.91 86.94 83.52 21 90.26 83.41 77.66 22 90.48 77.19 69.40 23 96.33 90.63 84.47 24 93.75 78.64 74.82 mean 92.23±4.22 87.86±9.16 83.29±15.32

[0143] The evaluation results based on the lung vessel voxel points are shown in Table 6. Among them, the average coefficient of Voxel-recall is 87.21±6.81%; the average coefficient of Voxel-precision is 88.56±7.79%; and the average coefficient of Voxel-accuracy is 88.22±7.20%. Among them, the Voxel-precision evaluation index is the highest, which shows that a part of the veins are misclassified as arteries, which may be caused by the mistake in classifying the lung vessel skeleton, or may be caused by the problem of occupying part of the voxel points during the reconstruction of the lung vessels, as shown in Figure 10

[0144] ​Table 6: Evaluation results based on lung vessel voxel points

[0145] Dataset Voxel-recall (%) Voxel-precision (%) Voxel-accuracy (%) 1 91.86 95.61 93.83 2 92.39 83.92 87.42 3 80.15 85.29 84.42 4 88.43 86.79 88.84 5 92.68 94.24 94.40 6 87.32 91.80 89.74 7 71.44 92.08 82.89 8 87.01 86.94 87.32 9 88.79 92.36 90.92 10 94.93 96.32 95.42 11 84.31 81.20 83.15 12 91.81 90.83 92.66 13 88.74 87.92 88.31 14 93.05 91.41 91.75 15 93.06 92.77 93.11 16 79.94 80.89 83.00 17 84.01 88.39 86.91 18 76.43 96.35 87.20 19 87.92 87.56 87.96 20 91.72 87.98 89.71 21 87.63 82.57 83.37 22 78.12 80.49 78.15 23 92.35 91.16 92.43 24 89.08 80.46 84.42 mean 87.22±6.81 88.56±7.79 88.22±7.20

[0146] The pulmonary artery carries venous blood with a high carbon dioxide content, and the pulmonary vein carries oxygenated blood. The topological characteristics of the pulmonary artery and the pulmonary vein are similar. The existing algorithm uses the proximity of the pulmonary artery and the pulmonary vein. In theory, this is correct, but in practice, the pulmonary artery and the pulmonary vein are intertwined, and there may be a situation where a pulmonary vein is closer to a pulmonary artery. Therefore, we believe that the pulmonary vein can only assist the pulmonary artery and vein classification algorithm, but not the main role or decisive feature. If only the pulmonary vein is used to classify the pulmonary artery and vein, this method may not have generalization ability. The method of the present application uses the topological characteristics of the pulmonary vein to constrain the pulmonary artery and vein classification algorithm, that is, to weight the edges when classifying.

[0147] To verify that the pulmonary vein guidance and distance constraint feature further improves the classification algorithm effect, the ablation experiment is performed. First, control the pulmonary vein guidance and distance constraint feature as a variable, and perform a comparative experiment based on the lung external vessel distance weighted evaluation index. The experimental results show that this feature improves the algorithm effect on most data, and the average improvement on this index is 3.13%, and the P value result is 0.00725, which is less than 0.01, and belongs to extremely significant difference, which proves the effectiveness of the pulmonary vein guidance and distance constraint. The specific results are shown in Table 7.

[0148] Table 7: Ablation experiment based on the pulmonary vein guidance and distance constraint feature

[0149] Dataset Distance-weighted indicators without lung airway features (%) Distance-weighted indicators with lung airway features (%) 1 93.42 96.75 2 89.77 92.44 3 86.57 89.94 4 92.13 94.16 5 94.13 95.24 6 91.62 95.87 7 95.01 97.95 8 93.45 97.12 9 97.97 97.95 10 95.76 97.72 11 86.44 90.50 12 93.40 95.40 13 91.66 95.00 14 92.02 94.97 15 91.97 96.25 16 88.43 90.16 17 86.43 91.19 18 99.24 99.24 19 89.33 93.09 20 90.26 94.60 21 85.03 90.38 22 80.62 85.12 23 90.27 95.62 24 86.92 90.39 mean 90.91±7.06 94.04±5.20

[0150] The pulmonary artery and vein classification results, the corresponding pulmonary vessel skeleton, and the corresponding gold standard are visualized from multiple angles, Figure 11 Some of the artery and vein classification results are shown in FIG. 8.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not limited thereto; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present application.

