A coronary artery intelligent auxiliary analysis method and system for CT images

Abstract

The present invention discloses a method and system for intelligent auxiliary analysis of coronary arteries for CT images. First, an MHD file or DICOM sequence containing CTA volume data is selected to obtain the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data, and to perform a visual display; then, a coronary centerline extraction method based on a three-dimensional void convolutional neural network tracker is used to complete the fully automatic extraction of the coronary centerline, generate a coronary centerline tree, and perform a visual display; finally, the coronary vessels to be analyzed are selected, and the coronary vessels are highlighted in the three-dimensional display area of ​​the centerline, and the selected coronary vessels are straightened and reconstructed at multiple angles and displayed visually. The present invention combines deep learning with an iterative tracking algorithm to complete the fully automatic extraction of the coronary centerline, making full use of existing advanced technologies to reduce the workload of clinical medical staff, and also reducing the accuracy of human intervention post-processing technology, and efficiently and accurately achieving medical auxiliary effects.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical electronics technology, and specifically relates to a coronary artery intelligent auxiliary analysis method and system. Background Art

[0002] Coronary heart disease is a cardiovascular disease with a high incidence and great harm. Its full name is coronary atherosclerotic heart disease. It is a heart disease caused by atherosclerotic lesions in the coronary arteries, which cause vascular stenosis or blockage, resulting in myocardial ischemia, hypoxia or necrosis. With the rapid development of my country's economy, the continuous improvement of people's living standards and changes in eating habits, the incidence of coronary heart disease has increased year by year. Now it has become the number one killer threatening the lives of the Chinese people. Therefore, early prevention and timely diagnosis of coronary artery disease are particularly important.

[0003] The current mainstream diagnosis and treatment option is coronary computed tomography angiography (CTA), also known as non-invasive vascular imaging technology, which injects iodine-containing contrast agents into the blood vessels, performs spiral CT scans on the patient, and then processes the data obtained by computer to obtain vascular images. The imaging results include two-dimensional tomographic images of the coronary vessels and adjacent tissues, providing clinicians with a lot of valuable information for diagnosing coronary artery disease.

[0004] The efficient and accurate extraction of the coronary artery centerline based on coronary CTA medical images is the cornerstone of coronary atherosclerotic heart disease lesion area visualization, coronary disease quantitative analysis, rating, and treatment plan planning. The coronary artery centerline can provide a quantitative description of the relevant characteristics of the vascular structure, such as morphological parameters and topological parameters such as vessel length, vessel diameter, vessel bifurcation point, and vessel curvature. At the same time, the effective information provided by the coronary centerline is used in combination with CPR technology to complete the visualization analysis of the coronary vessels and surrounding tissue information.

[0005] However, it is difficult to accurately analyze the inner lumen information of blood vessels from a complex background by only observing two-dimensional tomographic images, which can easily lead to misdiagnosis and missed diagnosis.

[0006] In the current diagnosis and treatment scheme, clinicians usually need to manually post-process coronary CTA medical images on a workstation and diagnose and analyze the condition based on the post-processing results. This manual operation is not only labor-intensive and inefficient, but also has a certain degree of subjective judgment due to a large amount of human intervention. Therefore, it is necessary to provide an intelligent post-processing technology for coronary CTA medical images to avoid the defects of the current technology.

[0007] In the prior art, the patent - A method and system for matching the centerline of a coronary artery (Chinese patent, application publication number CN 108335284 A, application publication date 2018.07.27) relates to the field of computer vision technology and provides a method and system for matching the centerline of a coronary artery, wherein the method comprises: obtaining a candidate starting point set and a candidate ending point set according to the image of the complete centerline of the coronary artery in the current frame; obtaining the candidate centerline using a depth-first search algorithm according to the candidate starting point set and the candidate ending point set; obtaining the characteristic distance between the candidate centerline and the coronary artery centerline matched in the image of the complete centerline of the coronary artery in the previous frame using a dynamic time warping algorithm; and determining the matching centerline of the coronary artery according to the characteristic distance. The system comprises: an endpoint acquisition module for obtaining a candidate starting point set and a candidate ending point set; a search module for obtaining a candidate centerline using a depth-first search algorithm; and a matching module for determining the matching centerline of the coronary artery using a dynamic time warping algorithm. Although the invention can extract the centerline of the coronary artery, it cannot achieve good results in terms of matching speed and accuracy. Summary of the invention

