A coronary point cloud data processing optimization method of a multi-task label model

Abstract

The application discloses a coronary point cloud data processing optimization method of a multi-task label model. A heart partition model which has been trained is used to segment a heart target region from a heart medical image; an import point position model which has been trained is used to search a coronary artery in a target region of an aorta to obtain an import point position of the coronary artery; the import point position and the heart medical image are combined with a center line tracking model to process and obtain a center line of the coronary artery; a deep learning coronary artery reconstruction algorithm is used to obtain new coronary point cloud data in combination with a coronary artery import point and the new coronary point cloud data is fed back to update and optimize. The application uses multi-task labels to completely realize multi-tasks of automatic heart medical image processing, overcomes the difficulty in manufacturing labels of multi-tasks, solves the problems of branch loss and adhesion in a coronary artery reconstruction process, and thus improves the accuracy of coronary artery reconstruction and also improves the efficiency.

Description

Technical Field

[0001] This invention relates to a method for optimizing cardiac medical image data processing, and more particularly to a method for optimizing coronary artery point cloud data processing using a multi-task labeling model. Background Technology

[0002] With the continuous development of medicine, medical imaging provides more intuitive evidence in the treatment of various diseases. Current technology usually relies on cardiac CTA data, and the three-dimensional reconstruction of tissues and organs in CTA data is crucial. Therefore, in the process of utilizing deep learning, the creation of labels will be particularly important.

[0003] The coronary arteries are the blood vessels that supply blood to the heart. Lesions in the coronary arteries can affect the function of the corresponding areas of the heart, and are a cause of coronary heart disease, angina pectoris, and myocardial infarction. Current techniques involved in cardiac CTA data include coronary artery reconstruction, coronary plaque analysis, coronary artery stenosis analysis, and cardiac segmentation, among which coronary artery reconstruction is particularly important. During the process of performing three-dimensional reconstruction of the coronary arteries, doctors more or less perform the reconstruction function semi-automatically.

[0004] Current technologies for cardiac CTA data research are mostly semi-automatic, relying heavily on the operator's experience. With the development of deep learning, the collection of sample datasets has become crucial, requiring the collection of corresponding labels under different task conditions. Currently, there is no fixed, standardized process for the collection and management of cardiac CTA datasets.

[0005] Moreover, existing technologies generally perform simple preprocessing (rotation, normalization, etc.) on CTA images and then use CNN networks to track the centerline radius and direction. They also use 3D Uet networks to directly segment the images, which can lead to problems such as broken or lost coronary artery branches. There may be issues such as missing branches or adhesions, resulting in the reconstructed coronary artery point cloud failing to meet the requirements of coronary artery tree labels. Summary of the Invention

[0006] To address the problems existing in the background technology, this invention proposes an optimization method for coronary artery point cloud data processing using a multi-task labeling model. This method solves problems such as small branch loss, atrial appendage adhesion, and pulmonary vein adhesion during coronary artery reconstruction, and overcomes the difficulty in creating multi-task labels.

[0007] Existing cardiac medical imaging data, when using excessive contrast agent, often leads to adhesion between the atrial appendage and coronary arteries during coronary remodeling. Conversely, in images with insufficient contrast agent, data showing deep coronary artery stenosis can result in coronary artery branch rupture. This invention effectively solves these problems.

[0008] The technical solution of the present invention is as follows: For a detailed flowchart of the method of this invention, please refer to [link / reference]. Figure 1 The flowchart of the multi-task label training model is as follows: Step 13 is to directly obtain coronary point cloud data by using a deep learning coronary artery reconstruction algorithm, given the original cardiac medical image, the location of the coronary artery inlet point and the centerline tracing.

[0009] Later, CTA-type cardiac medical images can be processed in batches for standardization. The system can directly utilize cardiac partitioning models to automatically obtain aortic tissue, left ventricle, left atrium, atrial appendage, and myocardium; directly utilize coronary artery inlet point location models to automatically search for the inlet points of the left and right coronary arteries; utilize deep learning coronary artery reconstruction algorithms to obtain coronary artery point cloud data; and utilize centerline extraction algorithms to automatically extract coronary artery tree labels. If the quality of various coronary artery tree labels is good, they can be added to the new label dataset to update the corresponding models and obtain new inlet point location models.

