Segmentation and calibration method of pulmonary bronchi
Through the combination of voxel sharpening filtering and deep learning convolutional neural network, the problem of inaccurate lung airway segmentation is solved, and the accurate automatic segmentation and name annotation of lung airways is realized, which improves the accuracy and efficiency of surgery .
Patent Information
- Application Number
- CN201910719059.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2019-08-05
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2039-08-05
AI Technical Summary
The existing lung airway segmentation methods are easily affected by noise, are inaccurate in segmentation, and it is difficult to accurately calibrate the airway structure in respiratory endoscopy, which increases the difficulty and risk of surgery.
Using a method based on voxel sharpening filtering and deep learning convolutional neural network, the lung airway tree area is extracted, and the leakage is removed through deep learning is classified and tracked. The contour is corrected in combination with the conditional random field, and the airway is accurately and automatically segmented. Automatically label the airway name through the graph neural network.
It significantly improves the accuracy and speed of lung airway segmentation, increases the number of detections of airway branches, reduces the impact of noise, and realizes accurate and automatic labeling of airway names, reducing the risk of surgical errors.
Smart Images
Figure CN112330686B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for segmenting and calibrating pulmonary bronchi, and in particular to a method for accurate automatic segmentation of pulmonary airways based on voxel sharpening tracking and deep learning convolutional neural network and automatic calibration of pulmonary bronchi names based on deep learning convolutional neural network. Background Art
[0002] The existing segmentation methods of lung airways are mainly based on the adaptive region growing method, that is, automatically or manually selecting seed points, starting from the seed points, selecting candidate points in the 6-neighborhood or 26-neighborhood, judging whether the candidate points are target points according to the growth criteria, and then using the target points as new seed points for the next adaptive growth. This type of method is implemented by a variety of different algorithms, and the difference between these algorithms mainly lies in the setting of the growth criteria. Some region growing algorithms directly compare the difference in voxel values, and voxels with differences within a certain range are considered to be in the same region. This method is relatively simple, but it is easily affected by noise and causes under-segmentation. At the same time, some people proposed a dual threshold algorithm to improve the above algorithm, but the overall effect is general, the speed is slow, and few airway branches can be found. In addition, some methods use the image gray value as the growth criterion for a growth, and then add secondary growth. The secondary growth criterion adds professional experience to correspond the relationship between the airway structure features and the calculated features (Hessian features), so as to judge whether the voxel meets the growth criterion; but the disadvantages of this algorithm are slow speed, setting secondary growth increases the amount of calculation, and the amount of calculation for judging the secondary growth criterion is large.
[0003] Because the lung airway is a complex tree-like structure, during the intervention of respiratory endoscopic surgery, it is difficult for doctors (especially young doctors) to grasp the position of the endoscope in the lungs, and they often need to refer to the structural drawings of the lung airway organs to control the direction of the endoscope. This increases the difficulty, risk and time of interventional surgery, and can also lead to surgical errors.
[0004] Therefore, it is urgent to develop a segmentation and calibration method for the lung bronchi that overcomes the above-mentioned defects. Summary of the invention
[0005] In view of the above problems, the present invention provides a method for segmenting and calibrating pulmonary bronchi to solve the above technical problems, which includes:
[0006] Step S1: extracting the pulmonary airway tree region from the image data;
[0007] Step S2: setting a voxel region of interest according to the pulmonary airway tree region, and performing sharpening filtering on the voxel region of interest;
[0008] Step S3: classifying the voxel region of interest through a deep learning convolutional neural network according to the processed voxel region of interest, removing leakage and tracing airway branches to obtain the optimized pulmonary airway tree region;
[0009] Step S4: extracting airway tree structure information from the optimized pulmonary airway tree region, and obtaining a set of candidate names for each airway and a weight of each candidate name through a pulmonary airway name labeling model according to the pulmonary airway tree region and the airway tree structure information;
[0010] Step S5: Optimizing the name of each group of airway trees according to the weights, accurately and automatically labeling the airway tree names, and obtaining airway name information.
