Coronary center point cloud extraction method, training method and related products

Abstract

The application provides a coronary center point cloud extraction method, a training method and related products. In the extraction method, three-dimensional CCTA volume data of a coronary artery is resampled along multiple spatial planes to obtain multiple two-dimensional slices; the two-dimensional slices are input into a single-stage multi-task network model, one task branch of which obtains a bounding box and a type label of each independent blood vessel region in the two-dimensional slice, and another task branch obtains a geometric center of each independent blood vessel region; the geometric center of each independent blood vessel region is spatially transformed to the three-dimensional CCTA volume data space of the coronary artery to obtain an initial point cloud; the initial point cloud is denoised to obtain a denoised point cloud, and the three types of points, i.e., a left coronary root node, a right coronary root node and a non-root node, in the denoised point cloud are independently subjected to spatial down-sampling to obtain a down-sampled point cloud, and the only left coronary root node and the only right coronary root node are determined in the down-sampled point cloud. The method has a strong ability to suppress artifacts and can obtain accurate root nodes.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method for extracting coronary artery center point clouds, a training method, and related products. Background Technology

[0002] Coronary computed tomography angiography (CCTA) is a core tool for diagnosing coronary artery disease. Obtaining a high-precision three-dimensional coronary artery center point cloud can provide valuable information for various subsequent analytical tasks. For example, the coronary artery center point cloud can be used to construct a coronary artery centerline tree. However, existing techniques for extracting coronary artery center point clouds are limited, and artifacts in CCTA volumetric data can significantly interfere with center point cloud extraction. Summary of the Invention

[0003] This invention provides a method, training method, and related products for extracting coronary artery center point clouds, so as to provide high-quality coronary artery center point clouds and extract accurate left and right coronary artery root nodes.

[0004] This invention provides the following technical solution: a method for extracting coronary artery center point clouds, comprising: The three-dimensional CCTA volume data of the coronary arteries were sliced ​​and resampled along multiple spatial planes to obtain multiple two-dimensional slices; The two-dimensional slice is input into a single-stage multi-task network model. One task branch of the single-stage multi-task network model obtains the bounding box and type label of each independent vascular region in the two-dimensional slice. The type label includes the left coronary root node, the right coronary root node and the non-root node. The other task branch of the single-stage multi-task network model obtains the geometric center of each independent vascular region. Transform the geometric center space of each independent vascular region to the three-dimensional CCTA volume data space of the coronary artery to obtain the initial point cloud; The initial point cloud is denoised to obtain a denoised point cloud. The three types of points in the denoised point cloud—left coronary root node, right coronary root node, and non-root node—are independently spatially downsampled to obtain a downsampled point cloud. A unique left coronary root node and a unique right coronary root node are determined in the downsampled point cloud.

[0005] In some implementations, the step of slicing and resampling the coronary three-dimensional CCTA volume data along multiple spatial planes includes: a horizontal plane, a coronal plane, and a sagittal plane.

[0006] More preferably, the plurality of spatial planes further include oblique sectional planes that are different from the horizontal plane, the coronal plane, and the sagittal plane.

[0007] In some implementations, the single-stage multi-task network model is selected from the YOLO model, RTMO model, FCOS model, CenterNet model, and CenterPose model.

[0008] In some implementations, the noise reduction of the initial point cloud includes: Calculate the Euclidean distance between any point in the initial point cloud and its surrounding points, and define two points whose Euclidean distance is less than a first set distance as adjacent points to obtain multiple local clusters; If the number of points in the local cluster is less than the set threshold number, then all points in the local cluster are determined to be noise points, and the local cluster is removed from the initial point cloud to obtain the denoised point cloud.

[0009] In some implementations, the step of independently performing spatial downsampling on the three types of points in the denoised point cloud—left coronary artery root node, right coronary artery root node, and non-root nodes—includes: The physical space where the noise-reduced point cloud is located is divided into voxel grids. The centroid of a single type of point in any single voxel grid is calculated. The point closest to the centroid is selected and retained. The remaining points in the voxel grid are deleted to obtain a set of downsampling points of the corresponding type. The single type is any one of the three types: left coronary root node, right coronary root node and non-root node. The downsampling point cloud is obtained by combining the sets of downsampling points of the three types: left coronary root node, right coronary root node, and non-root node.