[0010] I. A method for optimizing coronary artery point cloud data processing using a multi-task labeling model, characterized in that: Step 1: Heart Division The target cardiac region is segmented from cardiac medical images using a pre-trained cardiac segmentation model. More specifically, labeled cardiac medical images are processed to obtain labeled cardiac region data. The labeled cardiac region data is then input into a labeling neural network to train a cardiac partitioning model. The cardiac target region is then segmented from the cardiac medical images using the cardiac partitioning model.

[0011] Step 2: Identification of the coronary artery inlet location: The inlet point location of the coronary arteries is obtained by searching the aortic tissue in the target area of ​​the heart using a pre-trained inlet point location model. Step 3: Obtain coronary artery point cloud data: The initial coronary point cloud data is obtained by inputting the inlet point location of the coronary artery obtained in step two into a traditional coronary artery reconstruction algorithm. Specifically, the location of the coronary artery inlet point can be input into a traditional coronary artery reconstruction algorithm to perform region growing, segment out the coronary artery point cloud, and obtain the coronary artery point cloud data through repair and preservation.

[0012] Step 4: Centerline Tracking Coronary artery tree label data is obtained from coronary artery point cloud data. The coronary artery tree label data is used to calculate the target Q value, and the target Q value is used to train the centerline tracking model. At the same time, the cardiac medical image is preprocessed to obtain image data as the first target region. Then, the location of the inlet point of the coronary artery and the first target region are input together into the trained centerline tracking model for prediction and processing to obtain the centerline of the coronary artery. Step 5, Deep Learning Coronary Artery Reconstruction: A special deep learning coronary artery reconstruction algorithm is used in conjunction with the coronary artery centerline obtained in Step 4 to obtain new coronary artery point cloud data.

[0013] Specifically, step one involves first importing cardiac medical images into annotation software to obtain cardiac regions such as aortic tissue, left coronary artery, right coronary artery, left ventricle, left atrium, auricle, and myocardium as first label data; then inputting the cardiac medical images and the first label data into a labeling neural network to train a cardiac partitioning model; and finally inputting the cardiac medical images to be processed into the trained cardiac partitioning model for segmentation and extraction to obtain the target cardiac regions of aortic tissue, left ventricle, left atrium, auricle, and myocardium.

[0014] The cardiac regions obtained by segmenting cardiac medical image data of unknown cardiac regions using the trained cardiac partitioning model do not include the left and right coronary arteries. The left and right coronary arteries are not segmented because the tissues of the left and right coronary arteries are small, and there are some adhesions and coronary artery ruptures. If they are directly segmented, the marked coronary arteries will have lost small branches.

[0015] The left and right coronary artery results obtained by predicting and segmenting cardiac medical image data of unknown cardiac regions using a trained cardiac partitioning model cannot be used as labels for coronary arteries. The present invention further utilizes the data containing the predicted segmentation results of the left and right coronary arteries, combined with information on aortic tissues and organs, to determine the location of the coronary artery inlet point.

[0016] Step two further includes: Based on the aortic tissue segmentation results obtained in step one, an image region near a specific aortic tissue is selected. Non-coronary artery features are removed using a double dilation and connected component labeling method to obtain sample feature points. A portion of the sample feature points are used as the center to extract an image block near the aortic tissue. The extracted image block is then normalized and used as input parameters. These input parameters are then input into the pre-trained inlet point location model to calculate the inlet point location of the coronary artery.

[0017] The output parameters of the inlet point location model are the left coronary inlet point and the right coronary inlet point. During training, the left coronary inlet point and the right coronary inlet point are obtained by processing the aortic tissue, left coronary, and right coronary data obtained through step one labeling, and are used as labels.

[0018] The aforementioned inlet point location model is used to automatically locate the inlet point of the coronary arteries, which includes the inlet point of the left coronary artery and the inlet point of the right coronary artery.

[0019] Step four further includes: In current technologies, coronary artery centerline tracking generally involves performing simple preprocessing (rotation, normalization, etc.) on CTA images and then using a CNN network to track the centerline radius and direction.

[0020] The method of obtaining coronary artery tree label data through coronary artery point cloud data includes: Using coronary artery point cloud data, coronary artery centerline data is extracted, and the extracted coronary artery centerline data is converted into coronary artery tree label data; The step of preprocessing the cardiac medical image to obtain image data as the first target region includes: Using cardiac medical images and the cardiac target region obtained from the segmentation in step one, the aortic tissue, left ventricle, left atrium, atrial appendage, and myocardium are removed. Then, taking advantage of the geometric characteristics that the coronary arteries are always close to the surface of the myocardium, the left and right coronary artery related regions are determined in the remaining regions. The left and right coronary artery related regions are a single region that includes the left and right coronary artery related regions. The image data of the left and right coronary artery related regions are uniformly planned with isometric spacing, and the data below the grayscale threshold are retained and normalized to obtain the first target region.

