An image processing method, image processing device, and computing equipment based on graph neural networks.

Abstract

This application provides an image processing method, image processing apparatus, and computing device based on a graph neural network. The image processing method includes obtaining the centerline of a blood vessel from a three-dimensional medical image containing blood vessels; setting multiple nodes on each centerline and obtaining cross-sections corresponding to each node; obtaining rays pointing towards the vessel wall along multiple directions centered on each node within each cross-section; sampling multiple points along each ray, constructing a feature sequence from all sampling point data on each ray, and using each feature sequence as vertices to construct a graph structure, where each vertex includes intensity variation features corresponding to the ray direction; based on the graph structure and a trained graph neural network, predicting the target length from each node to the vessel wall in the corresponding ray direction, and reconstructing the blood vessel according to the target length in each ray direction. This significantly improves the prediction accuracy of the reconstructed blood vessel's luminal diameter, making the reconstructed blood vessel closer to the real anatomical morphology of blood vessels.

Description

Technical Field

[0001] This application relates to the field of image processing technology, specifically to an image processing method, image processing device, and computing equipment based on graph neural networks. Background Technology

[0002] Clinical diagnosis and treatment decisions for coronary atherosclerotic heart disease require physicians to accurately quantify the degree of coronary artery stenosis. Current coronary CTA imaging techniques for lumen quantification generally employ a two-step approach: segmentation followed by quantification.

[0003] These techniques first use a fully convolutional network to segment CTA images at the voxel level, then extract the centerline based on the segmentation results and fit indices such as lumen radius and area. Due to its inherent physical and imaging characteristics, coronary CTA has several inherent limitations in estimating lumen boundaries and radii, such as spatial resolution and partial volume effects leading to blurred edges, and the true lumen radius appearing as a systematic underestimation or overestimation in the image.

[0004] These methods directly rely on voxel segmentation results, and the boundary offsets generated during segmentation are continuously accumulated and amplified during subsequent quantization, especially near the narrowest segments and bifurcation points, where significant radius errors are more likely to occur. These methods are prone to segmentation topological anomalies in vessel bifurcation and tortuous regions, leading to unstable centerline extraction, abrupt changes in the luminal diameter curve, spikes, or jagged edges. Furthermore, these methods are highly dependent on thresholding and morphological post-processing. In addition, these methods perform inference on full 3D volumetric data, resulting in high computational overhead and memory usage, which is not conducive to large-scale clinical deployment.

[0005] In addition, some methods use a single radius scalar to characterize the lumen on orthogonal sections based on the centerline. However, this type of characterization cannot accurately describe complex lumen morphologies such as eccentric plaques, irregular narrowings, and local deformations.

[0006] In summary, existing methods rely excessively on voxel segmentation, are prone to error accumulation and topological instability, and the characterization of lumen morphology is mostly limited to the centerline and a single radius. Under complex anatomical and imaging conditions, they cannot stably output continuous, smooth, and highly consistent results of lumen diameter and stenosis quantification. Summary of the Invention

[0007] This application addresses the aforementioned technical problems in the existing technology. The purpose of this application is to provide an image processing method, image processing device, and computing equipment based on graph neural networks, which can reduce the strong dependence on fine voxel segmentation results, reduce the accumulation and amplification of segmentation errors during vascular reconstruction, improve the prediction accuracy of the lumen diameter, obtain continuous and stable lumen boundaries, make the reconstructed blood vessels closer to the real vascular anatomy, and improve the accuracy and consistency of vascular reconstruction.

[0008] According to a first aspect of this application, an image processing method based on a graph neural network is provided, the image processing method comprising: via a processor, The centerline of the blood vessel is obtained based on a three-dimensional medical image containing the blood vessel; Multiple nodes are set on each of the aforementioned center lines, and the cross-sections corresponding to each node are obtained; Within each of the aforementioned cross sections, rays pointing towards the pipe wall are obtained along multiple directions centered on the node; Samples are taken from multiple sampling points along each ray, and all sampling point data on each ray are used to construct a feature sequence. Each feature sequence is used as a vertex to construct a graph structure, wherein the vertex includes the intensity change feature of the corresponding ray direction. Based on the graph structure and the trained graph neural network, the target length of each node to the vessel wall in the corresponding ray direction is predicted, and the blood vessel is reconstructed according to the target length in each ray direction.

[0009] According to a second aspect of this application, an image processing apparatus based on a graph neural network is provided, the image processing apparatus including a processor, the processor being configured to... The centerline of the blood vessel is obtained based on a three-dimensional medical image containing the blood vessel; Multiple nodes are set on each of the aforementioned center lines, and the cross-sections corresponding to each node are obtained; Within each of the aforementioned cross sections, rays pointing towards the pipe wall are obtained along multiple directions centered on the node; Samples are taken from multiple sampling points along each ray, and all sampling point data on each ray are used to construct a feature sequence. Each feature sequence is used as a vertex to construct a graph structure, wherein the vertex includes the intensity change feature of the corresponding ray direction. Based on the graph structure and the trained graph neural network, the target length of each node to the vessel wall in the corresponding ray direction is predicted, and the blood vessel is reconstructed according to the target length in each ray direction.

[0010] According to a third aspect of this application, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the steps of the image processing method based on graph neural networks described in various embodiments of this application.

[0011] According to a fourth aspect of this application, a computer program product is provided, comprising a computer program, wherein when the computer program is executed by a processor, it implements the steps of the image processing method based on graph neural networks described in various embodiments of this application.

[0012] Compared with the prior art, the beneficial effects of the embodiments of this application are as follows: This application provides an image processing method based on graph neural networks. The method obtains the centerline of a blood vessel from a three-dimensional medical image containing blood vessels. Multiple nodes are set on each centerline, and cross-sections corresponding to each node are obtained. Rays pointing towards the vessel wall are obtained along multiple directions centered on each node within each cross-section. Multiple sampling points along each ray are sampled, and all sampling point data on each ray are constructed into a feature sequence. Each feature sequence is used as a vertex to construct a graph structure, with each vertex containing intensity variation features corresponding to the ray direction. The graph structure unifies the spatial features of the blood vessel, fully utilizing the correlation information within and between cross-sections to improve the completeness and robustness of feature representation.