[0010] In some implementations, determining a unique left coronary root node and a unique right coronary root node in the downsampling point cloud includes: By combining any left coronary artery root node and any right coronary artery root node in the downsampled point cloud, multiple root node pairs are obtained. Candidate root node pairs are selected from the plurality of root node pairs based on spatial distance constraints, lateral position constraints, and front-back position constraints. The unique left coronary artery root node and the unique right coronary artery root node are determined by the sum of the network prediction confidence of the candidate root node pairs, or by clustering the downsampled point cloud to obtain multiple local clusters, and the unique left coronary artery root node and the unique right coronary artery root node are determined by the sum of the number of points in the local clusters to which the candidate root node pairs belong.

[0011] This invention provides the following technical solution: a training method for a single-stage multi-task network model, comprising: A training dataset is provided, wherein the samples are multi-view two-dimensional slice samples of coronary artery three-dimensional CCTA volume data samples, and the corresponding labels are the bounding box, type label and geometric center of each independent vascular region in the two-dimensional slice sample. The type labels include left coronary artery root node, right coronary artery root node and non-root node. Some samples in the training dataset contain multiple levels of artifacts. The single-stage multi-task network model is trained based on the training dataset. One task branch of the single-stage multi-task network model obtains the bounding box and type label of each independent blood vessel region in the two-dimensional slice sample, and the other task branch of the single-stage multi-task network model obtains the geometric center of each independent blood vessel region.

[0012] The training dataset contains samples with varying degrees of artifacts, meaning some samples have milder artifacts while others have more severe ones. The artifacts can be classified as level 2, level 3, level 4, etc.

[0013] In some implementations, during the step of providing the training dataset, the image region containing the left coronary root node or containing the right coronary root node in the two-dimensional slice sample is copied and merged into other two-dimensional slice samples that do not contain the left coronary root node or the right coronary root node, to obtain a new two-dimensional slice sample.

[0014] The present invention provides the following technical solution: a computing device, including a memory and a processor, wherein the memory stores a program, and the processor runs the program to execute the above-described coronary artery center point cloud extraction method, or to execute the above-described single-stage multi-task network model training method.

[0015] The present invention provides the following technical solution: a program product, which executes the above-mentioned coronary artery center point cloud extraction method when running on a processor.

[0016] The technical solution of the present invention has the following advantages.

[0017] Resampling the 3D CCTA volumetric data of the coronary arteries along multiple spatial planes is equivalent to observing the 3D CCTA volumetric data from multiple perspectives. Artifacts are more pronounced at some perspectives and less so at others. Therefore, multi-view slice resampling can reduce blind spots and suppress the interference of artifacts on the extraction of the center point cloud. Furthermore, multi-view slices offer sufficient spatial representation and ensure strong continuity in the center point cloud extraction process.

[0018] Bounding boxes are crucial data in the training phase of single-stage multi-task network models. One function is to guide attention, allowing the network to first learn "where the blood vessels are roughly located," and then precisely pinpoint "where the center point is." The two task branches of the single-stage multi-task network model are named the bounding box branch and the center point branch, respectively. A second function is multi-task collaborative learning; the bounding box branch and the center point branch share a common backbone network, and the location priors provided by the bounding box branch can help the center point branch converge faster.

[0019] Spatial downsampling enables a more uniform distribution of multiple left coronary artery root nodes, multiple right coronary artery root nodes, and multiple non-root nodes in the downsampled point cloud, facilitating subsequent tasks. The coronary artery center point cloud extraction method not only extracts the coronary artery center point cloud but also simultaneously identifies a unique left coronary artery root node and a unique right coronary artery root node. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the coronary artery center point cloud extraction method according to an embodiment of the present invention.

[0021] Figure 2 This is a diagram illustrating key information in the coronary artery center point cloud extraction method of this invention.

[0022] Figure 3 This is a flowchart illustrating the training method of a single-stage multi-task network model according to an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the structure of a computing device according to an embodiment of the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0025] refer to Figure 1 The present invention provides a method for extracting coronary artery center point cloud, comprising the following steps.

[0026] S1. The three-dimensional CCTA volume data of the coronary artery is sliced ​​and resampled along multiple spatial planes to obtain multiple two-dimensional slices.