[0021] The centerline tracking model described uses a DQN reinforcement learning network, which consists of a backbone branch and a target branch.

[0022] Specifically, in step four, the centerline tracking model adopts a DQN reinforcement learning network, which includes a backbone branch and a target branch, and the topology of the backbone branch and the target branch is the same. Using the inlet point of the coronary artery obtained in step two as the initial coronary artery point, the centerline tracking model predicts the next coronary artery point position based on the initial coronary artery point. Then, a fixed-size image is selected from the first target region with the next coronary artery point position as the center. The fixed-size image is input into the trained centerline tracking model to obtain the centerline result of the coronary artery.

[0023] The target Q-value is an intermediate parameter used during the training of the DQN reinforcement learning network for the centerline tracking model.

[0024] In a more specific implementation, centerline tracking can be obtained by using multiple datasets and reinforcement learning networks to enhance the learning of centerline tracking.

[0025] The network parameters of the target branch do not need to be updated iteratively. Instead, they are copied from the current main branch at regular intervals, forming a delayed update.

[0026] Existing coronary artery reconstruction methods typically employ 3D Uet networks for direct image segmentation. However, due to the small size of coronary arteries, direct segmentation using 3D Uet networks often leads to issues such as broken or missing coronary artery branches. Consequently, existing coronary artery reconstruction algorithms also suffer from problems like lost branches or adhesions, resulting in reconstructed coronary artery point clouds that fail to meet the requirements for coronary artery number labeling.

[0027] This invention first utilizes a centerline tracing model to obtain the coronary artery centerlines, and then combines this with a region growing function to acquire coronary artery point cloud data, denoted as the deep learning coronary artery reconstruction method. The complete coronary artery point cloud data obtained through the deep learning coronary artery reconstruction method is then used to extract the coronary artery centerline data using a centerline extraction algorithm, transforming it into coronary artery tree label data. The centerline tracing model is then retrained to improve the accuracy of the final model, thereby increasing the precision of coronary artery reconstruction and resolving issues such as adhesion between the left atrial appendage and the coronary arteries.

[0028] In step five, the deep learning coronary artery reconstruction algorithm specifically includes: The centerline of the coronary artery is first expanded and binarized to obtain the connected domain range. Based on the connected domain range, non-coronary artery connected domains are extracted according to the inlet point position and excluded from the connected domain range. The connected domain ranges that are finally located within the centerline of the coronary artery are superimposed on the coronary artery and then traversed to obtain complete coronary artery point cloud data.

[0029] More specifically, in step five, the deep learning coronary artery reconstruction algorithm is as follows: S1. First, the centerline of the coronary artery is expanded K times by a sphere with a kernel of 3*3*3. The image range where the centerline is expanded is taken as the first range. The original medical image in the first range is then binarized with the cutoff gray level to obtain the second range. The original medical image data in the second range is processed by labeled connected components to obtain each connected component to form the third range. In the third range, the left coronary connected component and the right coronary connected component are extracted according to the left coronary inlet point and the right coronary inlet point obtained in step 2 as the fourth range. Specifically, the connected component closer to the left coronary inlet point is taken as the left coronary connected component, and the connected component closer to the right coronary inlet point is taken as the right coronary connected component. S2. Determine the number of connected components in the third range: If the third range contains more than two connected regions, then the connected regions in the third range excluding the fourth range are used as the fifth range; otherwise, all connected regions are coronary artery tissue and do not need to be removed, and the third range is used directly as the fifth range. S3. Determine if the fifth range is on the center line: If the fifth range is within the center line, then the fifth range belongs to the coronary artery tissue. The fifth range is superimposed on the fourth range of the coronary artery to obtain the sixth range. Traverse each point on the center line, find the point on the center line between the fifth range and the fourth range as a reference point, and superimpose the reference point on the sixth range to obtain the seventh range. The seventh range is used as the complete coronary artery point cloud data.