[0013] Based on a graph structure and a trained graph neural network, the target length from each node to the vessel wall in the corresponding ray direction is predicted, and the blood vessel is reconstructed based on the target length in each ray direction. Thus, constructing the vessel lumen boundary based on the predicted target length in each ray direction improves the reliability of lumen boundary prediction and reduces reliance on full voxel segmentation. The method provided in this application fully utilizes the spatial positional correlation within and between cross-sections of the blood vessel, improving the rationality of target length prediction. Reconstructing the blood vessel based on the target length significantly improves the accuracy of lumen diameter reconstruction, making the reconstructed blood vessel more closely resemble the actual anatomical morphology of the blood vessel, and obtaining continuous, stable, and more consistent lumen boundary and diameter measurement results. Attached Figure Description

[0014] In drawings that are not necessarily drawn to scale, the same reference numerals may describe similar parts in different views. Similar reference numerals with different letter suffixes may indicate different examples of similar components. The drawings generally illustrate various embodiments by way of example rather than limitation, and are used together with the specification and claims to illustrate the disclosed embodiments. Such embodiments are illustrative and exemplary, and are not intended to be exhaustive or exclusive embodiments of the method, apparatus, or non-transitory computer-readable medium having instructions for implementing the method.

[0015] Figure 1 A flowchart illustrating an image processing method based on a graph neural network according to an embodiment of this application is shown.

[0016] Figure 2 A schematic diagram of vascular reconstruction according to an embodiment of this application is shown.

[0017] Figure 3 This diagram illustrates a graph neural network-based prediction of target length according to an embodiment of this application.

[0018] Figure 4 A schematic cross-sectional view of a blood vessel reconstructed based on a target length according to an embodiment of this application is shown.

[0019] Figure 5 A schematic diagram of an image processing apparatus based on a graph neural network according to an embodiment of this application is shown. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of this application, the application will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments of this application will be further described in detail below with reference to the accompanying drawings and specific examples, but these are not intended to limit the scope of this application.

[0021] The terms "first," "second," and similar words used in this application do not indicate any order, quantity, or importance, but are merely used for distinction. The terms "including" or "comprising," etc., used in this application mean that the element preceding the word encompasses the elements listed after the word, and do not exclude the possibility of encompassing other elements. In this application, the arrows shown in the figures for each step are merely examples of the execution order, not limitations. The technical solution of this application is not limited to the execution order described in the embodiments. The steps in the execution order can be combined, broken down, or rearranged, as long as the logical relationship of the executed content is not affected.

[0022] All terms used in this application (including technical or scientific terms) have the same meaning as understood by one of ordinary skill in the art to which this application pertains, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant art, and not as idealized or highly formalized, unless expressly defined herein. Techniques and apparatus known to one of ordinary skill in the art may not be discussed in detail, but where appropriate, such techniques, methods, and apparatus should be considered part of the specification.

[0023] Figure 1 The flowchart illustrates an image processing method based on a graph neural network according to an embodiment of this application. Specifically, the steps S101-S105 are executed by a processor. In this application, the arrows shown in the figure for each step are merely examples of the execution order and not limitations. The technical solution of this application is not limited to the execution order described in the embodiment. The steps in the execution order can be combined, decomposed, or their order can be changed, as long as the logical relationship of the execution content is not affected.

[0024] In step S101, the centerline of the blood vessel is obtained based on a three-dimensional medical image containing the blood vessel.

[0025] Specifically, acquiring three-dimensional medical images containing blood vessels, the technical term "acquisition" means any method of obtaining them directly or indirectly, with or without additional image processing such as noise reduction, cropping, and reconstruction.

[0026] The three-dimensional medical images include CTA (Computed Tomographic Angiography) images, and the vessels include coronary arteries, aorta, or carotid arteries. For example, for CTA images containing coronary arteries, at least the major vessels such as the left main coronary artery (LM), left anterior descending artery (LAD), circumflex artery (LCX), and right coronary artery (RCA) are covered to ensure the integrity of the coronary vessels in the CTA images.

[0027] In some embodiments, the acquired three-dimensional medical images are preprocessed. Exemplarily, the preprocessing includes selecting cardiac cycle phases with smaller motion artifacts from multiphase ECG-gated data, such as selecting late diastolic phases; resampling the original voxels to an approximately isotropic resolution, such as setting the resolution to 0.4 mm to 0.6 mm; truncating and normalizing the CT values, such as mapping the CT value range from -100 HU to 1000 HU to a normalized value of 0 to 1; and processing the image using a little-kernel Gaussian filter or anisotropic filter to reduce high-frequency noise in the image.

[0028] There are several ways to obtain the centerline, such as using the minimum path method, the skeleton extraction method based on the coarse segmentation results, or the centerline tracking method. There are no restrictions on which method can be used.

[0029] For example, the minimum path method typically finds the shortest path through the interior of the blood vessel between the starting and ending points in the CTA image to extract the centerline. Skeleton extraction methods based on coarse segmentation results usually require first segmenting the blood vessel target and then using a geometric centroid model to calculate the centerline. Centerline-based tracking methods obtain the blood vessel orientation at the current location by determining the local orientation of the blood vessel. Then, combining the local spatial information and curvature changes of the blood vessel in the image, a cross-section of the blood vessel orthogonal to the current orientation and its center point are obtained. The current tracking point and direction are then corrected based on the center point of this cross-section, and the above local center point calculation process is repeated until the end of the blood vessel is tracked.

[0030] This is merely an illustrative example, and the method for obtaining the centerline of the blood vessel is not the core of the embodiments of this application, and is not specifically limited thereto.

[0031] In step S102, multiple nodes are set on each of the center lines, and the cross-sections corresponding to each node are obtained.

[0032] In other words, a series of ordered sampling points are selected as nodes along each centerline.

[0033] Specifically, a curvature-adaptive sampling method can be used to set up multiple nodes. For example, the curvature at each position on the centerline can be calculated first, and the node spacing can be dynamically adjusted according to the curvature. A larger sampling step size can be used in areas where the blood vessel is straight and the curvature is small, while a smaller sampling step size can be used in areas where the blood vessel is highly tortuous and the curvature is large. The node distribution can be further increased in the bifurcation region of the blood vessel, ultimately generating a node sequence with adaptive density.

[0034] Alternatively, a method of uniform sampling with equal arc length can be used to set the nodes. For example... Figure 2 As shown, centerline 202 is obtained by extracting the centerline from CTA image 201. Then, each centerline can be represented as a sequence of three-dimensional points arranged in arc length order. Through interpolation or resampling, the arc length between adjacent nodes on each centerline is made equal to a preset step size, such as 0.5 mm or 1.0 mm. The i-th node on each centerline can be represented as C... i And set N nodes for each centerline. Furthermore, it can identify bifurcation points based on each node and record the parent-child relationships of each vessel segment. For each node C... i It records the previous node, the next node, and the corresponding blood vessel segment identifier to construct the blood vessel tree topology.