[0011] The above segmentation and calibration method, wherein the step S1 includes:
[0012] Step S11: performing smoothing processing on the image data to remove noise in the image data;
[0013] Step S12: setting seed points on the processed image data to form a seed point set;
[0014] Step S13: extracting the pulmonary airway tree region from the image data using a region growing algorithm according to the seed point set.
[0015] The above segmentation and calibration method, wherein the step S2 includes:
[0016] Step S21: defining and setting an initial voxel region of interest according to the extracted pulmonary airway tree region, and determining a plurality of voxel region of interest parameters;
[0017] Step S22: introducing a sharpening filter and using a Laplacian of Gaussian operator to locally enhance the initial voxel region of interest;
[0018] Step S23: Laplace convolution is performed on the voxel grayscale values in the voxel region of interest that are smaller than the voxel grayscale value at the center of the Laplace kernel, so that the grayscale at the edge of the contour increases and the grayscale inside the contour decreases, thereby obtaining the final voxel region of interest.
[0019] The above segmentation and calibration method, wherein the step S3 includes:
[0020] Step S31: extracting the airway part in the final voxel region of interest by the deep learning convolutional neural network according to the final voxel region of interest to remove leakage, and updating the pulmonary airway tree region;
[0021] Step S32: determining the number of branches in the updated pulmonary airway tree region; if the number is 0, stopping the tracking of the pulmonary airway tree region; if the number is 1, moving the pulmonary airway tree region in the direction of bronchial extension by a distance of one voxel, and executing step S31; if the number of branches is greater than 1, setting a sub-airway region;
[0022] Step S33: Correcting the updated pulmonary airway tree region based on the airway contour correction method of the conditional random field;
[0023] Step S34: identifying the bronchial bifurcation site of the corrected pulmonary airway tree region to obtain the optimized pulmonary airway tree region.
[0024] The above segmentation and calibration method, wherein the step S4 includes:
[0025] Step S41: extracting airway tree structure information from the optimized pulmonary airway tree region, and obtaining a set of candidate names for each airway and a weight of each candidate name through a pulmonary airway name labeling model according to the pulmonary airway tree region and the airway tree structure information;
[0026] Step S42: constructing an airway tree name combination according to the candidate name set, each group of the airway tree name combination includes the weight of each airway candidate name;
[0027] Step S43: After optimizing each name of each group of airway trees according to the weight of each candidate airway name, the airway tree names are accurately and automatically labeled to obtain airway name information.
[0028] The above segmentation and calibration method, wherein the step S5 includes:
[0029] The name of each group of airway trees is optimized according to the weights, and accurate and automatic labeling of the airway tree names is performed to obtain airway name information.
[0030] The above segmentation and calibration method, wherein the step S41 includes:
[0031] Step S411: extracting airway tree structure information from the optimized lung airway tree region;
[0032] Step S412: obtaining a parameterized airway characterization value of each airway according to the airway tree structure information;
[0033] Step S413: calibrating the name of each airway according to the parameterized airway characterization value, and then learning and training the parameterized airway characterization value and the corresponding labeled airway name through a graph neural network depth to obtain a lung airway name labeling model;
[0034] Step S414: obtaining a set of candidate names for each airway and a weight of each candidate name through the pulmonary airway name labeling model according to the pulmonary airway tree region and the airway tree structure information.
[0035] The above segmentation and calibration method further includes:
[0036] Step S6: guiding the surgical intervention process according to the airway name information in combination with the real-time bronchoscopic video image.
[0037] The above-mentioned segmentation and calibration method, wherein the parameterized airway characterization value includes: the length of the airway, the direction of the airway, the number of sub-airways of the airway, the distance between the starting point of the airway and the end point of the main airway, the average direction of the sub-airways of the airway, the relative distance between the end point of the airway and the starting point of its father airway, the direction of the brother airways of the airway, the angle between the sub-airways of the airway, the angle between the airway and the father airway, the angle between the airway and the grandfather airway, the angle between the father airway of the airway and the grandfather airway, and the angle between the airway and the main airway.