[0030] The coronary artery point cloud data is then fed back to the coronary artery point cloud data from step three for updating, and a new coronary artery centerline is obtained through calculation and processing using a centerline extraction algorithm. This iterative process continuously yields better coronary artery point cloud data, improving the success rate and accuracy of coronary artery reconstruction and resolving issues such as adhesion between the left atrial appendage and the coronary arteries.

[0031] Preferably, the cardiac medical image can be a CTA image, the cardiac partition model / labeling neural network can be a 3D Unet neural network, and the inlet point location model can be a ResNet10 neural network, but is not limited thereto.

[0032] The method further includes step six: feeding the new coronary artery point cloud data back into step four, obtaining coronary artery tree label data through the new coronary artery point cloud data, and retraining the centerline tracking model. Preferably, steps four to five can be repeated iteratively until a preset number of iterations is reached.

[0033] II. A coronary artery tree label data processing optimization system for a multi-task labeling model, comprising: The cardiac segmentation module uses a pre-trained cardiac segmentation model to segment the target cardiac region from cardiac medical images. The coronary artery inlet location module uses the target cardiac region obtained by the cardiac partitioning module as input to the trained inlet location model to search for coronary arteries in the target aortic region and obtain the inlet location of the coronary arteries. The centerline tracking module uses the results of the heart target region obtained by segmentation in the heart partitioning module, combined with the coronary artery point cloud data, and then inputs them into the trained centerline tracking model to obtain the centerline of the coronary artery. The deep learning coronary artery reconstruction module uses a deep learning coronary artery reconstruction algorithm combined with the processing of the inlet point location of the coronary artery to obtain new coronary artery point cloud data, which is then fed back to the centerline tracking module for further updates and optimizations.

[0034] In this invention, a multi-task labeling model is composed of a heart partitioning module, a coronary artery inlet location module, a centerline tracking module, and a deep learning coronary artery reconstruction module.

[0035] 3. A storage medium storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0036] Specifically, the computer program is an instruction that implements the above method.

[0037] The multi-task tags of the present invention include cardiac partition tags (aortic tissue, left ventricle, left atrium, left atrial appendage, left coronary artery, right coronary artery, myocardium), left and right coronary artery inlet point location tags, coronary artery point cloud tags, and coronary artery tree tags.

[0038] This invention automates data processing by importing case data, resulting in a relatively standardized and unified dataset that can meet multiple research tasks. It provides dataset support for future development and offers a technological direction for cardiac medical diagnosis.

[0039] The beneficial effects of this invention are: This invention utilizes multi-task tags to fully realize fully automatic cardiac zoning, automatic search for coronary artery inlets, coronary artery centerline tracking, coronary artery three-dimensional reconstruction, and so on.

[0040] This invention provides a fixed workflow that overcomes the difficulty of creating labels for multi-task applications.

[0041] This invention proposes a streamlined method for processing cardiac medical image data, which solves the problems of branch loss and adhesion during coronary artery reconstruction, thereby improving the accuracy of coronary artery reconstruction.

[0042] This invention optimizes the coronary artery reconstruction algorithm with fully automated multi-task label generation, which not only solves the problem of adhesion between atrial appendage tissue and coronary arteries, but also the problem of ruptured vascular branches. It also unifies the processing method of sample datasets and improves the efficiency of the algorithm. Attached Figure Description

[0043] Figure 1 This is a flowchart of data optimization processing under the multi-task label training model of the present invention; Figure 2 (a) is a schematic diagram of the region growth markers; Figure 2 (b) is a schematic diagram of the Spline markings; Figure 2 (c) is a schematic diagram of three-dimensional cardiac partitioning; Figure 3 This is a schematic diagram of the image after bone removal; Figure 4 This is a diagram illustrating the location of the coronary artery inlet point automatically. Figure 5 (a) is a schematic diagram of coronary artery point cloud; Figure 5 (b) is a schematic diagram of coronary artery point cloud superimposed on VR image; Figure 5 (c) is a schematic diagram of the coronary artery centerline; Figure 6 This is a schematic diagram of the coronary artery inlet points after removing some cardiac partitions. Figure 7 This is a schematic diagram of the region after the maximum slice dilation; Figure 8 This is a schematic diagram of the rotation and translation of the maximum slice; Figure 9 This is a schematic diagram showing the preservation of coronary artery regions in different slices; Figure 10 This is a schematic diagram of the flowchart for updating parameters in reinforcement learning; Figure 11 This is a schematic diagram of centerline tracking; Figure 12 (a) is a schematic diagram of the point cloud data of the main coronary artery and coronary artery branches; Figure 12 (b) is a schematic diagram of coronary artery point cloud data at an extremely narrow location; Figure 13 This is a schematic diagram of the coronary artery point cloud and the coronary artery centerline; Figure 14 This is a flowchart illustrating the overall steps of the method of the present invention. Detailed Implementation

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 14 As shown, the embodiments of the present invention are as follows: The original cardiac medical images are multiple sets of human cardiac CTA data, with a size of [N,C,L], where N represents the number of slices in the image, and C and L represent the image resolution, respectively. The CTA data also includes a series of imaging parameters, including the slice spacing and pixel spacing.