[0027] The spatial planes can be arranged in as many directions as possible to suppress the influence of artifacts. In some embodiments, the step of slicing and resampling the coronary artery 3D CCTA volume data along multiple spatial planes includes: a horizontal plane, a coronal plane, and a sagittal plane. More preferably, the multiple spatial planes also include oblique sections different from the horizontal plane, the coronal plane, and the sagittal plane.

[0028] S2. Input the two-dimensional slice into a single-stage multi-task network model. One task branch of the single-stage multi-task network model obtains the bounding box and type label of each independent vascular region in the two-dimensional slice. The type label includes the left coronary root node, the right coronary root node, and the non-root node. Another task branch of the single-stage multi-task network model obtains the geometric center of each independent vascular region.

[0029] Multiple 2D slices from multiple perspectives input concurrently are encapsulated into variable-length tensors as batch input data. A single-stage multi-task network model performs underlying tensor batch processing inference. This mechanism of deep collaboration between multi-pipeline concurrency and dynamic tensor batch processing effectively eliminates the blind spots in traditional single-slice object detection and improves the spatial representation capability and overall inference efficiency for complex 3D vascular morphology by fully utilizing the parallel computing throughput of underlying hardware (such as GPUs).

[0030] In some implementations, the single-stage multi-task network model is selected from the YOLO model, RTMO model, FCOS model, CenterNet model, and CenterPose model.

[0031] S3. Transform the geometric center space of each independent vascular region to the three-dimensional CCTA volume data space of the coronary artery to obtain the initial point cloud.

[0032] The spatial geometric transformation matrix corresponding to each 2D slice in the original 3D volume data is called to perform inverse spatial mapping, accurately and uniformly mapping the coordinates of the set center of the vascular region in the discrete 2D slices back to the coordinate system of the original 3D CCTA volume data space, generating a high-density initial point cloud.

[0033] S4. Denoise the initial point cloud to obtain a denoised point cloud. Independently perform spatial downsampling on the three types of points in the denoised point cloud: left coronary root node, right coronary root node, and non-root node, to obtain a downsampled point cloud. In the downsampled point cloud, determine a unique left coronary root node and a unique right coronary root node.

[0034] Non-root nodes refer to the center points of the coronary artery pathway other than the left and right coronary artery root nodes.

[0035] For example, refer to Figure 2 The coronary artery center point cloud extraction method ultimately yields the coronary artery center point cloud, as well as a unique left coronary artery root node and a unique right coronary artery root node, providing a reliable starting anchor point for subsequent tasks.

[0036] In some implementations, the noise reduction of the initial point cloud includes: Calculate the Euclidean distance between any point in the initial point cloud and its surrounding points, and define two points whose Euclidean distance is less than a first set distance as adjacent points to obtain multiple local clusters; If the number of points in the local cluster is less than the set threshold number, then all points in the local cluster are determined to be noise points, and the local cluster is removed from the initial point cloud to obtain the denoised point cloud.

[0037] The first set distance can be, for example, 2mm, 2.5mm, or 3mm, and the number of set threshold points can be, for example, 3 or 4. Setting the first set distance too small or too large is not conducive to filtering noise points. Setting the number of set threshold points too small may result in noise points being missed, while setting it too large may lead to the accidental deletion of non-noise points.

[0038] The noise reduction operation removes isolated noise points.

[0039] In some implementations, the step of independently performing spatial downsampling on the three types of points in the denoised point cloud—left coronary artery root node, right coronary artery root node, and non-root nodes—includes: The physical space where the noise-reduced point cloud is located is divided into voxel grids. The centroid of a single type of point in any single voxel grid is calculated. The point closest to the centroid is selected and retained. The remaining points in the voxel grid are deleted to obtain a set of downsampling points of the corresponding type. The single type is any one of the three types: left coronary root node, right coronary root node and non-root node. The downsampling point cloud is obtained by combining the sets of downsampling points of the three types: left coronary root node, right coronary root node, and non-root node.

[0040] The voxel mesh is, for example, a cube. The size of the voxel mesh essentially corresponds to the local merging scale in the 3D point cloud space. The lower limit of the voxel mesh side length is 0.7 mm, and the upper limit is 2.4 mm. The preferred side length is 1.3 mm to 1.7 mm, more preferably 1.33 mm to 1.37 mm. This setting avoids excessive density and repetition in local point clouds while preserving the basic topology and bifurcation morphology of the main coronary artery and its proximal branches.