[0038] In the above-mentioned segmentation and calibration method, the deep learning convolutional neural network is one of a three-dimensional deep supervision network, a three-dimensional U-Net network and a three-dimensional V-Net network.
[0039] The present invention is directed to the prior art and has the following effects:
[0040] 1) Define and track the cuboid voxel region of interest, consider the relationship between voxels in the airway's three-dimensional space, introduce airway edge voxel sharpening filtering, deep learning convolutional neural network and conditional random field airway contour correction method, and accurately and automatically extract the lung airway, which can reduce the leakage, reduce the influence of image partial volume effect and image noise, greatly increase the number of detected airway branches and repair them to make them complete, and finally comprehensively improve the segmentation accuracy and speed;
[0041] 2) A parameterized airway characterization value that defines the lung airways is proposed, and a graph neural network deep learning method is introduced to generate a causal reasoning model that can automatically and accurately label the names of the lung airways, which can reduce optimization time and improve the accuracy and speed of automatic airway name labeling.
[0042] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0044] Figure 1 A flow chart of the segmentation and calibration method of the pulmonary bronchi of the present invention;
[0045] Figure 2 for Figure 1 A step-by-step flow chart of step S1;
[0046] Figure 3 for Figure 1 The step-by-step flow chart of step S2;
[0047] Figure 4 for Figure 1 A step-by-step flow chart of step S3;
[0048] Figure 5 for Figure 1 The step-by-step flow chart of step S4;
[0049] Figure 6 for Figure 5 The step-by-step flow chart of step S411;
[0050] Figure 7 Set up a schematic for the voxel region of interest;
[0051] Figure 8 Schematic diagram of the segmentation process;
[0052] Fig. 9 Schematic diagram of the position calibration process. DETAILED DESCRIPTION
[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0054] The exemplary embodiments and descriptions of the present invention are used to explain the present invention, but are not intended to limit the present invention. In addition, elements / components with the same or similar reference numerals used in the drawings and embodiments are used to represent the same or similar parts.
[0055] The terms “first”, “second”, etc. used in this document do not particularly refer to an order or sequence, nor are they used to limit the present invention. They are only used to distinguish elements or operations described with the same technical terms.
[0056] The directional terms used in this document, such as up, down, left, right, front or back, etc., are only used to refer to the directions of the drawings. Therefore, the directional terms used are used to illustrate and not to limit the present invention.
[0057] The words “include,” “including,” “have,” “contain,” etc. used in this document are open-ended terms, meaning including but not limited to.
[0058] As used herein, "and / or" includes any or all combinations of the items described.
[0059] As used herein, “plurality” includes “two” and “more than two”; as used herein, “plurality groups” includes “two groups” and “more than two groups”.
[0060] The terms "substantially" and "approximately" used herein are used to modify any quantity or error that may vary slightly, but these slight changes or errors do not change their essence. Generally speaking, the range of slight changes or errors modified by such terms may be 20% in some embodiments, 10% in some embodiments, 5% in some embodiments, or other values. Those skilled in the art should understand that the aforementioned values can be adjusted according to actual needs and are not limited thereto.
[0061] Please refer to Figure 1 and Figure 8 , Figure 1 A flow chart of the segmentation and calibration method of the pulmonary bronchi of the present invention; Figure 8 is a schematic diagram of the segmentation process. Figure 1 and Figure 8 As shown, the segmentation and calibration method of the lung bronchi of the present invention comprises the following steps:
[0062] Step S1: extracting the pulmonary airway tree region from the image data;
[0063] Please refer to Figure 2 , Figure 2 for Figure 1 The step-by-step flow chart of step S1 in FIG. Figure 2 As shown, step S1 includes:
[0064] Step S11: Smoothing the image data to remove noise in the image data.
[0065] Specifically, the average gradient value of the air voxels of the image data is calculated, the noise level is determined, and then the anisotropic diffusion filter based on the Level-Set method is used for denoising. The number of image filtering times depends on the noise level.
[0066] Step S12: Setting seed points on the processed image data to form a seed point set.
[0067] Step S13: extracting the pulmonary airway tree region from the image data using a region growing algorithm according to the seed point set.