[0035] This is merely an illustrative example, and other methods to set multiple nodes on each center line are not excluded. This embodiment does not constitute a limitation on any specific solution.

[0036] In some embodiments, a cross-section perpendicular to the tangential direction of the centerline is constructed at each node location. Specifically, the tangential vector of the centerline at the current node can be calculated first, and the initial normal vector and subnormal vector can be determined based on this tangential vector. Then, based on the cross-section coordinate system already constructed at the previous node, parallel movement and orientation alignment are performed to keep the cross-section coordinate system of the current node consistent with the cross-section coordinate system of the previous node, generating a cross-section that is perpendicular to the tangential direction of the centerline and has a continuous and unified orientation. This construction method can ensure that the cross-section orientation of each node remains consistent, avoiding abnormal rotation of the blood vessel cross-section during its extension along the centerline.

[0037] In some embodiments of this application, the cross-section is constructed based on a plane orthogonal to the tangential direction of each node.

[0038] As an example, each node C can be... i Calculate the tangential vector t i The tangential vector ti It can characterize the local extension direction of blood vessels. Node C can be calculated. i+1 With C i-1 The coordinate difference vectors are summed and normalized to obtain node C. i Tangential vector t at point i In the tangential vector t i Based on this, a local orthogonal coordinate system can be constructed. For example, a global reference vector (such as the patient's head-to-foot direction) can be selected first, and this global reference vector can be compared with the tangential vector t. i The cross product is used to obtain the normal direction. Then, this normal direction is compared with the tangent vector t. i After orthogonalization and normalization, the standard normal vector n is obtained. i Through the tangential vector t i With normal vector n i The cross product yields the binormal vector b. i This forms a set of three pairwise orthogonal unit vectors. Taking node C as an example... i Let n be the origin of the coordinate system, and the normal vector n be the origin of the coordinate system. i With the binormal vector b i Commonly spanned corresponding node C i A local cross-section. This cross-section is perpendicular to the centerline at node C. i Tangential vector t at point i Vertical, that is, the node C i The orthogonal cross-section of the blood vessel at the location, such as Figure 2 203 in the figure represents the local centerline containing the cross-section, and the area shown by the ellipse in the figure represents the schematic diagram of the cross-section.

[0039] In step S103, within each of the cross sections, rays pointing towards the tube wall are obtained along multiple directions centered on the node.

[0040] In other words, within the orthogonal cross section corresponding to each node, rays in multiple directions are generated with the node as the origin, and the rays extend towards the wall of the blood vessel.

[0041] Specifically, the CT values ​​of all pixels within each orthogonal cross section can be obtained. The transverse and longitudinal gradient components of the CT values ​​within the two-dimensional space of the cross section are then calculated. Combining these two sets of gradient components, the gradient magnitude and gradient direction at each location are calculated, thus completely extracting the CT value gradient distribution information of that cross section. Since there are significant differences in the CT values ​​of the lumen, wall, and surrounding soft tissue of a blood vessel, a significant abrupt change in grayscale gradient occurs at the interface between the lumen and wall. Therefore, the nodes of each cross section can be used as the starting points for ray generation, generating rays pointing towards the vessel wall along the high gradient direction.

[0042] In other embodiments of this application, based on the differences in the contour shape of the pipe wall in each cross section, the angular intervals between rays in different directions are set differently so that the density of the rays matches the changes in the pipe wall contour.

[0043] Specifically, a preliminary detection and analysis of the vascular contour can be performed on the current cross-section. Based on the CT value distribution characteristics, the approximate distribution area of ​​the vessel wall in the circumferential direction of the cross-section can be roughly delineated. The density of the X-ray array can be adaptively adjusted based on the degree of morphological changes in the vessel wall contour. For example, in areas with narrow lumens, large undulations in the vessel wall morphology, and irregular contours, the angular interval between the X-rays in that area can be reduced, thus increasing the X-ray density in that area. Conversely, in areas with smooth vessel wall morphology and gentle contour changes, the angular interval between the X-rays can be increased, resulting in a sparser X-ray density in that area. In this way, the sampling accuracy of key vascular lesion areas can be specifically enhanced, while reducing the redundant computational load in normal, regular vessel areas, effectively balancing the overall sampling effect and computational efficiency.

[0044] In other embodiments of this application, within the cross-section, multiple spoke-shaped rays are arranged at equal angular intervals around the node.

[0045] Using the node at the center of the cross-section as the common starting point of all rays, and with this node as the center, a fixed angular interval is set in the plane of the cross-section. Multiple rays are then arranged sequentially along the circumference according to this angular interval, so that all rays extend radially from the node toward the outside of the cross-section. The multiple rays are distributed in a regular spoke-shaped ring, and the rays completely cover the entire circumferential direction of the cross-section. The included angle between each adjacent ray is consistent. In this way, uniform sampling of the circumferential area of ​​the pipe wall can be achieved.

[0046] like Figure 2 Figure 204 shows the sampled feature map of the cross-section at each node along the centerline. Within the cross-sectional plane, node C... i Centered on a central axis, K spoke-shaped rays are arranged at equal angular intervals. K is a positive integer. In some embodiments, there are no fewer than 12 rays in each cross-section to ensure the accuracy of subsequent vessel wall contour localization and lumen morphology reconstruction. As a preferred embodiment, the number of rays K is 24, which ensures the smoothness of the vessel contour while avoiding difficulties in model training due to an excessively large K. For example, the angle corresponding to the k-th ray can be set as angle theta. k = 2π × k / K, from the angle and n i b i The linear combination yields the direction vector d in the k-th direction. i,k = cos(theta k )×n i +sin(thetak )×b i .

[0047] In step S103, multiple sampling points along each ray are sampled, and all sampling point data on each ray are used to construct a feature sequence. Each feature sequence is used as a vertex to construct a graph structure, wherein the vertex includes the intensity change feature of the corresponding ray direction.

[0048] In other words, pixel CT values, grayscale gradients, and other relevant data are sequentially collected at multiple equidistant locations along each spoke-shaped ray to form a one-dimensional feature sequence corresponding to that ray. The feature sequence of each ray is used as a vertex in the graph structure, and the vertex's attributes contain the numerical change information of all sampling points on that ray, that is, the intensity change characteristics along the ray direction from the central node to the pipe wall boundary.