[0041] After spatial downsampling, the number of root nodes in the left coronary artery decreased and their distribution became more uniform, the number of root nodes in the right coronary artery decreased and their distribution became more uniform, and the number of non-root nodes decreased and their distribution became more uniform.

[0042] Note that if the three types of points are not distinguished and a single spatial downsampling is performed, the information of two of the three types of points within a voxel grid will be lost.

[0043] In some implementations, determining a unique left coronary root node and a unique right coronary root node in the downsampling point cloud includes: By combining any left coronary artery root node and any right coronary artery root node in the downsampled point cloud, multiple root node pairs are obtained. Candidate root node pairs are selected from the plurality of root node pairs based on spatial distance constraints, lateral position constraints, and front-back position constraints. The unique left coronary artery root node and the unique right coronary artery root node are determined by the sum of the network prediction confidence of the candidate root node pairs. Alternatively, a density-based spatial clustering algorithm is performed on the downsampled point cloud to obtain multiple local clusters. The unique left coronary artery root node and the unique right coronary artery root node are determined by the sum of the number of points in the local clusters to which the candidate root node pairs belong.

[0044] Constraint 1 (Spatial distance constraint): The Euclidean distance between the left and right candidate root nodes in the same pair in three-dimensional space must be greater than or equal to the set physiological threshold (e.g., 10 mm).

[0045] Constraints 2 and 3 are described in the LPS coordinate system. In the LPS coordinate system, the positive X-axis points from the patient's right side to the patient's left side, the positive Y-axis points from the patient's front side to the patient's back side, and the positive Z-axis points from the patient's foot side to the patient's head side.

[0046] Constraint 2 (Lateral Position Constraint): The X-axis coordinate of the right coronary root node is sufficiently large, and the X-axis coordinate of the left coronary root node is sufficiently small. A criterion for determining that the X-axis coordinate of the right coronary root node is sufficiently large is, for example, that it is greater than the mean or median of the X-axis coordinates of all right coronary root nodes. A criterion for determining that the X-axis coordinate of the left coronary root node is sufficiently small is, for example, that it is less than the mean or median of the X-axis coordinates of all left coronary root nodes. The purpose of Constraint 2 is to bring the two root nodes as close as possible to the ascending aorta.

[0047] Constraint 3 (Front-back Position Constraint): Compare the front and back coordinate values ​​of the candidate nodes, and restrict the physical coordinates of the left coronary root node to be strictly located in the front direction of the right coronary root node, and the Y-axis coordinate value of the left coronary root node is less than the Y-axis coordinate value of the right coronary root node.

[0048] If multiple candidate root node pairs still exist after filtering by the above three constraints, then one of the following two methods will be used to select a unique left coronary artery root node and a unique right coronary artery root node.

[0049] Method 1: Select the candidate root node pair with the largest sum of network prediction confidence as the unique left coronary artery root node and the unique right coronary artery root node.

[0050] Network prediction confidence is the confidence level of a single-stage multi-task network model when it predicts that a point is either the root node of the left coronary artery or the root node of the right coronary artery.

[0051] Method 2: Cluster the downsampled point cloud to obtain multiple local clusters, and determine the unique left coronary artery root node and the unique right coronary artery root node based on the sum of the number of points in the local cluster to which the candidate root node belongs.

[0052] The candidate root node pair with the largest sum of points in the two local clusters can be selected as the unique left coronary artery root node and the unique right coronary artery root node.

[0053] The clustering operation specifically involves calculating the Euclidean distance between any point in the downsampled point cloud and its surrounding points, defining two points whose Euclidean distance is less than a second set distance as adjacent points, and obtaining multiple local clusters.

[0054] The recommended second distance setting is 1 to 2 times the diagonal length of the voxel mesh. A second distance that is too small or too large will hinder the differentiation of candidate root nodes.

[0055] The coronary artery center point cloud extraction method provided by this invention has the following advantages.

[0056] The multi-view fusion mechanism compensates for the blind spots and insufficient spatial representation that occur at specific angles, ensuring the integrity of the localization of each branch of the coronary tree.