[0008] In order to overcome the deficiencies of the prior art, the present invention provides a method and system for intelligent auxiliary analysis of coronary arteries for CT images. First, an MHD file or DICOM sequence containing CTA volume data is selected to obtain the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data, and to perform a visual display; then, a coronary centerline extraction method based on a three-dimensional void convolutional neural network tracker is used to complete the fully automatic extraction of the coronary centerline, generate a coronary centerline tree, and perform a visual display; finally, the coronary vessels to be analyzed are selected, and the coronary vessels are highlighted in the three-dimensional display area of ​​the centerline, and the selected coronary vessels are straightened and reconstructed at multiple angles and visualized. The present invention combines deep learning with iterative tracking algorithms to complete the fully automatic extraction of the coronary centerline, making full use of existing advanced technologies to reduce the workload of clinical medical staff, and also reducing the accuracy of human intervention post-processing technology, and efficiently and accurately achieving medical assistance.

[0009] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0010] Step 1: Select the MHD file or DICOM sequence containing CTA volume data through the file selector, obtain the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data, and perform visual display;

[0011] Step 2: Based on the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data obtained in step 1, the coronary centerline extraction method based on the three-dimensional hole convolutional neural network is used to complete the fully automatic extraction of the coronary centerline, generate the coronary centerline tree and visualize it;

[0012] Step 3: Select the coronary artery to be analyzed from the coronary arteries corresponding to the coronary centerline extracted in step 2, highlight the selected coronary artery in the three-dimensional display area of ​​the coronary centerline, and perform multi-angle straightening and surface reconstruction on the selected coronary artery for visual display.

[0013] Furthermore, when the DICOM sequence is selected in step 1, the grayscale value of the CTA volume data of the DICOM sequence is standardized to the grayscale value in the MHD format; then the file processing thread is started to obtain the voxel, spatial spacing, and three-dimensional image size information of the medical imaging data.

[0014] Further, the specific steps of step 2 are as follows:

[0015] Step 2-1: Data preprocessing;

[0016] Step 2-1-1: Manually mark the coronary artery centerline and the coronary artery entrance; randomly select a point x on the manually marked coronary artery centerline, and cut out an isotropic three-dimensional small volume on the CTA volume data with point x as the center voxel, and the size of the three-dimensional small volume is a×a×a;

[0017] Assume that the radius of the coronary artery at point x is r, and a point is found along the manually marked coronary artery centerline at a distance r from point x, recorded as x_0, and the displacement vector between point x and x_0 is Δx; mark multiple points on the tracker sphere of the three-dimensional void convolutional neural network, and designate the direction with the smallest angle with the displacement vector Δx in the direction of the line connecting the center point x and multiple points as the reference direction, and set the designated reference direction category probability to 0.5 in the reference distribution of the potential movement direction D of point x; use the same method to find another point at a distance r in another direction from point x along the manually marked coronary artery centerline, and obtain its reference direction and category probability; at this time, the probabilities of two reference direction categories in the reference distribution of the movement direction D are set to 0.5, and the probabilities of the remaining direction categories are set to 0.0;

[0018] Step 2-1-2: Randomly select a point x on the manually marked coronary artery centerline, use the random displacement d sampled in the three-dimensional normal distribution, translate the position point x by a distance d in any direction, record the translated point as x′, and cut out an isotropic three-dimensional small block on the CTA volume data with point x′ as the central voxel. The size of the three-dimensional small block is a×a×a;

[0019] Find the point x1 closest to point x′ on the manually marked coronary artery centerline, set the coronary artery radius at point x1 to be r, find a point x_0 at a distance r from point x1 along the manually marked coronary artery centerline, and record the displacement vector between point x1 and x_0 as Δx; mark multiple points on the tracker sphere of the three-dimensional void convolutional neural network, designate the direction with the smallest angle with the displacement vector Δx in the direction of the line connecting the center point x1 and multiple points as the reference direction, and set the reference direction category probability to 0.5 in the reference distribution of the potential movement direction D of point x′; use the same method to find another point at a distance r from the position point x1 in another direction along the manually marked coronary artery centerline, and obtain its reference direction and category probability; at this time, the probabilities of two direction categories in the reference distribution of the movement direction D are set to 0.5, and the probabilities of the remaining direction categories are set to 0.0;

[0020] Step 2-1-3: Randomly rotate the isotropic three-dimensional small blocks generated in step 2-1-1 and step 2-1-2 to expand the data set. Each isotropic three-dimensional small block is rotated around the x-axis, y-axis or z-axis at a random angle θ∈[0,2π] to generate three-dimensional small blocks in multiple directions;

[0021] Step 2-2: construct a coronary artery centerline extraction model and train it;