[0046] The network training involved in the examples all utilizes a unified multi-center dataset, totaling 740 datasets. This example uses a CTA dataset as an example, with a size of [224, 512, 512], an interlayer spacing of 0.625, and a pixel spacing of 0.378906. The specific implementation process is as follows: Heart regions.

[0047] The first step is to import the cardiac CTA image data into annotation software. The software uses region growing functionality to mark the aortic tissue, left coronary artery, right coronary artery, left ventricle, left atrium, and atrial appendage, such as... Figure 2 As shown in (a); the right coronary artery and left ventricular myocardium are marked using the Spline function, as shown. Figure 2 As shown in (b).

[0048] The second step involves using a 3D Unet neural network as the labeled neural network to train a model, obtaining a cardiac segmentation model, denoted as model1. The model's input parameters are [128,128,128,1], and its output parameters are [128,128,128,5]. Model1 segments the aorta, left ventricle (LV), left atrium (LA), atrial appendage (LAA), and myocardium. The segmentation result is denoted as the target cardiac region, such as... Figure 2 As shown in (c).

[0049] The third step, using the target area of ​​the heart as auxiliary information, involves removing excess vertebrae from different angles, preserving only the tissues and organs near the heart, such as... Figure 3 As shown, the heart tissue is visualized.

[0050] 2. Location of the coronary artery inlet.

[0051] The first step is to scale the original cardiac CTA images to ensure that the pixel spacing and slice spacing between cardiac CTA images are uniform to (0.5, 0.5, 0.5).

[0052] The second step is to select an area of ​​the image near the aortic tissue and use methods such as quadratic dilation and connected component labeling to remove non-coronary artery features to obtain sample feature points.

[0053] The third step is to use the ResNet10 neural network as the inlet point location model, which is denoted as model2.

[0054] Using a subset of feature points from sample feature point 1 as the center, nearby image patches are extracted and the data is normalized to serve as input parameters for the ResNet10 neural network. The aortic tissue, left coronary artery, and right coronary artery data labeled by software directly provide the entry points of the left and right coronary arteries, which can then be used as labels. Labels [1,0] represent locations that are not coronary artery entry points, and [0,1] represent locations that are coronary artery entry points. The labels are the output parameters of the ResNet10 neural network, where the model's input parameters are [64,64,64,1], and the output parameters are [n,2], where n is the number of non-zero index coordinates.

[0055] Fourthly, in this case, the ResNet10 neural network output values ​​are used to obtain the coordinates of the left and right inlet points of the coronary arteries using hierarchical clustering, such as... Figure 4 As shown, the ResNet10 neural network is used to automatically locate the coronary artery inlet point.

[0056] 3. Centerline tracking.

[0057] 3.1 Obtain coronary artery tree label data through coronary artery point cloud data.

[0058] The first step involves segmenting the coronary artery point cloud using the automatic coronary inlet point location and combining it with region growing in traditional coronary artery reconstruction algorithms. This segmented point cloud data is then manually repaired and preserved, and designated as the second target region. Figure 5 As shown in (a). Here, the second target area is overlaid on the VR display image, as shown. Figure 5 As shown in (b), some branches are missing.

[0059] The traditional coronary artery reconstruction algorithm mentioned here specifically refers to obtaining the coronary artery reconstruction algorithm by using region growing based on the location of the coronary artery inlet point. For details, please refer to the patent "CN 201910047390.0 An image processing method and device, storage medium".