[0057] Benefiting from the high efficiency of single-stage target detection networks (such as the YOLO architecture) and dual parallel processing mechanism, the efficiency of coronary post-processing is significantly improved while reducing memory overhead, supporting rapid clinical diagnosis.

[0058] The simultaneous location of the root node and the blood vessel center provides strong logical support for the subsequent construction of the blood vessel tree topology and reduces the misjudgment rate.

[0059] refer to Figure 3 Embodiments of the present invention also provide a training method for a single-stage multi-task network model. The trained single-stage multi-task network model can be used for the above-mentioned coronary artery center point cloud extraction method. To improve robustness to artifacts, samples containing artifacts of different levels are added during training, and targeted data augmentation is performed. The generation process of two-dimensional slice samples also aims to provide as many perspectives as possible.

[0060] The training method includes the following steps.

[0061] S10. Provide a training dataset, wherein the samples are multi-view two-dimensional slice samples of coronary three-dimensional CCTA volume data samples, and the corresponding labels are the bounding box, type label and geometric center of each independent vascular region in the two-dimensional slice sample. The type label includes the left coronary root node, the right coronary root node and non-root node. Some samples in the training dataset contain multiple levels of artifacts.

[0062] Multi-view slicing can be performed on 3D CCTA volumetric data of coronary arteries with annotated 3D coronary artery regions to obtain multi-view 2D slice samples and the coronary vessel regions within these 2D slice samples. This allows for the generation and annotation of the bounding boxes and geometric centers of the coronary vessel regions. The centers of the coronary ostia in the 3D CCTA volumetric data can be manually annotated or already annotated within the 3D CCTA volumetric data.

[0063] The plane in which the two-dimensional slice sample is located should have as many orientations as possible, including, for example, the horizontal plane, the coronal plane, the sagittal plane, and oblique planes that are different from the horizontal plane, the coronal plane, and the sagittal plane.

[0064] The training dataset includes a subset of separately provided 2D slice samples containing artifacts of varying severity. Coronary artery 3D CCTA volumetric data samples containing artifacts of different degrees and morphologies (such as typical crescent-shaped and bird-wing-shaped motion artifacts) are selectively mixed at a predetermined ratio. These 3D CCTA volumetric data samples are then sliced ​​into 2D slices to obtain 2D slice samples containing artifacts. This forces the model to consider both artifact noise and real vascular structure during feature extraction. The artifact grading can be assessed by medical imaging experts.

[0065] The ratio of the number of 2D slice samples containing artifacts to the total number of samples is recommended to be 0.45 to 0.6. In a preferred embodiment, the ratio of the number of 2D slice samples containing artifacts to the total number of samples is 0.5.

[0066] In two-dimensional slice samples containing artifacts, the higher the degree of artifacts, the fewer the corresponding two-dimensional slice samples.

[0067] In one embodiment, the ratio of the number of samples in the four categories—no artifacts, mild artifacts, moderate artifacts, and severe artifacts—is 50%:25%:15%:10%. This allows the model to learn how to remove the influence of artifacts while avoiding overfitting to samples with severe artifacts.

[0068] In the step of providing the training dataset, the image region containing the left coronary root node or the right coronary root node in the two-dimensional slice is copied and fused into other two-dimensional slices that do not contain the left coronary root node or the right coronary root node, to obtain a new two-dimensional slice sample.

[0069] Because root node samples in the training dataset are extremely scarce compared to typical vascular pathway point samples, this invention abandons the traditional uniform sampling method and instead performs targeted data compensation for root node samples. By accurately extracting scarce root node instances (i.e., images of the vascular region where the root node is located in a 2D slice), and randomly copying and fusing them across background constraints into other images with different vascular topologies, new 2D slice samples are obtained. This mechanism not only directly breaks down the barrier of extreme class imbalance at the data distribution level, but also artificially constructs a new context combination of "scarce anatomical nodes and diverse complex backgrounds," forcing the model to break free from its dependence on specific surrounding pixel environments, thereby significantly improving the network's high-sensitivity recognition and recall lower bound for key long-tail targets.

[0070] Data augmentation methods for training datasets may also include the following operations.

[0071] Enhanced Omnidirectional Spatial Transformation: Addressing the differences in the physiological orientation of blood vessels among different patients and the randomness of scanning angles in imaging equipment, this invention performs extensive nonlinear geometric transformations (including omnidirectional rotation, large-scale translation, flipping, and scaling) on ​​two-dimensional slice samples to obtain new two-dimensional slice samples. This forces the model to truly learn and extract the intrinsic morphology and orientation features of blood vessels, independent of specific local scanning angles and positions.