[0022] The coronary artery centerline extraction model includes three parallel three-dimensional hole convolutional neural networks, namely SeedCNN, OstiaCNN and TrackCNN; the network structures of the three three-dimensional hole convolutional neural networks are different only in the output part, and the input is a three-dimensional small block of size a×a×a. The last layer of the output part of SeedCNN performs regression to predict, and the predicted value is the proximity between the central voxel of the input three-dimensional small block and the coronary artery centerline; the last layer of the output part of OstiaCNN performs regression to predict, and the predicted value is the proximity between the central voxel of the input three-dimensional small block and the entrance of the coronary artery; the output part of TrackCNN includes two branches, the first branch uses classification to predict the direction category probability distribution of the central voxel of the input three-dimensional small block, and the second branch uses regression to predict the coronary radius at the central voxel of the input three-dimensional small block;

[0023] The SeedCNN is trained using the 3D small volume generated in step 2-1 and the proximity between the central voxel of the small volume and the manually marked coronary artery centerline;

[0024] OstiaCNN is trained using the 3D small volume generated in step 2-1 and the proximity between the central voxel of the small volume and the manually marked coronary artery entrance;

[0025] Train TrackCNN, and the training data are the three-dimensional small volume generated in step 2-1 and the reference direction of the central voxel of the three-dimensional small volume, as well as the category probability and the coronary artery radius r;

[0026] Step 2-3: Use the coronary centerline extraction model to obtain the entry point and seed point;

[0027] Step 2-3-1: Preprocess the input CTA volume data, remove the lung area, and retain the heart part;

[0028] Step 2-3-2: randomly set multiple center points in the retained heart part, and use each center point as the central voxel to extract a three-dimensional small block with a size of a×a×a;

[0029] Step 2-3-3: Input the extracted 3D small volume blocks into the SeedCNN and OstiaCNN trained in step 2-2 respectively; obtain the predicted value of the proximity between the central voxel of the input 3D small volume block and the centerline of the coronary artery through SeedCNN, and take the central voxels corresponding to the first N values ​​of the predicted proximity value from near to far as seed points to form a seed point queue; obtain the predicted value of the proximity between the central voxel of the input 3D small volume block and the entrance of the coronary artery through OstiaCNN, and take the central voxels corresponding to the first M values ​​of the predicted proximity value from near to far as entry points to form an entry point queue;

[0030] Step 2-4: Use TrackCNN combined with iterative tracking algorithm to simulate the centerline tracker to extract the centerline of a single coronary artery;

[0031] Step 2-4-1: randomly select a seed point from the seed point queue, extract a 3D small volume block with the seed point as the central voxel, input the extracted 3D small volume block into TrackCNN, and obtain the direction category probability distribution and coronary artery radius prediction value at the central voxel;

[0032] Step 2-4-2: Find two local maxima from the direction category probability distribution obtained in step 2-4-1 as the forward movement direction and reverse movement direction of the current central voxel. The angle between the forward movement direction and the reverse movement direction must be greater than 90 degrees.

[0033] Step 2-4-3: The tracker first tracks the center line along the forward motion direction. The tracker uses the predicted value of the coronary radius of the current central voxel as the step size and reaches the new position point along the forward motion direction of the current central voxel. Then, a three-dimensional small volume is extracted with the new position point as the central voxel, and the newly extracted three-dimensional small volume is input into the TrackCNN model to obtain the direction category probability distribution and radius prediction value at the new central voxel. Then, the local maximum value is found from the direction category probability distribution as the new tracking direction, and the angle between the new tracking direction and the initial forward motion direction is less than 60 degrees. Repeat this process until the forward motion direction tracking is terminated when the termination condition is met. The process of the tracker tracking the center line along the reverse motion direction is the same as the process of tracking the forward motion direction. The termination condition of the tracker is the normalized entropy of the direction classification probability distribution.

[0034] Connecting all the position points reached by the trackers during the tracking process will form the centerline of a single coronary artery;

[0035] Step 2-5: Generate coronary centerline tree;

[0036] Step 2-5-1: If the distance between the single coronary artery centerline generated in step 2-4 and any entry point in the entry point queue is less than or equal to the set value, the single coronary artery centerline is saved; otherwise, if the distance between the single coronary artery centerline generated in step 2-4 and all entry points in the entry point queue is greater than the set value, the single coronary artery centerline is discarded;

[0037] Step 2-5-2: remove the seed points in the seed point queue that overlap with the centerline of a single coronary artery from the seed point queue;

[0038] Step 2-6: Reselect a seed point from the seed point queue, and repeat steps 2-4 and 2-5 until all seed points in the seed point queue are traversed, multiple coronary centerlines are extracted, and a coronary centerline tree is formed.