[0060] The second step involves using coronary artery point cloud data and a centerline extraction algorithm. For details, please refer to the method proposed in patent CN202110278953.4, "A Method for Calculating the Radius of a Coronary Artery, Terminal, and Storage Medium," to extract the coronary artery centerline. Figure 5 As shown in (c), the coronary artery centerline data is stored as [NN, W], where NN is the number of center points on the coronary artery centerline, and W is the information of the center points on the coronary artery centerline, denoted as the first target point. This includes, but is not limited to, the following: coordinates in the pixel coordinate system, coordinates in the physical coordinate system, tangent direction of the centerline, curvature direction, radius of curvature, radius, traversal index of the centerline, etc.

[0061] The third step is to convert the coronary artery centerline data into coronary artery tree label data at equal intervals (0.5, 0.5, 0.5) along the X, Y, and Z axes, and record it as the second target point.

[0062] The coronary artery tree is stored in a dictionary format, with each segment of the coronary artery stored separately. The key is stored as the index of the start and end position, which in this example can be '(0,20)'. The value is stored as the coordinate of the center line position of the coronary artery segment. The center point of the coronary artery at the bifurcation position is both the end position of the previous coronary artery segment and the start point of the next coronary artery segment. From this, the coronary artery tree label data can be obtained.

[0063] 3.2 Preprocessing of raw CTA images.

[0064] Subsequently, a DQN reinforcement learning network was used as the centerline tracking model. During the DQN network reinforcement learning process, the original image CTA was preprocessed as follows, and the resulting image data was denoted as the first target region: First, remove the left ventricle, left atrium, and atrial appendage from the segmentation results of the cardiac partition model. Then, remove the aortic tissue at the location of the coronary artery inlet, such as... Figure 6 As shown.

[0065] Then, due to issues such as adhesions to the atrial appendages and pulmonary veins in the coronary arteries, these problems generally occur in the left coronary artery, rarely in the right. To address these issues, based on the characteristic that the left coronary artery always surrounds the surface of the myocardium, the relevant area of ​​the left coronary artery is pre-defined, which also significantly improves computation time. By maximizing the shared area of ​​the left ventricle and myocardium, the largest two-dimensional slice is determined. On this largest two-dimensional slice, pixels are dilated according to a certain threshold T (T = 50 in this case) to define a fixed region, such as... Figure 7 As shown, the fixed region is simultaneously updated to the three-dimensional space.

[0066] Secondly, through the above steps, the relevant region on the largest two-dimensional slice is obtained, denoted as Plane1. The centroid P0 is calculated, where P1 is the inlet point of the right coronary artery. By rotating it 90 degrees, the region with P2 as its centroid is obtained, denoted as Plane2. Then, Plane2 is translated to obtain the relevant region with P1 as its centroid, denoted as Plane3, as follows... Figure 8 As shown. Since the right coronary artery generally does not have excessive adhesions, the relevant area of ​​Plane3 can be expanded outward by a certain threshold pixel to ensure that the entire right coronary artery falls within the relevant area.

[0067] Finally, combining the above operations, the present invention can frame the coronary arteries in the heart within the following first target region, such as... Figure 9 As shown, the grayscale values ​​of the data in this area are kept between [-100, 700] and then normalized. The interlayer spacing and pixel spacing are then uniformly planned to [0.5, 0.5, 0.5].

[0068] In the image preprocessing process, the spatial features of the coronary arteries were combined to make corresponding constraints, while reducing the amount of computation and specifically tracking the center line of the coronary arteries.

[0069] 3.3 Model Construction.

[0070] The DQN reinforcement learning network is used as the topology of the centerline tracking model.

[0071] Using the coronary artery inlet point as the initial position, the system moves along a certain region (in this example, a 4x4x4 box) from the initial position to obtain the next point position. This is then updated with image data of size [19,19,19] selected from the first target region, centered at the initial position. This data serves as the input parameters for the Policy net (backbone branch) and Target net (target branch) of the DQN reinforcement learning network. Figure 10 As shown.

[0072] To illustrate the parameter update framework for reinforcement learning, given the current state S(t), the main branch, the Policy Net, predicts the Q-values ​​of different actions corresponding to the current state. The policy function then selects the actions that maximize the Q-value for transition. These different actions refer to the different directions the next point will fall in.

[0073] The state at the next time step is set to S(t+1). The Q value Q(t+1) corresponding to the next time step is calculated through the target branch Target net. Then, the loss is calculated and the main branch Policy net is updated.

[0074]

[0075] in, The reward value in the policy function. As a discount factor, For the current action, This is the current state. The next action, This is the next state.