[0049] Specifically, taking node C i Starting from the direction vector d of the single ray i,k Multiple sampling points can be selected according to a preset step size (such as 1 pixel), and the maximum sampling radius R can be preset. max Each ray is divided into M sampling layers (e.g., Figure 2 (204 in the text), the sampling radii corresponding to each sampling layer are r1, r2, r3..., r... M The radius difference between adjacent layers is a preset step size r, and the coordinates of each sampling point are determined by node C. i Coordinates, sampling radius, and the direction vector d of the current ray i,k get.

[0050] The CT values ​​of each sampling point can be obtained sequentially using trilinear interpolation, nearest neighbor interpolation, or other methods. These values ​​are then concatenated into a one-dimensional vector according to the sampling order, representing the feature sequence of each ray. This feature sequence describes the intensity variation characteristics from the node to the vessel wall. Specifically, the intensity variation characteristics characterize the trend of CT value changes along the ray direction from the center of the blood vessel to the vessel wall, reflecting the boundary information between the lumen and the vessel wall.

[0051] In some embodiments, the intensity variation characteristics can be obtained by calculating radial voxel features on the one-dimensional intensity sequence corresponding to each ray. For example, the intensity variation characteristics include the average intensity, variance, maximum value, and / or minimum value obtained statistically in different radial radius intervals; the gradient statistics formed by the intensity differences between adjacent sampling points, used to characterize the edge change information at the junction of the lumen and the pipe wall; and statistical descriptive features obtained based on piecewise histograms or multi-scale methods. Based on radial voxel features, the intensity distribution law and local abrupt change characteristics along the ray direction from the node to the pipe wall can be comprehensively described.

[0052] The feature sequence corresponding to each ray is constructed as a vertex in the graph structure. The attribute of the vertex is the intensity change feature along the ray direction, specifically the trend of the CT value of the sampling point with the radial distance (such as the value is stable inside the lumen and the CT value changes abruptly at the boundary of the tube wall). Each vertex uniquely corresponds to a ray, and the number of vertices is consistent with the number of rays in the cross section.

[0053] In some embodiments of this application, the vertex further includes global geometric features of the blood vessels at the corresponding node location.

[0054] Specifically, this global geometric feature is used to characterize the overall structure of the blood vessel's morphology at the current node's location. For example, the global geometric feature can include spatial location global features, morphological structure global features, and topological relationship global features. The spatial location global feature includes the normalized arc length position of the current node within the entire blood vessel segment. This feature clarifies the relative coordinates of the current cross-section on the overall extension path of the blood vessel, reflecting whether the position is near the beginning, end, or middle of the vessel. The morphological structure global features include features describing the macroscopic morphology of the blood vessel, such as local curvature, local bending angle, and tubularity. Local curvature reflects the degree of curvature of the blood vessel at this node, and tubularity characterizes the regularity of the blood vessel cross-section. The topological relationship global features include labeling information indicating whether it is located in a bifurcation neighborhood. This topological relationship global feature encodes the topological structure of the blood vessel (e.g., whether it is a bifurcation point, branch vessel, etc.), clarifying the topological position of the current cross-section and avoiding boundary recognition errors in complex structural regions such as bifurcation.

[0055] In a preferred embodiment, the vertex includes intensity variation features corresponding to the ray direction, and global geometric features of the blood vessel at the corresponding node location.

[0056] In other words, the radial intensity variation features are concatenated with the global geometric features at the vector level to form the node C. i The eigenvector f in the direction of the k-th ray i,k .

[0057] Specifically, for example, one can first sample and calculate radial voxel features along each ray to obtain intensity variation features (such as local features like mean, variance, gradient difference, and piecewise histogram for different radius intervals). At the same time, based on the vessel centerline, one can calculate the global geometric features of each node (such as normalized arc length position, local curvature, bending angle, tubularity, and whether it is in a bifurcation neighborhood). Then, through vector concatenation, weighted fusion, element-wise addition, or convolutional fusion, the two types of features are integrated into the feature vector of the vertex. Finally, the fused feature vector is used as the feature of the corresponding vertex in the graph structure, and the edges between vertices are constructed based on the circumferential adjacency of the rays, forming a graph structure for the input of the graph neural network.

[0058] In this way, each vertex in the graph structure can not only capture the local intensity changes along each ray direction, but also associate the global spatial location, macroscopic shape and topological relationship of the blood vessel. This enables the graph neural network to improve the accuracy of blood vessel contour recognition during the subsequent learning process.

[0059] In some embodiments of this application, edges between vertices are constructed based on the adjacent orientation relationship of each ray in the circumferential direction of the cross section, forming a complete graph structure.

[0060] Specifically, based on the spatial adjacency of each ray in the circumferential direction of the cross-section, edge connections are constructed for the vertices corresponding to adjacent rays in the circumferential direction. For example, the vertex corresponding to the k-th ray establishes edges with the vertices corresponding to the (k-1)-th and (k+1)-th rays, respectively, thus forming a structured graph structure. This graph structure not only preserves the radial intensity variation information from the node to the vessel wall on a single ray, but also encodes the spatial positional association of adjacent rays in the circumferential direction through edge connections. Transforming the raw image information of the two-dimensional cross-section of the blood vessel into structured graph data enables the graph neural network to efficiently learn and capture the local density abrupt changes in the vessel wall interface region, as well as the global continuity features of the vessel contour in the circumferential direction, thereby significantly improving the accuracy and robustness of vessel wall boundary localization, lumen stenosis detection, and three-dimensional morphological reconstruction of the blood vessel.

[0061] In other embodiments of this application, vertices of adjacent cross sections corresponding to the same ray angle are connected to each other, vertices of adjacent ray angles within the same cross section are connected to each other, and vertices of the same ray angle on cross sections with parent-child relationships at bifurcation points are connected to each other to form the edges of the graph structure.

[0062] Where the ray angle k remains constant, the adjacent nodes C along the arc length of the blood vessel are... i With C i+1 The vertex v of the corresponding cross section i,k With v i+1,k Connecting them is used to model the spatial continuity of blood vessels along the direction of their extension along the centerline, enabling graph neural networks to learn the morphological features of blood vessels in the longitudinal direction.

[0063] At the same node C i Within the same cross section, the vertex v corresponding to the angles k and k+1 of adjacent rays... i,k With v i,k+1 Connected and arranged in a closed loop, these structures reflect the circumferential relationship of the pipe wall profile within a single cross-section.

[0064] In the bifurcation region of blood vessels, vertices with the same ray angle at corresponding nodes on the parent and child blood vessels are connected, thereby establishing the structural dependency relationship between the parent and child blood vessels in the bifurcation region in the graph neural network and improving the accuracy of vessel wall identification at the bifurcation location.