[0068] Specifically, the airway tree is extracted using a region growing algorithm. The region growing algorithm refers to the process of developing groups of pixels or regions into larger regions. Specifically, it starts from a set of seed points and merges adjacent pixels with similar attributes to each seed point, such as intensity, grayscale, texture color, etc., into this region. The seed point setting can be manually set or automatically obtained. The specific method of automatic seed point acquisition is as follows: by setting a threshold, a binary image is obtained on an image section of a certain image data; then the connected area in the binary image is automatically detected, and the main trachea position is found by judging the size of the connected area, and the center position of the main trachea is obtained as the seed point. After obtaining the seed point, any region growing method is used to preliminarily extract all lung airway tree regions in the image data.
[0069] Step S2: setting a voxel region of interest according to the pulmonary airway tree region, and performing sharpening filtering on the voxel region of interest.
[0070] Please refer to Figure 3 and Figure 7 , Figure 3 Step S2 Figure 1 A step-by-step flow chart of Figure 7 Set up a schematic for the voxel region of interest. Figure 3 and Figure 7 As shown, step S2 includes:
[0071] Step S21: defining and setting an initial voxel region of interest according to the extracted pulmonary airway tree region, and determining a plurality of voxel region of interest parameters.
[0072] In this embodiment, the voxel region of interest is a cuboid voxel region of interest ( Figure 7 ), the parameters of the cuboid voxel region of interest include 8 nodes D 1 -D 8 Determine the boundaries, the center of gravity G of the lumen, the upper and lower sections and their two center points P 1 , P 2 .
[0073] Step S22: Introduce sharpening filtering and use the Laplacian of Gaussian operator to locally enhance the initial voxel region of interest.
[0074] Step S23: Laplace convolution is performed on the voxel grayscale values in the voxel region of interest that are smaller than the voxel grayscale value at the center of the Laplace kernel, so that the grayscale at the edge of the contour increases and the grayscale inside the contour decreases, thereby obtaining the final voxel region of interest.
[0075] Specifically, due to the partial volume effect of medical imaging, lumen features are lost, especially the bronchial airway contour is blurred. To address this problem, the present invention intends to introduce a sharpening filter, using the Gaussian Laplacian operator to locally enhance the cuboid voxel region of interest defined above, and then perform Laplacian convolution on the voxel grayscale value in the cuboid voxel region of interest that is less than the voxel grayscale value at the center of the Laplacian kernel, so that the grayscale of the contour edge increases and the grayscale of the contour interior decreases. Finally, the processing result of sharpening the cuboid voxel region of interest is obtained.
[0076] Step S3: Classify the voxel region of interest through a deep learning convolutional neural network according to the processed voxel region of interest, remove leakage and trace airway branches, and obtain an optimized lung airway tree region.
[0077] Among them, in this embodiment, the deep learning convolutional neural network is a three-dimensional deep supervision network, but the present invention is not limited to this. In other embodiments, the deep learning convolutional neural network can also be a three-dimensional U-Net network or a three-dimensional V-Net network. Taking the three-dimensional deep supervision network as an example, the three-dimensional deep supervision network of the present invention includes a three-dimensional full convolutional network, a three-dimensional deconvolution layer and a three-dimensional deep supervision mechanism. The three-dimensional full convolutional network mainly includes a three-dimensional convolutional layer and a three-dimensional subsampling layer. Three-dimensional convolution uses a nonlinear kernel function (activation function) based on random gradient descent. In the three-dimensional subsampling process, the maximum pooling function is introduced to reduce the output features and parameters of the convolutional layer, while keeping the learned features in the three-dimensional space. The purpose of introducing the three-dimensional deconvolution layer is to solve the sparsity problem of the output results of the three-dimensional full convolutional network. There are gradient vanishing and explosion in deep network training, which leads to failure of back propagation and non-convergence of training. In response to these problems, a three-dimensional deep supervision mechanism and an auxiliary loss error classification method are introduced. Using the three-dimensional deep learning convolutional neural network method to extract and optimize the airway can speed up the optimization speed and greatly improve the segmentation accuracy. In the classification process of deep learning convolutional neural networks, the three-dimensional rectangular region of interest is input instead of single voxel information. This takes into account the regional relationship between voxels, reduces information loss, and reduces the impact of noise.