[0076] The Q-value is calculated using the coronary tree label data obtained in 3.1 and used as the output parameter of the DQN reinforcement learning network. During the data training process of the DQN reinforcement learning network, the sample dataset is expanded by randomly rotating along the X, Y, and Z axes. Every once in a while, the target branch Target net is updated with the parameters of the main branch Policy net. After training multiple times, the centerline tracking model model3 can be obtained. This model model3 is used to track centerline data.

[0077] Deep learning coronary artery reconstruction algorithm.

[0078] The cardiac CTA data to be processed in this example, after the above data preprocessing process, directly obtains the centerline through the centerline tracking model 3, as follows: Figure 11 As shown, it can be observed that the first-order branches of the anterior descending branch and the branches on the right coronary trunk both emerge.

[0079] Using the traced centerline, coronary artery reconstruction is performed through region growth at a specific cutoff grayscale value (here, the cutoff grayscale value is 180). Often, there are cases of extremely narrowed coronary arteries, where the grayscale value is lower than the cutoff grayscale value, yet the coronary artery branches appear normal after the narrowing. For such cases, centerline tracing typically yields a centerline, but during the region growth process, it fails to grow the branches that have become narrowed.

[0080] To address this issue, this invention first expands the tracking centerline K times into a 3x3x3 sphere (K=10 in this example) within the deep learning coronary artery reconstruction algorithm. Essentially, the coronary arteries lie within the range defined by the expanded centerline; this region is denoted as the first range. Then, the original image within the first range is binarized using the cutoff grayscale, with the white portion designated as the second range. Within the second range, the data is labeled with different connected components using the `measure.label` function from the `skimage` library in Python, resulting in the third range.

[0081] At this point, the third region must contain at least two connected domains, one for the left coronary artery and the other for the right coronary artery, denoted as the fourth region. If the third region contains more than two connected domains, it is possible that coronary artery branch rupture may occur due to coronary artery stenosis, or the connected domain may not be tissue on the coronary artery. The left and right coronary artery connected domains are extracted based on the left and right coronary artery inlet points and used as the fourth region. The connected domains remaining after removing the fourth region from the third region are the fifth region.

[0082] To determine if the fifth region lies on the center line after tracking, it is usually within the center line. Since the fifth region belongs to the coronary artery tissue, it can be superimposed on the fourth region of the coronary artery and designated as the sixth region. Figure 12 As shown in (a).

[0083] Iterate through each point on the center line after tracing, find the point on the center line between the fifth and fourth ranges, and superimpose it onto the sixth range, denoted as the seventh range, as follows: Figure 12 As shown in (b).

[0084] The seventh region obtained at this point represents a complete coronary artery point cloud dataset. We then return to step 3 and use the centerline extraction algorithm to calculate the coronary artery geometric information, such as... Figure 13 As shown.

[0085] This improves the success rate and accuracy of coronary artery remodeling. This method also solves problems such as adhesions between the left atrial appendage and the coronary arteries.

[0086] In specific implementation, for other cardiac CTA image datasets in the future, data processing can be carried out according to the above-mentioned cardiac partitioning model, coronary artery inlet location model, centerline tracking model, traditional coronary artery reconstruction algorithm and deep learning coronary artery reconstruction algorithm. The data obtained in the later stage can meet the requirements of its corresponding label and serve as the training set for the network, which facilitates the collection of a larger dataset.

Claims

1. A method for optimizing coronary artery point cloud data processing using a multi-task labeling model, characterized in that: Step 1: Cardiac Segmentation: The trained cardiac segmentation model is used to segment the cardiac medical image to obtain the target cardiac region; Step 2, Coronary Artery Inlet Location Identification: Using the trained inlet location model, the aortic tissue in the target area of ​​the heart is searched to obtain the location of the coronary artery inlet. Step two includes: on the aortic tissue results obtained in step one, selecting the image region of the aortic tissue, removing non-coronary artery features using a double dilation and connected component labeling method to obtain sample feature points, using some feature points in the sample feature points as the center to extract image blocks near the aortic tissue, and normalizing the extracted image blocks as input parameters, and inputting the input parameters into the trained inlet point location model to calculate the inlet point location of the coronary artery; Step 3: Obtain coronary artery point cloud data: Use the inlet point location of the coronary artery obtained in Step 2 to input the coronary artery reconstruction algorithm to obtain coronary artery point cloud data; Step 4, Centerline Tracking: Coronary artery tree label data is obtained from coronary artery point cloud data. The coronary artery tree label data is used to calculate the target Q value, which is used to train the centerline tracking model. Simultaneously, the cardiac medical image is preprocessed to obtain image data as the first target region. Then, the location of the coronary artery inlet point and the first target region are input together into the trained centerline tracking model for prediction and processing to obtain the centerline of the coronary artery. Step 5, Deep Learning Coronary Artery Reconstruction: New coronary artery point cloud data is obtained by combining the deep learning coronary artery reconstruction algorithm with the coronary artery centerline obtained in Step 4. The deep learning coronary artery reconstruction algorithm is as follows: The centerline of the coronary artery is first expanded and binarized to obtain the connected domain range. Based on the connected domain range, non-coronary artery connected domains are extracted according to the inlet point position and excluded from the connected domain range. The connected domain ranges that are finally located within the centerline of the coronary artery are superimposed on the coronary artery and then traversed to obtain complete coronary artery point cloud data.