Claims

1. A method for automatically extracting pulmonary arteries and veins from chest plain CT images, characterized in that: Includes the following steps: Step 1: Acquire data on the trachea, pulmonary vessels, and extrapulmonary arteries and veins, and perform data preprocessing to standardize slice thickness; Step 2: Based on the reconstructed CT scan images, perform pulmonary vascularization. Step 3: Classify the pulmonary vascular skeleton using the pulmonary arteriovenous growth algorithm; Step 3.1: Determine the preliminary seed points for extrapulmonary arteries and veins; Step 3.2: Use Euclidean space vector features to constrain the multi-branch nodes caused by adhesion points in the pulmonary vascular skeleton, so as to weight the correct growth direction of the pulmonary vascular skeleton. Step 3.3: Skeletalize the pulmonary trachea and use pulmonary trachea guidance and distance features to apply weighted constraints on the correct pulmonary vascular skeleton growth direction; Step 3.4: Combine the pulmonary vascular skeleton with the spherical movement algorithm to grow pulmonary blood vessels; Step 3.5: Use a voting decision-making method to classify the unclassified pulmonary vessels and skeleton; Step 4: Reconstruct the pulmonary arteries and veins based on the arteries and veins marked by the pulmonary vascular skeleton; A preemptive breadth-first search algorithm is used to put all nodes that make up the pulmonary vascular skeleton and their corresponding categories into a queue and mark them. Then, nodes are continuously removed from the queue. When a node is removed, its corresponding voxel point in the 3D image data is assigned the corresponding category, and all unmarked neighborhood points of the voxel point and their corresponding categories are put into the queue and marked. This process continues until all voxel points in the queue are removed, thus completing the pulmonary arteriovenous reconstruction. Step 5: Evaluate the classified pulmonary vessels based on the marked vascular skeleton of the pulmonary arteries and veins, and the vascular voxel points after the reconstruction of the pulmonary arteries and veins.

2. The method for automatic extraction of pulmonary arteries and veins in chest plain CT images according to claim 1, characterized in that: The specific method for step 1 is as follows: Step 1.1: Obtain data on the trachea, pulmonary vessels, and extrapulmonary arteries and veins; A dataset containing multiple sets of CT scan images was obtained, and each set of CT scan images was reconstructed using a 512*512 matrix. The reconstruction method adopted the mediastinal window algorithm, and the slice thickness parameters of the reconstruction were different. Each set of CT scan images contained different lung tissues of the same lung and their corresponding labels, specifically including lung parenchyma, trachea, lobes, pulmonary arteries and pulmonary veins. The extrapulmonary arterial and venous labels are obtained by performing an intersection and union operation on the lung parenchyma label and the pulmonary arterial and venous labels; the categories of the pulmonary arterial and venous labels are then reset as unclassified pulmonary vessels. Step 1.2: For pulmonary trachea, pulmonary blood vessels, and extrapulmonary arteries and veins and their corresponding labels, the nearest neighbor method is used to unify the layer thickness.