[0065] Specifically, adjacent cross-sections (corresponding to node C) along the arc length extension direction of the blood vessel centerline and at the same ray angle k are... i With C i+1 Vertex v on ) i,k With v i+1 , k They are interconnected, forming adjacent edges along the arc length; simultaneously, at the same node C i Within a single cross-section, the vertices corresponding to the angle k of adjacent rays arranged in clockwise or counterclockwise order (e.g., v) i,k With v i,k+1 The rays are interconnected, and because they are distributed in a ring shape within the cross-section, a closed ring connection is ultimately achieved. Furthermore, at the bifurcation points of the blood vessels, for parent and child blood vessels with a parent-child topological relationship, their corresponding nodes C are... i Vertices with the same ray angle are connected to each other (e.g., the parent vessel vertex v). 父,i,k With the apex of the sub-vessel v 子，j,k This is used to model the structural relationship between parent and child blood vessels in the bifurcation region, solve the problem of easy distortion of the contour at the bifurcation point, and ensure the natural transition of blood vessel morphology in the bifurcation region.

[0066] Returning to the embodiment of this application, in step S105, based on the graph structure and the trained graph neural network, the target length of each node to the vessel wall in the corresponding ray direction is predicted, and the blood vessel is reconstructed according to the target length in each ray direction.

[0067] In other words, based on the graph structure and the trained graph neural network, using the feature vector corresponding to each vertex as input, the features of adjacent vertices are aggregated and updated through multiple graph convolutional layers. Finally, the regression output layer predicts the distance from each node to the vessel wall along the corresponding ray direction, thus obtaining the predicted value of the target length. Centered on each node, the contour points of the vessel wall are determined according to each ray direction and its corresponding target length. The contour points within the same cross-section are sequentially connected to form the complete cross-sectional contour of the blood vessel. All cross-sectional contours are then sequentially stitched together along the arc length of the centerline, ultimately achieving the three-dimensional morphological reconstruction of the entire blood vessel.

[0068] Specifically, such as Figure 3 As shown, in this embodiment of the application, node C i The vertex v of the graph is defined by the ray angle k. i,k The graph edges are constructed according to three types of adjacency relationships: arc-length adjacency edges (connecting vertices of adjacent nodes under the same ray angle k, corresponding to purple connections), cross-sectional angle adjacency edges (connecting vertices of adjacent angles under the same node, corresponding to green connections), and bifurcation adjacency edges (connecting vertices of the same ray angle on the parent and child blood vessels in the bifurcation region), which together constitute the graph structure input.

[0069] Vertex v i,k eigenvector f i,k The input is fed into a graph neural network structure consisting of multiple graph convolutional layers (or graph attention layers) and non-linear activation layers. Each graph convolutional layer, based on the adjacency relationships of three types of edges, starts from the adjacent vertex v. i,k Aggregate feature information; if ordinary graph convolution is used, it directly applies the features to adjacent vertices v. i,k After summing the features, add the summation to the linear transformation of its own features. If a graph attention mechanism is used, first calculate the adjacent vertex v. i,k The attention weights between nodes are calculated, and the adjacent features are summed using a weighted average. Then, the current vertex v is updated using an activation function. i,k The network, through multi-layer feature propagation and fusion, comprehensively considers the angles of each ray within the same cross-section and the information of upstream and downstream blood vessel segments in adjacent cross-sections, ensuring that each vertex v... i,k The features simultaneously include local radial details and global vascular morphological constraints.

[0070] Graph neural networks output each vertex v through a regression layer. i,k Output predicted target length L i,k In this embodiment, monotonically increasing nonlinear functions such as ReLU or Softplus can be used in the output layer to ensure the predicted target length L. i,k Non-negative. In the neural network model training of this application, a multi-constraint loss function is used, including regression error, smoothing constraint, bifurcation constraint, and physical prior. The regression error aligns with the true target length, the smoothing constraint ensures contour continuity, the bifurcation constraint stabilizes the bifurcation region morphology, and the physical prior constraint results conform to vascular physiological laws, collectively optimizing the target length L. i,k The accuracy of the prediction.

[0071] This application retrieves all vertices v i,k Target length L i,k , with each node C i Centered on the cross-section, the angular directions of each ray are related to the corresponding target length L. i,k Determine the coordinates P of the contour point along the ray direction. i,k , the same node C i Connect all corresponding contour points to form a closed loop, such as... Figure 2 The outline of the blood vessel lumen shown in 205 is as follows. Figure 4 An enlarged view of the lumen profile is shown, which shows that the target length in each ray direction of the lumen profile is close to the actual blood vessel radius.

[0072] In this embodiment, the lumen contours of all cross-sections are sequentially arranged and surface-fitted along the arc length extension direction of the vessel centerline. Simultaneously, topological constraints of adjacent bifurcation edges ensure the transition between the parent and child vessel contours in the bifurcation region, thereby reconstructing the three-dimensional structure of the entire vessel (e.g., ...). Figure 2 (206 in the middle).

[0073] In other words, based on the nodes of each cross section and the target length of each node in the corresponding ray direction, the contour points of each node in the corresponding ray direction are obtained; the contour points of adjacent ray angles in the same cross section are connected, and the contour points of the same ray angle in adjacent cross sections are connected to obtain the reconstructed blood vessel.

[0074] Based on the lumen profile, geometric quantities such as cross-sectional area and equivalent diameter can be calculated, thereby obtaining node C. i area sequence A i In the area sequence A i The minimum value is found within the vascular structure to obtain the minimum lumen area and its corresponding node location. Related indicators such as diameter stenosis, area stenosis, and lesion length are then calculated. The predicted results are overlaid on a CPR straightening image or a 3D vascular tree, providing physicians with integrated morphological and quantitative visualization results.

[0075] The three-dimensional blood vessels reconstructed based on the embodiments of this application can realistically reflect the fine geometric relationships of blood vessels in curved, torsional, and bifurcated areas, avoiding problems such as spatial structural distortion and difficulty in corresponding lesion locations that occur when observing traditional straightened images. This three-dimensional blood vessel is directly reconstructed from the predicted lumen contour, ensuring a high degree of consistency between quantitative calculation results (such as stenosis and minimum lumen area) and the visualized structure, reducing the inconsistency between measured values ​​and displayed structures in traditional methods. By overlaying the lumen contour in the three-dimensional blood vessel tree and highlighting the lesion location with the minimum lumen area or minimum diameter, doctors can intuitively locate the spatial position of the lesion segment within the overall coronary anatomy, significantly improving the accuracy and efficiency of stenosis localization. Simultaneously, this three-dimensional blood vessel can serve as the basis for subsequent geometric analysis and functional expansion, supporting applications such as local cross-sectional change rate calculation, lumen volume statistics, and hemodynamic simulation. Doctors can also directly observe the spatial correspondence between the predicted vessel wall boundary and the original image in the three-dimensional blood vessel, facilitating structural-level verification and validation of the automatic prediction results, further enhancing the interpretability and clinical credibility of the system output.