[0078] Please refer to Figure 4 , Figure 4 for Figure 1 The step-by-step flow chart of step S3 in FIG. Figure 4 As shown, step S3 includes:
[0079] Step S31: extract the airway part in the final voxel region of interest through a deep learning convolutional neural network according to the final voxel region of interest, remove the leakage, and update the lung airway tree region.
[0080] Step S32: Determine the number of branches in the updated pulmonary airway tree region. If it is 0, stop tracking the pulmonary airway tree region. If it is 1, move the pulmonary airway tree region in the direction of bronchial extension by a distance of one voxel and execute step S31. If the number of branches is greater than 1, set the sub-airway region.
[0081] Step S33: Correcting the updated pulmonary airway tree region based on the airway contour correction method of the conditional random field.
[0082] Specifically, although most of the bronchial contours have been accurately segmented after image sharpening and deep learning convolutional neural network voxel classification, the contour segmentation of some specific fuzzy areas is still inaccurate. Therefore, the present invention intends to introduce an airway contour correction method based on conditional random fields to further improve the segmentation accuracy. The contour correction method based on the conditional random field model first defines an energy cost function, which includes a voxel conditional probability function and the output after bilateral filtering of the voxels, wherein the voxel conditional probability function is initialized as a linear combination of the final result based on the three-dimensional deep supervision network classification and the intermediate prediction result; then optimization is implemented to finally obtain the corrected result.
[0083] Step S34: identifying the bronchial bifurcation site of the corrected pulmonary airway tree region to obtain an optimized pulmonary airway tree region.
[0084] Specifically, the identification of bronchial bifurcation sites is based on the analysis of the number of connected regions in the correction results. The number of bifurcation sites depends on the number of connected regions N. c After each cuboid is sharpened, classified and corrected, D 5 D 6 D 7 D 8 section( Figure 7 ), using the binary search method, detect the number of connected areas N on the cross section c :(1) If N c = 0, there is no bifurcation, and detection stops; (2) If N c = 1, then expand the cuboid area and repeat the above steps to segment until the trachea is completely segmented; (3) If N c = 2 or 3, the bifurcation has been successfully detected, and then connect G and P 2( Figure 7 ), in this direction GP 2 Continue to track (set) the sub-rectangular area and repeat the above steps for segmentation until all bronchial areas are segmented.
[0085] Please refer to Fig. 9 Please combine Figure 1 , Fig. 9 Step S4: extracting airway tree structure information from the optimized pulmonary airway tree region, and obtaining a set of candidate names for each airway and a weight of each candidate name through a pulmonary airway name labeling model according to the pulmonary airway tree region and airway tree structure information.
[0086] Please refer to Figure 5 , Figure 5 for Figure 1 The step-by-step flow chart of step S4 in FIG. Figure 5 As shown, step S4 includes:
[0087] Step S41: extracting airway tree structure information from the optimized pulmonary airway tree region, and obtaining a set of candidate names for each airway and a weight of each candidate name through a pulmonary airway name labeling model according to the pulmonary airway tree region and the airway tree structure information.
[0088] Step S42: construct airway tree name combinations according to the candidate name set, each set of airway tree name combinations includes the weight of each airway candidate name.
[0089] Step S43: After optimizing each name of each group of airway trees according to the weight of each candidate airway name, the airway tree names are accurately and automatically labeled, and the airway name information is obtained.
[0090] For further information, please refer to Figure 6 , Figure 6 for Figure 5 The step-by-step flow chart of step S411 in FIG. Figure 6 As shown, step S41 includes:
[0091] Step S411: extracting airway tree structure information from the optimized lung airway tree region.
[0092] Step S412: Obtaining a parameterized airway characterization value of each airway according to the airway tree structure information.
[0093] Step S413: According to the parameterized airway characterization value, the name of each airway is calibrated, and then the parameterized airway characterization value and its corresponding labeled airway name are deeply learned and trained through the graph neural network to obtain a lung airway name labeling model.