3. The method for automatic extraction of pulmonary arteries and veins from chest plain CT images according to claim 2, characterized in that: The specific method for step 2 is as follows: Using numerical analysis software, a set of reconstructed two-dimensional CT scan images were converted into three-dimensional image data. In the three-dimensional image data, voxel points that do not belong to the pulmonary vessels were defined as background points, voxel points on the surface of the pulmonary vessels that contact the background points were defined as boundary points, and voxel points that belong to the pulmonary vessel category were defined as foreground points. During the pulmonary vascular skeletonization process, all pulmonary vascular voxel points in the 3D image data are traversed, and the category of the current voxel point is determined, and different choices are made accordingly: for a certain voxel point, if it is a background point, the next voxel point is scanned; otherwise, it is determined whether it is a boundary point. If it is a boundary point, count the number of foreground points in the 26 neighborhoods of the boundary point. If there is only one foreground point, the boundary point cannot be deleted. Otherwise, continue to check the Euler invariance of the boundary point. If the Euler property value is unchanged, continue to check whether the boundary point is a simple point. Checking whether it is a simple point means checking whether removing the point will not create a cavity or void. If it is not a simple point and the boundary point has not been marked, add the boundary point to the queue and mark it. Then, check if the boundary points in the queue are simple boundary points. If they are, delete them; otherwise, keep them. Continue deleting all simple boundary points until the queue is empty. The data structure after pulmonary vascular skeletonization is described as: for each pulmonary vascular skeleton, a branch node object. Its attributes include identification number i, three-dimensional coordinates Identity number of all connected edges Is it a leaf node? For each edge object of the pulmonary vascular skeleton Its attributes include the identity number j and the two endpoints it possesses. The set of discrete points that make up the edge .

4. The method for automatic extraction of pulmonary arteries and veins from chest plain CT images according to claim 3, characterized in that: The specific method for step 3.1 is as follows: An expansion algorithm is used to fill the gap between the extrapulmonary arterial and venous labels and the intrapulmonary vascular labels. The expansion scale does not exceed 20 voxels to maintain the connectivity between pulmonary vessels. The process for determining the prior seed points is as follows: the images of extrapulmonary arterial and venous vessels and pulmonary vessels are superimposed, and the part where the two meet is obtained at the hilum. The voxels in the part where they meet are set as the prior seed points of extrapulmonary arterial and venous vessels, and the prior seed point category is marked according to the label of extrapulmonary arterial and venous vessels.

5. The method for automatic extraction of pulmonary arteries and veins in chest plain CT images according to claim 4, characterized in that: The specific method for step 3.2 is as follows: Traverse all adjacent edges of the multi-branch node caused by adhesion points in the pulmonary vascular skeleton, and select the adjacent edge direction with the largest angle with the current pulmonary vascular skeleton growth direction as the highest priority growth direction. To avoid erroneous growth, the cosine value between the highest priority growth direction and the current pulmonary vascular bone growth direction must be less than a set threshold C; when calculating the angle, the starting point of the growth direction is a multi-branch node, and the cosine value is used to evaluate whether the growth direction has consistency. For the current pulmonary vascular bone growth direction vector The vectors of all adjacent edges of its multi-way nodes are Then the weight of the growth direction The calculation is shown in the following formula: ； Where N is the total number of all adjacent edges of the multi-branch endpoint.

6. The method for automatic extraction of pulmonary arteries and veins in chest plain CT images according to claim 5, characterized in that: The specific method for step 3.3 is as follows: The pulmonary trachea was skeletonized using a pulmonary vascular skeletonization method; and pulmonary trachea guidance and distance constraints were achieved using a weighted approach using Euclidean space vector angles and spatial distances; among which, the pulmonary trachea With pulmonary vessels Distance features Subtract the Euclidean distance between the two from the threshold D. The criterion for judging Euclidean distance is the trachea. Midpoint of skeletal structure to pulmonary vessels Midpoint of skeletal structure The Euclidean distance between two points; the closer the Euclidean distance between them, the larger their distance feature. If the Euclidean distance between them exceeds a threshold, the distance feature is set to 0; trachea With pulmonary vessels vector features The reciprocal of the sine of the skeletal structure between the two; vector The direction vector between the two endpoints of the pulmonary vessels; the direction vector of the pulmonary trachea. and The closer the growth directions, the smaller the sine value and the larger the eigenvalue; the overall weight eigenvalue of the pulmonary trachea to the artery is obtained based on the guidance and distance characteristics of the pulmonary trachea. As shown in the formula below: ； Where L is the threshold of the guiding feature.