[0076] In some embodiments of this application, training the graph neural network to obtain a trained graph neural network includes: determining the reference length of each node in the corresponding ray direction based on the intersection position of each ray with the actual pipe wall surface; constructing a first loss function based on the first difference between the target length and the reference length; optimizing the parameters of the graph neural network through backpropagation until the first loss function converges to obtain a trained graph neural network.

[0077] Specifically, a graph neural network is trained based on a labeled dataset of blood vessel lumens. For cases in the dataset, doctors can delineate the contours of the lumens to generate realistic 3D reference blood vessels. Then, the intersection points of each ray with the real vessel wall surface of the reference 3D blood vessel are calculated along each centerline and ray direction. Based on the distance between the intersection point and the corresponding node, the reference length of each node in the corresponding ray direction is determined, and this reference length serves as a supervision signal for training the graph neural network model.

[0078] The target length output by the graph neural network is compared with the reference length, and the first difference between them is calculated. A first loss function is constructed based on this first difference, which can be a regression loss form such as mean absolute error or mean squared error. When the first difference is greater than a preset threshold, it is converted into the corresponding loss value. The gradient of the first loss function is passed to the parameters of each layer of the graph neural network through backpropagation, iteratively updating the parameters of the graph neural network model. This continuous cycle of feature forward propagation, loss calculation, and parameter update continues until the first loss function converges to a preset range, ultimately resulting in a trained graph neural network.

[0079] In some embodiments of this application, the method further includes: constructing a smooth loss along the centerline arc length based on a second difference in the target lengths of adjacent cross sections in the same ray angle direction; weighting and fusing the smooth loss with the first difference to obtain a second loss function; and performing backpropagation optimization on the graph neural network based on the second loss function.

[0080] Specifically, the target length of adjacent cross sections along the arc length of the centerline can be extracted for the same ray angle direction. The second difference between the two can be calculated. After squaring the second difference, the average value is calculated over all ray angles and adjacent node pairs to obtain the smoothing loss along the arc length of the centerline. This smoothing loss is used to constrain the smoothness of the gradual change in the diameter of the blood vessel along the longitudinal direction, so as to suppress the violent fluctuations in places without real stenosis and avoid abrupt changes in the diameter between adjacent cross sections.

[0081] The smoothing loss described above is weighted and fused with the first loss function to obtain the second loss function. The second loss function is used to balance the priority of regression accuracy and smoothness constraints. The gradient is calculated based on the second loss function, and the gradient is passed to the parameters of each layer of the graph neural network through the backpropagation algorithm. The parameters of the graph neural network model are iteratively updated until the second loss function converges, resulting in a well-trained graph neural network that balances the accuracy of lumen length prediction and the smoothness of longitudinal morphology.

[0082] In some embodiments of this application, a bifurcation consistency loss is constructed based on the third difference in target lengths of cross sections with a parent-child relationship at the bifurcation location on the blood vessel in the same ray angle direction; the bifurcation consistency loss is weighted and fused with a first loss function and / or a second loss function to obtain a third loss function, and the graph neural network is optimized by backpropagation based on the third loss function.

[0083] Specifically, the bifurcation points on the blood vessels are first located. For the cross-sections of the parent and child blood vessels at these bifurcation points, which have a parent-child topological relationship, the target lengths corresponding to each other in the same ray angle direction are extracted, and a third difference between these two target lengths is calculated. To reduce this third difference, a bifurcation consistency loss is constructed to constrain the continuous transition of the diameter between the parent and child blood vessels in the corresponding direction, avoiding abrupt diameter changes, contour breaks, or unnatural structures in the bifurcation region. This bifurcation consistency loss is then weighted and fused with either a first or second loss function to obtain a third loss function. The gradient is calculated based on this third loss function, and the parameters of the graph neural network are iteratively optimized using a backpropagation algorithm. This ensures that the model accurately predicts the target length while maintaining a smooth overall blood vessel contour and continuous bifurcation morphology that conforms to physiological structure, resulting in a more stable and reliable trained graph neural network.

[0084] In some embodiments of this application, training the graph neural network to obtain a trained graph neural network further includes: penalizing negative target lengths and target lengths exceeding a threshold to construct a physical prior constraint loss; weighting and fusing the physical prior constraint loss with a first loss function, a second loss function, and / or a third loss function to obtain a fourth loss function; and performing backpropagation optimization on the graph neural network based on the fourth loss function.

[0085] In other words, when the target length is negative or exceeds the threshold, such prediction results are penalized. For example, the penalty term can be calculated by squaring the value of the part that exceeds the threshold, thereby constructing a physical prior constraint loss. This physical prior constraint loss can effectively suppress meaningless prediction results, so that the target length output by the model always remains within a reasonable range that conforms to the vascular anatomy.

[0086] The physical prior constraint loss is weighted and fused with the first, second, and / or third loss functions according to preset weights to obtain a fourth loss function. The preset weights of each loss term can be adaptively adjusted through validation set experiments to achieve an optimal balance among multiple constraints. The gradient is calculated based on this fourth loss function, and backpropagation and parameter iterative updates are performed on the graph neural network using optimization algorithms such as stochastic gradient descent and Adam until the model converges on the validation set, resulting in a well-trained graph neural network that combines prediction accuracy, structural smoothness, bifurcation continuity, and physical plausibility.

[0087] Thus, by introducing smoothing loss along the centerline arc length and bifurcation consistency loss into the loss function, and combining it with physical prior constraint loss, we can achieve continuous, stable, and relatively low computational cost quantitative analysis of the vascular lumen diameter, and improve the quantitative accuracy and stability of structural indicators such as minimum lumen area, stenosis, and lesion extent.

[0088] Figure 5 A schematic diagram of an image processing apparatus based on a graph neural network according to an embodiment of this application is shown. Figure 5 As shown, the image processing apparatus 500 may include an interface 507 and a processor 501. The interface 507 may be configured to receive a three-dimensional medical image containing blood vessels. The processor 501 may be configured to execute an image processing method based on a graph neural network according to various embodiments of the present invention.