[0039] Furthermore, in step 3, the selected coronary artery is straightened and reorganized along the center line of the coronary artery using a surface reconstruction method, and the reorganized data is cut at 0 degrees, 45 degrees and 90 degrees along the central axis to obtain multi-angle surface reconstruction results, and finally the obtained vascular cross-section is visualized.

[0040] Furthermore, a=19.

[0041] A coronary artery intelligent auxiliary analysis system for CT images, comprising:

[0042] The medical imaging data module is used to select the MHD file or DICOM sequence containing CTA volume data through the file selector, obtain the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data, and perform visual display;

[0043] The centerline extraction module is used to automatically extract the coronary centerline based on the voxel, spatial spacing, and three-dimensional image size information of the acquired CTA volume data using a coronary centerline extraction method based on a three-dimensional hole convolutional neural network, generate a coronary centerline tree, and visualize it;

[0044] The surface reconstruction module is used to select the coronary artery to be analyzed from the coronary arteries corresponding to the coronary centerline extracted by the centerline extraction module, highlight the selected coronary artery in the three-dimensional display area of ​​the coronary centerline, and simultaneously perform multi-angle straightening and surface reconstruction on the selected coronary artery and perform visual display.

[0045] The beneficial effects of the present invention are as follows:

[0046] The present invention combines deep learning with iterative tracking algorithms to complete the fully automatic extraction of the coronary centerline, making full use of existing advanced technologies to reduce the workload of clinical medical staff, and also reducing the accuracy of human intervention post-processing technology, and efficiently and accurately achieves medical assistance. At the same time, based on the above results, the coronary vessel straightening surface reconstruction technology is realized, and multi-angle cutting is performed, so that doctors can more accurately and comprehensively analyze the degree of stenosis and the number of plaques in the blood vessels, thereby assisting doctors to complete the diagnosis of the disease more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a block diagram of the implementation of the present invention.

[0048] Figure 2 This is a flow chart of extracting the centerline of a single coronary artery according to the present invention.

[0049] Figure 3 This is a flow chart of the automatic positioning of the coronary artery centerline entry point and the seed point initialization of the present invention.

[0050] Figure 4 FIG. 4 is a diagram of medical imaging data according to an embodiment of the present invention.

[0051] Figure 5 A coronary artery centerline tree and a selected coronary artery centerline are displayed with highlighting in an embodiment of the present invention.

[0052] Figure 6 Schematic diagram of the coronary artery multi-angle straightening curved surface reconstruction results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0054] The present invention is based on coronary CTA medical images combined with deep learning technology to complete the fully automatic extraction of the coronary blood vessel centerline, and uses CPR technology to display the tomographic information at different levels on the same plane, so that the vascular structure in the three-dimensional space can be comprehensively and meticulously displayed in one image, thereby better assisting doctors in observing and analyzing the inner cavity information of the coronary artery, and ultimately completing the diagnosis and treatment of coronary artery disease.

[0055] like Figure 1 As shown, a method and system for intelligent auxiliary analysis of coronary arteries for CT images include the following steps:

[0056] Step 1: Select the MHD file or DICOM sequence containing CTA volume data through the file selector, start the file processing thread, obtain the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data, and perform visual display;

[0057] Step 2: Based on the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data obtained in step 1, the coronary centerline extraction method based on the three-dimensional hole convolutional neural network is used to complete the fully automatic extraction of the coronary centerline, generate the coronary centerline tree and visualize it;

[0058] Step 3: Select the coronary artery to be analyzed from the coronary arteries corresponding to the coronary centerline extracted in step 2, highlight the selected coronary artery in the three-dimensional display area of ​​the coronary centerline, and perform multi-angle straightening and surface reconstruction on the selected coronary artery for visual display.

[0059] Furthermore, when the DICOM sequence is selected in step 1, the grayscale value of the CTA volume data of the DICOM sequence is standardized to the grayscale value in the MHD format; then the file processing thread is started to obtain the voxel, spatial spacing, and three-dimensional image size information of the medical imaging data.