[0072] Multi-scale adaptive training: This design addresses the practical challenges of varying image sizes in clinical reconstructions and the multi-scale geometric features of vascular trees, characterized by "thick trunks and thin distal branches." In practice, the resolution and scaling of the input images are dynamically and randomly adjusted during training iterations (covering multiple resolution ranges that the target application device may output), forcing the model to extract features within a varying receptive field. This mechanism significantly enhances the model's strong generalization and adaptation capabilities to different input resolutions and reconstructed image sizes. Furthermore, it enables the model to both sensitively capture weak signals from distal, fine branches at high resolution to prevent missed detections and establish a global spatial perception of the overall structure of the main vascular trunk at low resolution.

[0073] Brightness Enhancement: Addressing real-world physical differences in clinical scanning imaging, this feature specifically simulates inconsistent lumen brightness or overall image contrast attenuation caused by uneven contrast agent filling, equipment voltage fluctuations, or individual patient differences. During model iteration, the grayscale of 2D slide samples is dynamically adjusted. This can involve uniform linear or non-linear transformations of the grayscale of the 2D slide samples, or linear or non-linear transformations of the grayscale of localized areas of the 2D slide samples. This enables the model to learn to extract robust features under complex and non-uniform brightness distributions.

[0074] All three data augmentation techniques described above can be implemented dynamically during model iteration. That is, the samples used in each iteration of the model can be different from those used in the previous iteration.

[0075] S20. The single-stage multi-task network model is trained according to the training dataset. One task branch of the single-stage multi-task network model obtains the bounding box and type label of each independent blood vessel region in the two-dimensional slice sample, and the other task branch of the single-stage multi-task network model obtains the geometric center of each independent blood vessel region.

[0076] In a specific example, in a single-stage multi-task network model, the backbone feature extraction network extracts features from the two-dimensional slice sample, then inputs the extracted features into a multi-scale feature fusion network for feature fusion, and finally inputs the feature fusion result into a multi-task decoupled detection head. One task branch in the multi-task decoupled detection head is responsible for detecting and generating bounding boxes and type labels for each independent blood vessel region in the two-dimensional slice sample, while another task branch is responsible for calculating and outputting the two-dimensional pixel coordinates of the geometric center (lumen center point) of the blood vessel cross-section within the bounding box.

[0077] Based on the same inventive concept, and referring to Figure 4 Embodiments of the present invention also provide a computing device, including at least one memory and at least one processor, wherein the memory stores a program, and the processor runs the program to execute the above-described coronary artery center point cloud extraction method, or to execute the above-described training method for a single-stage multi-task network model.

[0078] Based on the same inventive concept, embodiments of the present invention also provide a program product that executes the above-described coronary artery center point cloud extraction method when running on a processor.

[0079] It should be noted that the processors mentioned above may include, but are not limited to, one or more processors or microprocessors. Each processor may be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor (MCU), central processing unit (CPU), graphics processing unit (GPU), data processing unit (DPU), or other electronic components to execute the methods in the above embodiments.

[0080] The above-mentioned storage devices can be various media capable of storing program code, such as USB flash drives, hard disk drives, solid-state drives, portable hard drives, read-only memory (ROM), random access memory (RAM), flash memory, EPROM, EEPROM, registers, magnetic disks, or optical disks. The storage device can be volatile, non-volatile, or a combination of both.

[0081] In one example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in a computing device.

[0082] This invention is not limited to the above-described embodiments. Any modifications, improvements, or substitutions that can be conceived by those skilled in the art without departing from the essential content of this invention fall within the scope of this invention.

Claims

1. A method for extracting coronary artery center point clouds, characterized in that, include: The three-dimensional CCTA volume data of the coronary arteries were sliced ​​and resampled along multiple spatial planes to obtain multiple two-dimensional slices; The two-dimensional slice is input into a single-stage multi-task network model. One task branch of the single-stage multi-task network model obtains the bounding box and type label of each independent vascular region in the two-dimensional slice. The type label includes the left coronary root node, the right coronary root node and the non-root node. The other task branch of the single-stage multi-task network model obtains the geometric center of each independent vascular region. Transform the geometric center space of each independent vascular region to the three-dimensional CCTA volume data space of the coronary artery to obtain the initial point cloud; The initial point cloud is denoised to obtain a denoised point cloud. The three types of points in the denoised point cloud—left coronary root node, right coronary root node, and non-root node—are independently spatially downsampled to obtain a downsampled point cloud. A unique left coronary root node and a unique right coronary root node are determined in the downsampled point cloud.