[0089] Through interface 507, the device for extracting the centerline of a three-dimensional medical image containing blood vessels can be connected to a network (not shown), such as, but not limited to, a local area network in a hospital or the Internet. However, the communication methods implemented by interface 507 are not limited to a network and may include NFC, Bluetooth, WIFI, etc.; it can be a wired connection or a wireless connection. Taking a network as an example, interface 507 can connect the device for extracting the centerline of a three-dimensional medical image to external devices such as an image acquisition device (not shown), a medical image database 508, and an image data storage device 509. The image acquisition device can be any type of imaging modality, such as, but not limited to, computed tomography (CT), digital subtraction angiography (DSA), magnetic resonance imaging (MRI), functional MRI, dynamic contrast-enhanced MRI, diffusion MRI, spiral CT, cone-beam computed tomography (CBCT), positron emission tomography (PET), single-photon emission computed tomography (SPECT), X-ray imaging, optical tomography, fluorescence imaging, ultrasound imaging, and radiotherapy field imaging.

[0090] In some embodiments, the image processing apparatus 500 may be a dedicated intelligent device or a general-purpose intelligent device. For example, the image processing apparatus 500 may be a computer customized for processing tasks such as extracting and extending the centerline, or a server located in the cloud. For example, the image processing apparatus 500 may be integrated into an image acquisition device.

[0091] The image processing apparatus 500 may include a processor 501 and a memory 504, and may additionally include at least one of an input / output 502 and an image display 503.

[0092] Processor 501 may be a processing device that includes one or more general-purpose processing devices, such as a microprocessor, a central processing unit (CPU), a graphics processing unit (GPU), etc. More specifically, processor 501 may be a Complex Instruction Set Computing (CISC) microprocessor, a Reduced Instruction Set Computing (RISC) microprocessor, a Very Long Instruction Word (VLIW) microprocessor, a processor running other instruction sets, or a processor running a combination of instruction sets. Processor 501 may also be one or more special-purpose processing devices, such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), system-on-a-chip (SoCs), etc. As those skilled in the art will understand, in some embodiments, processor 501 may be a special-purpose processor rather than a general-purpose processor. Processor 501 may include one or more known processing devices, such as microprocessors from the Pentium™, Core™, Xeon™, or Itanium series manufactured by Intel™, the Turion™, Athlon™, Sempron™, Opteron™, FX™, or Phenom™ series manufactured by AMD™, or various processors manufactured by SunMicrosystems. Processor 501 may also include a graphics processing unit, such as a GPU from GeForce®, the Quadro® or Tesla® series manufactured by Nvidia™, the GMA or Iris™ series manufactured by Intel™, or the Radeon™ series manufactured by AMD™. Processor 501 may also include an accelerated processing unit, such as the Desktop A-4 (6,6) series manufactured by AMD™, or the Xeon Phi™ series manufactured by Intel™. The disclosed embodiments are not limited to any type of processor or processor circuitry, which may otherwise be configured to perform the image processing methods according to various embodiments of the present invention. The method may include: obtaining the centerline of the blood vessel based on a three-dimensional medical image containing the blood vessel; setting multiple nodes on each centerline and obtaining a cross-section corresponding to each node; within each cross-section, obtaining rays pointing towards the vessel wall along multiple directions with the node as the center; sampling along multiple sampling points on each ray, constructing a feature sequence from all sampling point data on each ray, and constructing a graph structure using each feature sequence as a vertex, wherein the vertex includes intensity change features corresponding to the ray direction; predicting the target length from each node to the vessel wall in the corresponding ray direction based on the graph structure and a trained graph neural network, and reconstructing the blood vessel according to the target length in each ray direction.

[0093] Additionally, the term "processor" or "graphics processor" may include more than one processor, such as a multi-core design or multiple processors, each having a multi-core design. Processor 501 can execute a sequence of computer program instructions stored in memory 504 to perform the various operations, processes, and methods disclosed herein.

[0094] Processor 501 may be communicatively coupled to memory 504 and configured to execute computer-executable instructions stored therein. Memory 504 may include read-only memory (ROM), flash memory, random access memory (RAM), dynamic random access memory (DRAM) such as synchronous DRAM (SDRAM) or Rambus DRAM, static memory (e.g., flash memory, static random access memory), etc., on which the computer-executable instructions are stored in any format. In some embodiments, memory 504 may store computer-executable instructions for one or more image processing programs 505. The computer program instructions may be accessed by processor 501, read from ROM or any other suitable memory location, and loaded into RAM for execution by processor 501. For example, memory 504 may store one or more software applications. The software applications stored in memory 504 may include, for example, an operating system (not shown) for a general computer system and an operating system for a soft control device.

[0095] Furthermore, memory 504 may store the entire software application or only a portion of the software application that can be executed by processor 501 (e.g., image processing program 505). Additionally, memory 504 may store multiple software modules for implementing various steps of the image processing method consistent with the present invention.

[0096] Furthermore, memory 504 may store data generated / buffered during the execution of a computer program, such as medical image data 506, which may include medical images transmitted from an image acquisition device, a medical image database 508, an image data storage device 509, etc. In some embodiments, medical image data 506 may include a three-dimensional medical image containing blood vessels for which centerline extraction is to be performed, from which the image processing program 505 will extract the centerline.

[0097] In some embodiments, an image data storage device 509 may be provided to exchange image data with a medical image database 508, and a memory 504 may communicate with the medical image database 508 to obtain a three-dimensional medical image for centerline extraction. For example, the image data storage device 509 may reside in another medical image acquisition device (e.g., a CTA scan performed on the patient). The patient's three-dimensional medical image may be transmitted and saved to the medical image database 508, and the image processing device 500 may obtain a specific patient's three-dimensional medical image from the medical image database 508 and perform centerline extraction on that specific patient's three-dimensional medical image. Furthermore, trained graph neural network models may be stored on the medical image database 508 for access, acquisition, and utilization by other image processing devices when needed.

[0098] In some embodiments, the graph neural network model may be stored in memory 504. Alternatively, the learning network may be stored in a remote device, a separate database (such as a medical image database 508), a distributed device, and may be used by an image processing program 505.

[0099] In addition to displaying three-dimensional medical images, the image display 503 can also display other information, such as the centerline of blood vessels, reconstructed blood vessels, and related calculation results. For example, the image display 503 can be an LCD, CRT, or LED display.

[0100] Input / output 502 can be configured to allow image processing device 500 to receive and / or transmit data. Input / output 502 may include one or more digital and / or analog communication devices that allow the device to communicate with a user or other machines and devices. For example, input / output 502 may include a keyboard and mouse that allow the user to provide input.

[0101] In some embodiments, the image display 503 may present a user interface so that the user can conveniently and intuitively correct (such as edit, move, modify, etc.) related results using the input / output 502 together with the user interface.