[0060] Further, the specific steps of step 2 are as follows:

[0061] Step 2-1: Data preprocessing;

[0062] Step 2-1-1: Figure 4 As shown, the coronary artery centerline and the coronary artery entrance are manually marked; a point x is randomly selected on the manually marked coronary artery centerline, and an isotropic three-dimensional small volume is cut out on the CTA volume data with point x as the central voxel, and the size of the three-dimensional small volume is 19×19×19;

[0063] Assume that the radius of the coronary artery at point x is r, and find a point along the manually marked coronary artery centerline at a distance r from point x, record it as x_0, and the displacement vector between point x and x_0 is Δx; mark multiple points on the tracker sphere of the three-dimensional void convolutional neural network, mark multiple points, and designate the direction with the smallest angle with the displacement vector Δx in the direction of the line connecting the center point x and multiple points as the reference direction, and set the designated reference direction category probability to 0.5 in the reference distribution of the potential movement direction D of point x; use the same method to find another point along the manually marked coronary artery centerline at another direction distance r from point x, and obtain its reference direction and category probability; at this time, the probabilities of two reference direction categories in the reference distribution of the movement direction D are set to 0.5, and the probabilities of the remaining direction categories are set to 0.0;

[0064] Step 2-1-2: Randomly select a point x on the manually marked coronary artery centerline, use the random displacement d sampled in the three-dimensional normal distribution, translate the position point x by a distance d in any direction, record the translated point as x′, and cut out an isotropic three-dimensional small block on the CTA volume data with point x′ as the central voxel. The size of the three-dimensional small block is 19×19×19;

[0065] Find the point x1 closest to point x′ on the manually marked coronary artery centerline, set the coronary artery radius at point x1 to be r, find a point x_0 at a distance r from point x1 along the manually marked coronary artery centerline, and record the displacement vector between point x1 and x_0 as Δx; mark multiple points on the tracker sphere of the three-dimensional void convolutional neural network, designate the direction with the smallest angle with the displacement vector Δx in the direction of the line connecting the center point x1 and multiple points as the reference direction, and set the reference direction category probability to 0.5 in the reference distribution of the potential movement direction D of point x′; use the same method to find another point at a distance r from the position point x1 in another direction along the manually marked coronary artery centerline, and obtain its reference direction and category probability; at this time, the probabilities of two direction categories in the reference distribution of the movement direction D are set to 0.5, and the probabilities of the remaining direction categories are set to 0.0;

[0066] Step 2-1-3: Randomly rotate the isotropic three-dimensional small blocks generated in step 2-1-1 and step 2-1-2 to expand the data set. Each isotropic three-dimensional small block is rotated around the x-axis, y-axis or z-axis at a random angle θ∈[0,2π] to generate three-dimensional small blocks in multiple directions;

[0067] Step 2-2: construct a coronary artery centerline extraction model and train it;

[0068] The coronary artery centerline extraction model includes three parallel three-dimensional hole convolutional neural networks, namely SeedCNN, OstiaCNN and TrackCNN; the network structures of the three three-dimensional hole convolutional neural networks are different only in the output part, and the input is a three-dimensional small block of size a×a×a. The last layer of the output part of SeedCNN performs regression to predict, and the predicted value is the proximity between the central voxel of the input three-dimensional small block and the coronary artery centerline; the last layer of the output part of OstiaCNN performs regression to predict, and the predicted value is the proximity between the central voxel of the input three-dimensional small block and the entrance of the coronary artery; the output part of TrackCNN includes two branches, the first branch uses classification to predict the direction category probability distribution of the central voxel of the input three-dimensional small block, and the second branch uses regression to predict the coronary radius at the central voxel of the input three-dimensional small block;

[0069] The SeedCNN is trained using the 3D small volume generated in step 2-1 and the proximity between the central voxel of the small volume and the manually marked coronary artery centerline;

[0070] OstiaCNN is trained using the 3D small volume generated in step 2-1 and the proximity between the central voxel of the small volume and the manually marked coronary artery entrance;

[0071] Train TrackCNN, and the training data are the three-dimensional small volume generated in step 2-1 and the reference direction of the central voxel of the three-dimensional small volume, as well as the category probability and the coronary artery radius r;

[0072] Step 2-3: If Figure 2 As shown, the coronary centerline extraction model is used to obtain the entry point and seed point;

[0073] Step 2-3-1: Preprocess the input CTA volume data, remove the lung area, and retain the heart part;

[0074] Step 2-3-2: randomly set multiple center points in the retained heart part, and use each center point as the central voxel to extract a 3D small block with a size of 19×19×19;

[0075] Step 2-3-3: Input the extracted 3D small volume blocks into the SeedCNN and OstiaCNN trained in step 2-2 respectively; obtain the predicted value of the proximity between the central voxel of the input 3D small volume block and the centerline of the coronary artery through SeedCNN, and take the central voxels corresponding to the first N values ​​of the predicted proximity value from near to far as seed points to form a seed point queue; obtain the predicted value of the proximity between the central voxel of the input 3D small volume block and the entrance of the coronary artery through OstiaCNN, and take the central voxels corresponding to the first M values ​​of the predicted proximity value from near to far as entry points to form an entry point queue;

[0076] Step 2-4: If Figure 3 As shown, TrackCNN combined with iterative tracking algorithm is used to simulate the centerline tracker to extract the centerline of a single coronary artery;

[0077] Step 2-4-1: randomly select a seed point from the seed point queue, extract a 3D small volume block with the seed point as the central voxel, input the extracted 3D small volume block into TrackCNN, and obtain the direction category probability distribution and coronary artery radius prediction value at the central voxel;

[0078] Step 2-4-2: Find two local maxima from the direction category probability distribution obtained in step 2-4-1 as the forward movement direction and reverse movement direction of the current central voxel. The angle between the forward movement direction and the reverse movement direction must be greater than 90 degrees.