2. The method for extracting coronary artery center point clouds according to claim 1, characterized in that, In the step of resampling the three-dimensional CCTA volume data of the coronary arteries along multiple spatial planes, the multiple spatial planes include: the horizontal plane, the coronal plane, and the sagittal plane.

3. The method for extracting coronary artery center point clouds according to claim 1, characterized in that, The single-stage multi-task network model is selected from the YOLO model, RTMO model, FCOS model, CenterNet model, and CenterPose model.

4. The method for extracting coronary artery center point clouds according to claim 1, characterized in that, The noise reduction of the initial point cloud includes: Calculate the Euclidean distance between any point in the initial point cloud and its surrounding points, and define two points whose Euclidean distance is less than a first set distance as adjacent points to obtain multiple local clusters; If the number of points in the local cluster is less than the set threshold number, then all points in the local cluster are determined to be noise points, and the local cluster is removed from the initial point cloud to obtain the denoised point cloud.

5. The method for extracting coronary artery center point clouds according to claim 1, characterized in that, The independent spatial downsampling of the three types of points in the denoised point cloud—left coronary artery root node, right coronary artery root node, and non-root nodes—includes: The physical space where the noise-reduced point cloud is located is divided into voxel grids. The centroid of a single type of point in any single voxel grid is calculated. The point closest to the centroid is selected and retained. The remaining points in the voxel grid are deleted to obtain a set of downsampling points of the corresponding type. The single type is any one of the three types: left coronary root node, right coronary root node and non-root node. The downsampling point cloud is obtained by combining the sets of downsampling points of the three types: left coronary root node, right coronary root node, and non-root node.

6. The method for extracting coronary artery center point clouds according to claim 1, characterized in that, The process of determining a unique left coronary root node and a unique right coronary root node in the downsampled point cloud includes: By combining any left coronary artery root node and any right coronary artery root node in the downsampled point cloud, multiple root node pairs are obtained. Candidate root node pairs are selected from the plurality of root node pairs based on spatial distance constraints, lateral position constraints, and front-back position constraints. The unique left coronary artery root node and the unique right coronary artery root node are determined by the sum of the network prediction confidence of the candidate root node pairs, or by clustering the downsampled point cloud to obtain multiple local clusters, and the unique left coronary artery root node and the unique right coronary artery root node are determined by the sum of the number of points in the local clusters to which the candidate root node pairs belong.

7. A training method for a single-stage multi-task network model, comprising: A training dataset is provided, wherein the samples are multi-view two-dimensional slice samples of coronary artery three-dimensional CCTA volume data samples, and the corresponding labels are the bounding box, type label and geometric center of each independent vascular region in the two-dimensional slice sample. The type labels include left coronary artery root node, right coronary artery root node and non-root node. Some samples in the training dataset contain multiple levels of artifacts. The single-stage multi-task network model is trained based on the training dataset. One task branch of the single-stage multi-task network model obtains the bounding box and type label of each independent blood vessel region in the two-dimensional slice sample, and the other task branch of the single-stage multi-task network model obtains the geometric center of each independent blood vessel region.

8. The training method for a single-stage multi-task network model according to claim 7, characterized in that, In the step of providing the training dataset, the image region containing the left coronary root node or the right coronary root node in the two-dimensional slice sample is copied and fused into other two-dimensional slice samples that do not contain the left coronary root node or the right coronary root node, to obtain a new two-dimensional slice sample.

9. A computing device, characterized in that, The device includes a memory and a processor, the memory storing a program, and the processor running the program to perform the coronary artery center point cloud extraction method according to any one of claims 1 to 6, or to perform the training method of the single-stage multi-task network model according to claim 7 or 8.

10. A program product, characterized in that, The program product executes the coronary artery center point cloud extraction method according to any one of claims 1 to 6 when running on a processor.

Images