Vascular fluid dynamic parameter identification method and model
By constructing vascular map data and using the deep map neural network model to identify vascular fluid dynamics parameters, the problem of complex operation and high cost of traditional methods is solved, and high-precision and low-cost intravascular fluid flow velocity measurement is achieved.
Patent Information
- Application Number
- CN202411983598.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-31
- Publication Date
- 2025-04-08
AI Technical Summary
The traditional intravascular fluid flow rate measurement method is complex, costly and time-consuming, making it difficult to meet the non-invasive, fast and low-cost measurement needs.
By constructing vascular map data, including node features and edge features, the three-dimensional blood vessels are mapped to one-dimensional feature maps, and using the deep map neural network model, including at least a convolutional layer, edge residual connection layer and node residual connection layer, vascular fluid dynamics parameters are identified.
It improves the measurement accuracy of fluid dynamic parameters, reduces costs, and ensures the accuracy and efficiency of measurement.
Smart Images
Figure CN119784737A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present invention relate to the field of medical image processing, and in particular to a method and model for identifying vascular fluid dynamics parameters. Background Art
[0002] With the development of medical imaging technology, medical images play an increasingly important role in disease diagnosis and treatment planning. As one of the important structures of the human body, the accurate identification of blood vessels in images is crucial for a variety of medical procedures. The intravascular fluid flow rate is an important parameter for assessing vascular health and diagnosing related diseases. Although traditional methods for measuring intravascular fluid flow rate, such as Doppler ultrasound and magnetic resonance imaging (MRI), have been widely used in clinical practice, they have limitations such as complex operation, high cost, and long time consumption. With the development of medical imaging technology, the demand for non-invasive, fast, and low-cost methods for measuring intravascular fluid flow rate is growing.
[0003] Based on the above, there is an urgent need to provide a method for identifying vascular fluid dynamics parameters to reduce costs while ensuring the measurement accuracy of fluid dynamics parameters. Summary of the invention
[0004] In view of this, the embodiments of the present invention provide a method and model for identifying vascular fluid dynamics parameters, which can improve the measurement accuracy of fluid dynamics parameters.
[0005] To achieve the above objectives, in a first aspect, an embodiment of the present invention provides a method for identifying vascular fluid dynamics parameters, comprising:
[0006] Constructing blood vessel map data, wherein the blood vessel map data includes node features and edge features, and the blood vessel map data is a feature map that maps three-dimensional blood vessels to one-dimensional through blood vessel centerlines;
[0007] A deep graph neural network model is used to identify vascular fluid dynamics parameters. The deep graph neural network model includes at least a convolution layer, an edge residual connection layer and a node residual connection layer. The convolution layer is used to integrate adjacent nodes and edge features. The edge residual connection layer is used to realize edge features of edge areas and outputs of this layer, and perform residual connection processing; the node residual connection layer is used to realize node features outside the edge area and outputs of this layer, and perform residual connection processing.
[0008] Optionally, the node features include the cross-sectional area of the blood vessel and the node type, and the node type includes an entrance node, an exit node, a bifurcation node, and a node located inside a branch.
[0009] Optionally, the edge feature includes a true distance between two points on the centerline of the blood vessel and an edge type.
[0010] Optionally, the edge residual connection layer is also used to update the edge features when the edge features are updated.
[0011] Optionally, the edge is updated multiple times, and after the first edge update, graph normalization and nonlinear activation function are used for processing, wherein the graph normalization is used to normalize the neighborhood features of the nodes in the graph, and the nonlinear activation function is used to retain more marginalized features.
[0012] Optionally, it also includes: using a ten-fold cross validation method to validate the deep graph neural network model.
[0013] Optionally, the vascular fluid dynamics parameter includes any one of a fluid velocity parameter and a pressure parameter.
[0014] In a second aspect, an embodiment of the present invention provides a vascular fluid dynamics parameter recognition model, such as the deep graph neural network model described above, wherein the deep graph neural network model identifies vascular fluid dynamics parameters based on vascular map data, wherein the vascular map data includes node features and edge features, and the vascular map data is a feature map that maps three-dimensional blood vessels to one dimension through vascular centerlines; the deep graph neural network model includes at least a convolutional layer, an edge residual connection layer, and a node residual connection layer, wherein the convolutional layer is used to integrate adjacent nodes and edge features, and the edge residual connection layer is used to realize edge features of edge areas and outputs of this layer, and perform residual connection processing; the node residual connection layer is used to realize node features outside the edge area and outputs of this layer, and perform residual connection processing.
[0015] It can be seen that the present invention provides a method for identifying vascular fluid dynamics parameters, by constructing vascular map data, the vascular map data includes node features and edge features, the vascular map data is a feature map that maps three-dimensional blood vessels to one-dimensional through the vascular centerline; using a deep graph neural network model to identify vascular fluid dynamics parameters, the deep graph neural network model at least includes a convolution layer, an edge residual connection layer and a node residual connection layer, the convolution layer is used to integrate adjacent nodes and edge features, the edge residual connection layer is used to realize the edge features of the edge area and the output of this layer, and perform residual connection processing; the node residual connection layer is used to realize the node features outside the edge area and the output of this layer, and perform residual connection processing. The edge residual connection layer in the vascular fluid dynamics parameter identification method provided by the embodiment of the present invention can integrate the upper layer network features and the current layer features to maximize the utilization of edge features, retain the influence of edge features as much as possible, and obtain the final recognition result after superimposing with the node features outside the edge area. The retention of edge features can effectively improve the measurement accuracy of fluid dynamics parameters.