[0079] Step 2-4-3: The tracker first tracks the center line along the forward motion direction. The tracker uses the predicted value of the coronary radius of the current central voxel as the step size and reaches the new position point along the forward motion direction of the current central voxel. Then, a three-dimensional small volume is extracted with the new position point as the central voxel, and the newly extracted three-dimensional small volume is input into the TrackCNN model to obtain the direction category probability distribution and radius prediction value at the new central voxel. Then, the local maximum value is found from the direction category probability distribution as the new tracking direction, and the angle between the new tracking direction and the initial forward motion direction is less than 60 degrees. Repeat this process until the forward motion direction tracking is terminated when the termination condition is met. The process of the tracker tracking the center line along the reverse motion direction is the same as the process of tracking the forward motion direction. The termination condition of the tracker is the normalized entropy of the direction classification probability distribution.

[0080] Connecting all the position points reached by the trackers during the tracking process will form the centerline of a single coronary artery;

[0081] Step 2-5: Generate coronary centerline tree;

[0082] Step 2-5-1: If the distance between the single coronary artery centerline generated in step 2-4 and any entry point in the entry point queue is less than or equal to the set value, the single coronary artery centerline is saved; otherwise, if the distance between the single coronary artery centerline generated in step 2-4 and all entry points in the entry point queue is greater than the set value, the single coronary artery centerline is discarded;

[0083] Step 2-5-2: remove the seed points in the seed point queue that overlap with the centerline of a single coronary artery from the seed point queue;

[0084] Step 2-6: Reselect a seed point from the seed point queue and repeat steps 2-4 and 2-5 until all seed points in the seed point queue are traversed and multiple coronary artery centerlines are extracted. Figure 5 As shown, a coronary centerline tree is formed.

[0085] Furthermore, in step 3, the selected coronary artery is straightened and reorganized along the center line of the coronary artery using a surface reconstruction method, and the reorganized data is cut at 0 degrees, 45 degrees and 90 degrees along the central axis to obtain multi-angle surface reconstruction results, and finally the obtained vascular cross-section is visualized.

[0086] A coronary artery intelligent auxiliary analysis system for CT images, comprising:

[0087] The medical imaging data module is used to select the MHD file or DICOM sequence containing CTA volume data through the file selector, obtain the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data, and perform visual display;

[0088] The centerline extraction module is used to automatically extract the coronary centerline based on the voxel, spatial spacing, and three-dimensional image size information of the acquired CTA volume data using a coronary centerline extraction method based on a three-dimensional hole convolutional neural network, generate a coronary centerline tree, and visualize it;

[0089] The surface reconstruction module is used to select the coronary artery to be analyzed from the coronary arteries corresponding to the coronary centerline extracted by the centerline extraction module, highlight the selected coronary artery in the three-dimensional display area of ​​the coronary centerline, and simultaneously perform multi-angle straightening and surface reconstruction on the selected coronary artery and perform visual display.