[0094] Step S414: Obtain a set of candidate names for each airway and a weight of each candidate name through a pulmonary airway name labeling model according to the pulmonary airway tree region and airway tree structure information.
[0095] Specifically, first, based on the above-mentioned airway automatic segmentation results, 12 parameterized airway characterization values of each airway are calculated according to the airway tree structure information: the length of the airway, the direction of the airway, the number of sub-airways of the airway, the distance between the starting point of the airway and the end point of the main airway, the average direction of the sub-airways of the airway, the relative distance (position) between the end point of the airway and the starting point of its father airway, the direction of the brother airways of the airway, the angle between the sub-airways of the airway, the angle between the airway and the father airway, the angle between the airway and the grandfather airway, the angle between the father airway of the airway and the grandfather airway, and the angle between the airway and the main airway. Among them, the airway tree structure information at least includes information such as the position of the starting point and the end point of each airway, the center line, the radius size, and the bifurcation angle.
[0096] The 12 parameterized airway characterization values defined in the present invention can accurately express and quantify each bronchus, and while accurately segmenting the airway tree, can obtain airway characteristic (characterization) information such as the airway starting point and end point, center line, length, radius, center point of the airway area of each image slice, and the structural relationship between the upper-level airway and the lower-level airway (the structural relationship between the parent airway and the child airway), and express and save these airway characterization characteristic information in a unique XML file form (format).
[0097] Secondly, based on the above-mentioned parameterized airway characterization values, the name of each airway is manually calibrated, and the parameterized airway characterization values and their corresponding labeled airway names are input into the graph neural network for deep learning and training to obtain a lung airway name labeling model.
[0098] Again, for an image data, the above-mentioned automatic lung airway segmentation method based on voxel sharpening tracking and deep learning convolutional neural network is used to extract the airway area in the image and obtain the airway tree structure information.
[0099] Finally, based on the segmentation results and airway tree structure information, combined with the lung airway name labeling model trained based on graph neural network deep learning, the set of candidate names for each airway is calculated, and the weight of each candidate name is calculated.
[0100] Step S5: Optimize the name of each group of airway trees according to the weights, implement accurate and automatic labeling of the airway tree names, and obtain airway name information.
[0101] Furthermore, the segmentation and calibration method of the present invention further includes:
[0102] Step S6: guiding the surgical intervention process based on the airway name information and the real-time bronchoscopic video image.
[0103] In summary, the present invention proposes an accurate and complete solution for automatic segmentation of lung airways by medical image voxel sharpening, voxel tracking, deep learning convolutional neural network and contour correction of conditional random fields. At the same time, it also provides an automatic airway name labeling method based on graph neural network (GNN) deep learning, thereby realizing the consideration of the relationship between voxels in the spatial three-dimensional area of the airway, introducing voxel sharpening filtering and deep learning convolutional neural network methods, and accurately and automatically extracting the lung airways, which can reduce the infiltration, reduce the influence of image partial volume effect and image noise, greatly increase the number of detected airway branches and repair them to make them complete, and finally comprehensively improve the segmentation accuracy and speed. At the same time, a parameterized airway characterization value that defines the lung airway is proposed, and a graph neural network deep learning method is introduced to generate a causal reasoning model, which can automatically and accurately label the lung airway names, can reduce the optimization time, improve the accuracy and speed of automatic airway name labeling, and can guide bronchoscopic surgical intervention.