7. The method for automatic extraction of pulmonary arteries and veins in chest plain CT images according to claim 6, characterized in that: The specific method for step 3.4 is as follows: Using the pulmonary vascular skeleton as a growth guide, a sphere tangent to the pulmonary vascular wall continuously moves forward, marking the pulmonary vascular skeleton it passes through. Starting from a prior seed point, the sphere adjusts its radius according to the thickness of the pulmonary vessels during its journey, maintaining tangency to the pulmonary vascular wall. Once the sphere grows to the adhesion point (i.e., encountering a multi-branch node during its journey), it prioritizes selecting a path that minimizes its volume change while still tangent to the pulmonary vascular wall. The sphere's direction of movement is guided by pulmonary tracheal guidance, distance constraints, and Euclidean space vector constraints. Specifically, when the sphere is at a multi-branch node, the radius of the pulmonary vessels at the current location of the sphere is... The radius of the remaining blood vessels at the multi-branch node is The radius of the pulmonary vessels is used as an indicator to judge the magnitude of the spherical change, and Euclidean space vectors are used to constrain the features. and tracheal guidance and distance constraint characteristics Finally, the weights between the edge where the sphere is currently located and the other adjacent edges of the multi-way node are obtained. The formula is shown below: ； in, , , All are weighting coefficients; This represents the gradient difference between the thickness of the lung vessels to be developed and the current thickness of the lung vessels. The gradient difference in the distance traveled by the sphere. This represents the gradient of change in pulmonary vessel wall thickness as the sphere travels; After combining Euclidean spatial vector features, tracheal guidance and distance constraint features, the sphere selects a growth direction and continues to move forward until the pulmonary blood vessels grow to the terminal node.

8. The method for automatic extraction of pulmonary arteries and veins in chest plain CT images according to claim 7, characterized in that: The specific method for step 3.4 is as follows: For pulmonary vascular skeletons without a defined category, their adjacent edges are obtained based on their data structure. All classified adjacent edges of the pulmonary vascular skeleton are then evaluated using metrics including Euclidean distance, Euclidean vector, pulmonary vessel thickness deviation, and pulmonary vessel thickness gradient variation difference. Each classified adjacent edge... It will calculate the sum of four assessment indicators between the current unclassified pulmonary vascular skeleton and its own. , When the value is greater than 0, it indicates that the features of both belong to the same category; All adjacent edges are summed after evaluation. If the pulmonary vascular skeleton belongs to an artery, the summation is positive; otherwise, it is negative. The final category of the pulmonary vascular skeleton, whether it belongs to an artery or a vein, is determined by the sign of the summation, as shown in the following formula: ； ； Finally, all the pulmonary vascular skeletons were labeled as arteries and veins.

9. The method for automatic extraction of pulmonary arteries and veins in chest plain CT images according to claim 8, characterized in that: Step 5 uses the following two methods to evaluate the classified pulmonary vessels: (1) Distance-weighted assessment based on extrapulmonary vascular distance; First, the entire pulmonary vascular data voxels are traversed, and the center and farthest voxels of the entire pulmonary vascular data are obtained, thus obtaining the overall radius *r* of the pulmonary vascular data. For each pulmonary vascular skeleton, the median point in its discrete point set index is selected, i.e., the midpoint of the skeleton, and its three-dimensional coordinates are obtained, yielding the distance *d* between this point and the center point. For the midpoint of the pulmonary vascular skeleton, the class of this point is verified against the predicted class in the gold standard data. If they match, the pulmonary vascular skeleton is considered correctly predicted; otherwise, it is considered incorrectly predicted. Whether it is correctly predicted is determined by... This indicates that a distance-weighted assessment result based on the distance to extrapulmonary vessels is obtained. As shown in the formula below: ； ； The result value The closer the result is to 1, the better the classification effect of pulmonary vessels; the closer the result is to 0, the worse the classification effect. (2) Growth grade grouping assessment: Let the pulmonary vascular skeleton near the hilum be level 1. Each time a new edge grows, the level of that edge increases by one level. The set of pulmonary vascular skeletons of level i is... Every 4 levels of growth, these edges are grouped together. Specifically, starting from level a, growing for 4 levels, and ending at level a+3, the classification accuracy of each group is evaluated, as shown in the following formula. ； in, Indicates whether a particular pulmonary vessel skeleton was correctly predicted; The result value The closer the result is to 1, the better the classification effect of the pulmonary vessels in that group; the closer the result is to 0, the worse the classification effect of the pulmonary vessels in that group.

Images