2. The method for optimizing coronary artery point cloud data processing using a multi-task labeling model according to claim 1, characterized in that: Specifically, step one involves first importing cardiac medical images into annotation software to obtain cardiac regions including aortic tissue, left coronary artery, right coronary artery, left ventricle, left atrium, auricle, and myocardium as first label data; then inputting the cardiac medical images and the first label data into a labeling neural network to train a cardiac partitioning model; and finally inputting the cardiac medical images to be processed into the trained cardiac partitioning model for segmentation to obtain the target cardiac regions including aortic tissue, left ventricle, left atrium, auricle, and myocardium.

3. The method for optimizing coronary artery point cloud data processing using a multi-task labeling model according to claim 1, characterized in that: The method of obtaining coronary artery tree label data through coronary artery point cloud data includes: Using coronary artery point cloud data, coronary artery centerline data is extracted, and the extracted coronary artery centerline data is converted into coronary artery tree label data; The step of preprocessing the cardiac medical image to obtain image data as the first target region includes: Using cardiac medical images and the cardiac target region obtained from the segmentation in step one, the cardiac regions of aortic tissue, left ventricle, left atrium, atrial appendage, and myocardium are removed. Then, taking advantage of the geometric characteristics that the coronary arteries are always close to the surface of the myocardium, the relevant regions of the left and right coronary arteries are determined in the remaining regions. The image data of the relevant regions of the left and right coronary arteries are uniformly planned with equiaxial spacing, and the data below the grayscale threshold are retained and normalized to obtain the first target region.

4. The method for optimizing coronary artery point cloud data processing using a multi-task labeling model according to claim 1 or 3, characterized in that: In step four, the centerline tracking model uses a DQN reinforcement learning network, which includes a backbone branch and a target branch, and the topology of the backbone branch and the target branch is the same. Using the inlet point of the coronary artery obtained in step two as the initial coronary artery point, the centerline tracking model predicts the next coronary artery point position based on the initial coronary artery point. Then, a fixed-size image is selected from the first target region with the next coronary artery point position as the center. The fixed-size image is input into the trained centerline tracking model to obtain the centerline result of the coronary artery.

5. The method for optimizing coronary artery point cloud data processing using a multi-task labeling model according to claim 1, characterized in that: The method further includes the following steps: Step 6: Feed the new coronary artery point cloud data back into Step 4. Obtain coronary artery tree label data through the new coronary artery point cloud data and retrain the centerline tracking model.

6. A coronary artery tree label data processing and optimization system for a multi-task label model applied to the method of any one of claims 1 to 5, characterized in that... include: The cardiac segmentation module uses a trained cardiac segmentation model to segment cardiac medical images to obtain the target cardiac region. The coronary artery inlet location module uses the target cardiac region obtained by the cardiac partitioning module as input to the trained inlet location model to search for coronary arteries in the target aortic region and obtain the inlet location of the coronary arteries. The centerline tracking module uses the results of the heart target region obtained by segmentation in the heart partitioning module, combined with the coronary artery point cloud data, and then inputs them into the centerline tracking model to obtain the centerline of the coronary artery. The deep learning coronary artery reconstruction module uses a deep learning coronary artery reconstruction algorithm combined with coronary artery centerline processing to obtain new coronary artery point cloud data, which is then fed back to the centerline tracking module for further updates and optimizations.

7. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the method described in any one of claims 1 to 5.

8. A storage medium according to claim 7, characterized in that, The computer program thereunder is an instruction that implements the method of any one of claims 1 to 5.

Images