Claims

1. A method for intelligent auxiliary analysis of coronary arteries for CT images, characterized in that: The following steps are involved: Step 1: Select the MHD file or DICOM sequence containing CTA volume data through the file selector, obtain the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data, and perform visual display; Step 2: Based on the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data obtained in step 1, the coronary centerline extraction method based on the three-dimensional hole convolutional neural network is used to complete the fully automatic extraction of the coronary centerline, generate the coronary centerline tree and visualize it; the specific steps are as follows: Step 2-1: Data preprocessing; Step 2-2: construct a coronary artery centerline extraction model and train it; The coronary artery centerline extraction model includes three parallel three-dimensional atrous convolutional neural networks, namely SeedCNN, OstiaCNN and TrackCNN; the network structures of the three three-dimensional atrous convolutional neural networks differ only in the output part, and the input is a three-dimensional small block of size a×a×a. The last layer of the output part of SeedCNN performs regression for prediction, and the prediction value is the proximity between the central voxel of the input three-dimensional small block and the coronary artery centerline; The last layer of the output part of OstiaCNN performs regression to make predictions, and the predicted value is the proximity between the central voxel of the input 3D small volume and the entrance of the coronary artery. The output part of TrackCNN contains two branches. The first branch uses classification to predict the probability distribution of the direction category of the central voxel of the input 3D small volume, and the second branch uses regression to predict the coronary radius at the central voxel of the input 3D small volume. Step 2-3: Use SeedCNN and OstiaCNN extraction models to obtain the seed point queue and entry point queue; Step 2-4: Use TrackCNN combined with iterative tracking algorithm to simulate the centerline tracker to extract the centerline of a single coronary artery; Step 2-5: Generate coronary centerline tree; Step 2-5-1: If the distance between the single coronary artery centerline generated in step 2-4 and any entry point in the entry point queue is less than or equal to the set value, the single coronary artery centerline is saved; otherwise, if the distance between the single coronary artery centerline generated in step 2-4 and all entry points in the entry point queue is greater than the set value, the single coronary artery centerline is discarded; Step 2-5-2: remove the seed points in the seed point queue that overlap with the centerline of a single coronary artery from the seed point queue; Step 2-6: reselect a seed point from the seed point queue, and repeat steps 2-4 and 2-5 until all seed points in the seed point queue are traversed, and multiple coronary centerlines are extracted to form a coronary centerline tree; Step 3: Select the coronary artery to be analyzed from the coronary arteries corresponding to the coronary centerline extracted in step 2, highlight the selected coronary artery in the three-dimensional display area of ​​the coronary centerline, and perform multi-angle straightening and surface reconstruction on the selected coronary artery for visual display.

2. The method for intelligent auxiliary analysis of coronary arteries for CT images according to claim 1, characterized in that: When the DICOM sequence is selected in step 1, the grayscale value of the CTA volume data of the DICOM sequence is standardized to the grayscale value in the MHD format; then the file processing thread is started to obtain the voxel, spatial spacing, and three-dimensional image size information of the medical imaging data.