[0104] Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent substitutions for some of the technical features therein; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for segmenting and calibrating pulmonary bronchi, characterized in that: include: Step S1: extracting the pulmonary airway tree region from the image data; Step S2: setting a voxel region of interest according to the pulmonary airway tree region, and performing sharpening filtering on the voxel region of interest; Step S3: classifying the voxel region of interest through a deep learning convolutional neural network according to the processed voxel region of interest, removing leakage and tracing airway branches to obtain the optimized pulmonary airway tree region; Step S4: extracting airway tree structure information from the optimized pulmonary airway tree region, and obtaining a set of candidate names for each airway and a weight of each candidate name through a pulmonary airway name labeling model according to the pulmonary airway tree region and the airway tree structure information; The step S4 includes: Step S41: extracting airway tree structure information from the optimized pulmonary airway tree region, and obtaining a set of candidate names for each airway and a weight of each candidate name through a pulmonary airway name labeling model according to the pulmonary airway tree region and the airway tree structure information; The step S41 includes: Step S411: extracting airway tree structure information from the optimized lung airway tree region; Step S412: obtaining a parameterized airway characterization value of each airway according to the airway tree structure information; Step S413: calibrating the name of each airway according to the parameterized airway characterization value, and then performing deep learning and training through a graph neural network according to the parameterized airway characterization value and the corresponding labeled airway name to obtain a lung airway name labeling model; Step S414: obtaining a set of candidate names for each airway and a weight of each candidate name through the pulmonary airway name labeling model according to the pulmonary airway tree region and the airway tree structure information; Step S42: constructing an airway tree name combination according to the candidate name set, each group of the airway tree name combination includes the weight of each airway candidate name; Step S43: After optimizing each name of each group of airway trees according to the weight of each candidate name of the airway, the names of the airway trees are accurately and automatically labeled to obtain airway name information; Step S5: Optimizing the name of each group of airway trees according to the weights, accurately and automatically labeling the airway tree names, and obtaining airway name information.
2. The segmentation and calibration method according to claim 1, characterized in that: The step S1 includes: Step S11: performing smoothing processing on the image data to remove noise in the image data; Step S12: setting seed points on the processed image data to form a seed point set; Step S13: extracting the pulmonary airway tree region from the image data using a region growing algorithm according to the seed point set.
3. The segmentation and calibration method according to claim 1 or 2, characterized in that: The step S2 includes: Step S21: defining and setting an initial voxel region of interest according to the extracted pulmonary airway tree region, and determining a plurality of voxel region of interest parameters; Step S22: introducing a sharpening filter and using a Laplacian of Gaussian operator to locally enhance the initial voxel region of interest; Step S23: Laplace convolution is performed on the voxel grayscale values in the voxel region of interest that are smaller than the voxel grayscale value at the center of the Laplace kernel, so that the grayscale at the edge of the contour increases and the grayscale inside the contour decreases, thereby obtaining the final voxel region of interest.
4. The segmentation and calibration method according to claim 3, characterized in that: The step S3 includes: Step S31: extracting the airway part in the final voxel region of interest by the deep learning convolutional neural network according to the final voxel region of interest to remove leakage, and updating the pulmonary airway tree region; Step S32: determining the number of branches in the updated pulmonary airway tree region; if the number is 0, stopping the tracking of the pulmonary airway tree region; if the number is 1, moving the pulmonary airway tree region in the direction of bronchial extension by a distance of one voxel, and executing step S31; if the number of branches is greater than 1, setting a sub-airway region; Step S33: Correcting the updated pulmonary airway tree region based on the airway contour correction method of the conditional random field; Step S34: identifying the bronchial bifurcation site of the corrected pulmonary airway tree region to obtain the optimized pulmonary airway tree region.
5. The segmentation and calibration method according to claim 1, characterized in that: The step S5 includes: The name of each group of airway trees is optimized according to the weights, and accurate and automatic labeling of the airway tree names is performed to obtain airway name information.
6. The segmentation and calibration method according to claim 1, characterized in that: The parameterized airway characterization values include: The length of the airway, the direction of the airway, the number of sub-airways of the airway, the distance between the starting point of the airway and the end point of the main airway, the average direction of the sub-airways of the airway, the relative distance between the end point of the airway and the starting point of its father airway, the direction of the brother airways of the airway, the angle between the sub-airways of the airway, the angle between the airway and the father airway, the angle between the airway and the grandfather airway, the angle between the father airway of the airway and the grandfather airway, and the angle between the airway and the main airway.
7. The segmentation and calibration method according to claim 1, characterized in that: The deep learning convolutional neural network is one of a three-dimensional deep supervision network, a three-dimensional U-Net network and a three-dimensional V-Net network.