[0102] Interface 507 may include network adapters, cable connectors, serial connectors, USB connectors, parallel connectors, high-speed data transmission adapters such as fiber optics, USB 6.0, Lightning, wireless network adapters such as Wi-Fi adapters, and telecom (6G, 4G / LTE, etc.) adapters. Devices can connect to a network via interface 507. The network can provide a local area network (LAN), wireless network, cloud computing environment (e.g., software as a service, platform as a service, infrastructure as a service, etc.), client-server, wide area network (WAN), etc.

[0103] Embodiments of this application also provide a computing device, including: a memory, a processor, and a computer program stored on the memory, wherein the processor is configured to execute the computer program to implement the steps of the graph neural network-based image processing method described in various embodiments of this application.

[0104] Embodiments of this application also provide a computer storage medium storing computer program instructions. When executed by a processor, the computer program instructions cause the processor to perform the steps of the graph neural network-based image processing method according to various embodiments of the present invention. The storage medium may include read-only memory (ROM), flash memory, random access memory (RAM), dynamic random access memory (DRAM) such as synchronous DRAM (SDRAM) or Rambus DRAM, static memory (e.g., flash memory, static random access memory), etc., and may store computer-executable instructions thereon in any format.

[0105] Embodiments of this application also provide a computer program product, including a computer program, wherein when the computer program is executed by a processor, it implements the steps of the image processing method based on graph neural networks described in various embodiments of this application.

[0106] Various modifications and alterations can be made to the methods, apparatus, and systems of the present invention. In view of the description and practice of the disclosed systems and related methods, other embodiments can be derived by those skilled in the art. Each claim of the present invention is to be understood as an independent embodiment, and any combination thereof is also used as an embodiment of the present invention, and such embodiments are considered to be included in the present invention.

[0107] The examples are to be considered as exemplary only, and the true scope is indicated by the appended claims and their equivalents.

Claims

1. An image processing method based on graph neural networks, characterized in that, The image processing method includes: via a processor, The centerline of the blood vessel is obtained based on a three-dimensional medical image containing the blood vessel; Multiple nodes are set on each of the aforementioned center lines, and the cross-sections corresponding to each node are obtained; Within each of the aforementioned cross sections, rays pointing towards the pipe wall are obtained along multiple directions centered on the node; Samples are taken from multiple sampling points along each ray, and all sampling point data on each ray are used to construct a feature sequence. Each feature sequence is used as a vertex to construct a graph structure, wherein the vertex includes the intensity change feature of the corresponding ray direction. Based on the graph structure and the trained graph neural network, the target length of each node to the vessel wall in the corresponding ray direction is predicted, and the blood vessel is reconstructed according to the target length in each ray direction.

2. The image processing method according to claim 1, characterized in that, Within each of the cross sections, rays pointing towards the pipe wall are obtained in multiple directions centered on the node, including: within the cross section, multiple spoke-shaped rays are arranged at equal angular intervals centered on the node.

3. The image processing method according to claim 1, characterized in that, The vertex further includes the global geometric features of the blood vessels at the corresponding node location.

4. The image processing method according to claim 1, characterized in that, The image processing method further includes: The vertices of adjacent cross sections corresponding to the same ray angle are connected to each other, the vertices of adjacent ray angles within the same cross section are connected to each other, and the vertices of the same ray angle on cross sections with parent-child relationships at the bifurcation position are connected to each other to form the edges of the graph structure.

5. The image processing method according to claim 1, characterized in that, Training the graph neural network to obtain a trained graph neural network includes: Based on the intersection point of each ray with the actual pipe wall surface, the reference length of each node in the corresponding ray direction is determined; A first loss function is constructed based on the first difference between the target length and the reference length. The parameters of the graph neural network are optimized through backpropagation until the first loss function converges, thus obtaining a trained graph neural network.

6. The image processing method according to claim 5, characterized in that, Training the graph neural network to obtain a trained graph neural network further includes: Based on the second difference in the target length of adjacent cross-sections in the same ray angle direction, a smooth loss along the centerline arc length is constructed. The smoothing loss is weighted and fused with the first difference to obtain a second loss function, and the graph neural network is optimized by backpropagation based on the second loss function.

7. The image processing method according to claim 5 or 6, characterized in that, Training the graph neural network to obtain a trained graph neural network further includes: Based on the third difference in the target length of cross sections with parent-child relationship at the bifurcation location on the blood vessel in the same ray angle direction, a bifurcation consistency loss is constructed. The bifurcation consistency loss is weighted and fused with the first loss function and / or the second loss function to obtain the third loss function, and the graph neural network is optimized by backpropagation based on the third loss function.

8. The image processing method according to any one of claims 5-7, characterized in that, Training the graph neural network to obtain a trained graph neural network further includes: Penalize negative target lengths and target lengths exceeding the threshold to construct a physical prior constraint loss; The physical prior constraint loss is weighted and fused with the first loss function, the second loss function, and / or the third loss function to obtain the fourth loss function, and the graph neural network is optimized by backpropagation based on the fourth loss function.

9. The image processing method according to claim 1, characterized in that, Reconstructing blood vessels based on the target length in each ray direction, including: Based on the nodes of each cross section and the target length of each node in the corresponding ray direction, the contour points of each node in the corresponding ray direction are obtained. Connect the contour points with adjacent ray angles within the same cross section, and connect the contour points with the same ray angle in adjacent cross sections to obtain the reconstructed blood vessel.

10. An image processing device based on a graph neural network, characterized in that, The image processing device includes a processor, the processor being configured to... The centerline of the blood vessel is obtained based on a three-dimensional medical image containing the blood vessel; Multiple nodes are set on each of the aforementioned center lines, and the cross-sections corresponding to each node are obtained; Within each of the aforementioned cross sections, rays pointing towards the pipe wall are obtained along multiple directions centered on the node; Samples are taken from multiple sampling points along each ray, and all sampling point data on each ray are used to construct a feature sequence. Each feature sequence is used as a vertex to construct a graph structure, wherein the vertex includes the intensity change feature of the corresponding ray direction. Based on the graph structure and the trained graph neural network, the target length of each node to the vessel wall in the corresponding ray direction is predicted, and the blood vessel is reconstructed according to the target length in each ray direction.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by the processor, it implements claim 1. The steps of the image processing method based on graph neural networks as described in any one of the 9.

12. A computer program product comprising a computer program, wherein, When the computer program is executed by the processor, it implements claim 1. The steps of the image processing method based on graph neural networks as described in any one of the 9.

Images