3. The intelligent auxiliary analysis method for coronary arteries based on CT images according to claim 1, characterized in that: The specific steps of data preprocessing in step 2 are as follows: Step 2-1-1: Manually mark the coronary artery centerline and the coronary artery entrance; randomly select a point x on the manually marked coronary artery centerline, and cut out an isotropic three-dimensional small volume on the CTA volume data with point x as the center voxel, and the size of the three-dimensional small volume is a×a×a; Assume that the radius of the coronary artery at point x is r, and a point is found along the manually marked coronary artery centerline at a distance r from point x, recorded as x_0, and the displacement vector between point x and x_0 is Δx; mark multiple points on the tracker sphere of the three-dimensional void convolutional neural network, and designate the direction with the smallest angle with the displacement vector Δx in the direction of the line connecting the center point x and multiple points as the reference direction, and set the designated reference direction category probability to 0.5 in the reference distribution of the potential movement direction D of point x; use the same method to find another point at a distance r in another direction from point x along the manually marked coronary artery centerline, and obtain its reference direction and category probability; at this time, the probabilities of two reference direction categories in the reference distribution of the movement direction D are set to 0.5, and the probabilities of the remaining direction categories are set to 0.0; Step 2-1-2: Randomly select a point x on the manually marked coronary artery centerline, use the random displacement d sampled in the three-dimensional normal distribution, translate the position point x by a distance d in any direction, record the translated point as x / , and cut out an isotropic three-dimensional small block on the CTA volume data with point x / as the central voxel. The size of the three-dimensional small block is a×a×a; Find the point x1 closest to point x / on the manually marked coronary artery centerline, set the coronary artery radius at point x1 to be r, find a point x_0 at a distance r from point x1 along the manually marked coronary artery centerline, and record the displacement vector between point x1 and x_0 as Δx; mark multiple points on the tracker sphere of the three-dimensional void convolutional neural network, designate the direction with the smallest angle with the displacement vector Δx in the direction of the line connecting the center point x1 and multiple points as the reference direction, and set the reference direction category probability to 0.5 in the reference distribution of the potential movement direction D of point x / ; use the same method to find another point at a distance r from the position point x1 in another direction along the manually marked coronary artery centerline, and obtain its reference direction and category probability; at this time, the probabilities of two direction categories in the reference distribution of the movement direction D are set to 0.5, and the probabilities of the remaining direction categories are set to 0.0; Step 2-1-3: Randomly rotate the isotropic three-dimensional small blocks generated in step 2-1-1 and step 2-1-2 to expand the data set. Each isotropic three-dimensional small block is rotated around the x-axis, y-axis or z-axis at a random angle θ∈[0, 2π] to generate three-dimensional small blocks in multiple directions; The specific steps of constructing the coronary artery centerline extraction model and training it are as follows: The SeedCNN is trained using the 3D small volume generated in step 2-1 and the proximity between the central voxel of the small volume and the manually marked coronary artery centerline; OstiaCNN is trained using the 3D small volume generated in step 2-1 and the proximity between the central voxel of the small volume and the manually marked coronary artery entrance; Train TrackCNN, and the training data are the three-dimensional small volume generated in step 2-1 and the reference direction of the central voxel of the three-dimensional small volume, as well as the category probability and the coronary artery radius r; The specific steps of using SeedCNN and OstiaCNN extraction models to obtain the seed point queue and entry point queue are as follows: Step 2-3-1: Preprocess the input CTA volume data, remove the lung area, and retain the heart part; Step 2-3-2: randomly set multiple center points in the retained heart part, and use each center point as the central voxel to extract a three-dimensional small block with a size of a×a×a; Step 2-3-3: Input the extracted 3D small volume blocks into the SeedCNN and OstiaCNN trained in step 2-2 respectively; obtain the predicted value of the proximity between the central voxel of the input 3D small volume block and the centerline of the coronary artery through SeedCNN, and take the central voxels corresponding to the first N values ​​of the predicted proximity value from near to far as seed points to form a seed point queue; obtain the predicted value of the proximity between the central voxel of the input 3D small volume block and the entrance of the coronary artery through OstiaCNN, and take the central voxels corresponding to the first M values ​​of the predicted proximity value from near to far as entry points to form an entry point queue; The specific steps of using TrackCNN combined with an iterative tracking algorithm to simulate a centerline tracker to extract a single coronary artery centerline are as follows: Step 2-4-1: randomly select a seed point from the seed point queue, extract a 3D small volume block with the seed point as the central voxel, input the extracted 3D small volume block into TrackCNN, and obtain the direction category probability distribution and coronary artery radius prediction value at the central voxel; Step 2-4-2: Find two local maxima from the direction category probability distribution obtained in step 2-4-1 as the forward movement direction and reverse movement direction of the current central voxel. The angle between the forward movement direction and the reverse movement direction must be greater than 90 degrees; Step 2-4-3: The tracker first tracks the center line along the forward motion direction. The tracker uses the predicted value of the coronary radius of the current central voxel as the step size and reaches the new position point along the forward motion direction of the current central voxel. Then, a three-dimensional small volume is extracted with the new position point as the central voxel, and the newly extracted three-dimensional small volume is input into the TrackCNN model to obtain the direction category probability distribution and radius prediction value at the new central voxel. Then, the local maximum value is found from the direction category probability distribution as the new tracking direction, and the angle between the new tracking direction and the initial forward motion direction is less than 60 degrees. Repeat this process until the forward motion direction tracking is terminated when the termination condition is met. The process of the tracker tracking the center line along the reverse motion direction is the same as the process of tracking the forward motion direction. The termination condition of the tracker is the normalized entropy of the direction classification probability distribution. Connecting all the position points reached by the trackers during the tracking process will form the centerline of a single coronary artery.

4. The method for intelligent auxiliary analysis of coronary arteries for CT images according to claim 3, characterized in that: In step 3, the selected coronary artery is straightened and reorganized along the centerline of the coronary artery using a surface reconstruction method, and the reorganized data is cut at 0 degrees, 45 degrees and 90 degrees along the central axis to obtain multi-angle surface reconstruction results, and finally the obtained vascular cross-section is visualized.

5. The method for intelligent auxiliary analysis of coronary arteries for CT images according to claim 4, characterized in that: Said a=19.

6. A coronary artery intelligent auxiliary analysis system for CT images, used to implement the coronary artery intelligent auxiliary analysis method for CT images according to any one of claims 1 to 5; characterized in that: include: The medical imaging data module is used to select the MHD file or DICOM sequence containing CTA volume data through the file selector, obtain the voxel, spatial spacing, and three-dimensional image size information of the CTA volume data, and perform visual display; The centerline extraction module is used to automatically extract the coronary centerline based on the voxel, spatial spacing, and three-dimensional image size information of the acquired CTA volume data using a coronary centerline extraction method based on a three-dimensional hole convolutional neural network, generate a coronary centerline tree, and visualize it; The surface reconstruction module is used to select the coronary artery to be analyzed from the coronary arteries corresponding to the coronary centerline extracted by the centerline extraction module, highlight the selected coronary artery in the three-dimensional display area of ​​the coronary centerline, and simultaneously perform multi-angle straightening and surface reconstruction on the selected coronary artery and perform visual display.

Images