[0016] It should also be noted that the deep graph neural network model used in this application is low in cost and can accurately identify dynamic parameters. Moreover, when the present application further adopts graph normalization processing and nonlinear activation function processing in the recognition process, the graph normalization processing in this application improves the stability and convergence speed of the model, and the nonlinear activation function processing in this application increases the expression ability of the model and improves performance. For this reason, the above processing greatly improves the performance of the deep graph neural network model in this application, so that the vascular fluid dynamics parameter identification method provided in the embodiment of the present invention can reduce costs while ensuring the measurement accuracy of fluid dynamics parameters. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0018] Figure 1 A schematic diagram of the steps of a method for identifying vascular fluid dynamics parameters provided by an embodiment of the present invention.
[0019] Figure 2 A schematic diagram of the three-dimensional and one-dimensional mapping relationship corresponding to the blood vessel map provided in an embodiment of the present invention.
[0020] Figure 3 A schematic diagram of the architecture of a deep graph neural network model provided for an embodiment of the present invention.
[0021] Figure 4 A schematic diagram of the processing process of the deep graph neural network model provided by an embodiment of the present invention.
[0022] Figure 5 A parameter diagram of vascular fluid dynamics parameters provided by an embodiment of the present invention.
[0023] Figure 6 A schematic diagram of the results of verification by ten-fold cross validation provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0025] In order to solve the technical problem as described in the background technology, the present invention proposes a novel method to improve the measurement accuracy of fluid dynamics parameters by constructing a vascular map dataset and training a deep graph neural network model.
[0026] The embodiment of the present invention provides a method for identifying vascular fluid dynamics parameters. Figure 1 As shown, including:
[0027] Step S1, constructing blood vessel map data, wherein the blood vessel map data includes node features and edge features, and the blood vessel map data is a feature map that maps three-dimensional blood vessels to one-dimensional through blood vessel centerlines;
[0028] Step S2: determine an initial deep graph neural network model, where the initial deep graph neural network model is a neural network model that has not yet been trained.
[0029] Step S3: Using the database to train the initial deep graph neural network model to generate a deep graph neural network model;
[0030] The samples in the database are divided into a training set and a test set. The training set is used to input the initial deep graph neural network model for training, and the test set is used to detect the training degree of the initial deep graph neural network model. When it meets the preset recognition performance, it can be retained as a deep graph neural network model.
[0031] Step S4, using a deep graph neural network model to identify vascular fluid dynamics parameters, the deep graph neural network model includes at least a convolution layer, an edge residual connection layer and a node residual connection layer, the convolution layer is used to integrate adjacent nodes and edge features, the edge residual connection layer is used to realize edge features of edge areas and outputs of this layer, and perform residual connection processing; the node residual connection layer is used to realize node features outside the edge area and outputs of this layer, and perform residual connection processing.
[0032] Based on the above, the present invention provides a method for identifying vascular fluid dynamics parameters, by constructing vascular map data, the vascular map data includes node features and edge features, the vascular map data is a feature map that maps three-dimensional blood vessels to one-dimensional through the vascular centerline; using a deep graph neural network model to identify vascular fluid dynamics parameters, the deep graph neural network model at least includes a convolution layer, an edge residual connection layer and a node residual connection layer, the convolution layer is used to integrate adjacent nodes and edge features, the edge residual connection layer is used to realize the edge features of the edge area and the output of this layer, and perform residual connection processing; the node residual connection layer is used to realize the node features outside the edge area and the output of this layer, and perform residual connection processing. The edge residual connection layer in the vascular fluid dynamics parameter identification method provided by the embodiment of the present invention can integrate the upper layer network features and the current layer features to maximize the utilization of edge features, retain the influence of edge features as much as possible, and after superimposing with the node features outside the edge area, the final recognition result is obtained, and the retention of edge features can effectively improve the measurement accuracy of fluid dynamics parameters.
[0033] The vascular map data in the embodiment of the present invention may specifically be vascular map data of hepatic blood vessels, which may be a feature map that maps a three-dimensional blood vessel to a one-dimensional feature map through the vascular centerline. Figure 2 As shown, it is a schematic diagram of the three-dimensional and one-dimensional mapping relationship corresponding to the vascular map provided by an embodiment of the present invention. In this mapping relationship, the left side is a three-dimensional schematic diagram, which has a certain curvature. In this schematic diagram, A, A1, and A2 constitute three branches in the vascular map, and the right side is a one-dimensional schematic diagram. In this schematic diagram, a, a1, and a2 constitute three branches in the vascular map data. The one-dimensional schematic diagram on the right can be used as the vascular map data in this application.
[0034] It should also be noted that the blood vessel map data in the embodiment of the present application is composed of node features and edge features.
[0035] The node features include the cross-sectional area of the blood vessel and the node type, and the node type includes the inlet node, the outlet node, the bifurcation node, and the node located inside the branch. Generally speaking, the calculation of one-dimensional blood vessel fluid dynamics parameters is related to the length and topological relationship of the blood vessel, and has nothing to do with the spatial position of the blood vessel. The spatial position of each node is not considered in the node features. i ∈R + is the cross-sectional area of the section perpendicular to the centerline of the blood vessel. Each center point on the centerline is divided into an entry node, an exit node, a bifurcation node, and a node inside a branch, and one-hot encoding α is used. i ∈R 4 To represent different node types. For boundary conditions, is the inlet velocity of the blood vessel at time t, and the simplified model of hemodynamic characteristics is used for calculation at the outlet. R∈R is equal to the vascular resistance R at the outlet endpoint. 1 +R 2 The sum of R 1 The default value of is the characteristic impedance at the endpoint, C∈R is the compliance value of the blood vessel, and for the internal nodes of the blood vessel, the value is 0. In summary, the node features of the blood vessel graph data can be expressed as follows:
[0036]
[0037] It should also be noted that for each node in the vascular graph data, the target state and for The corresponding pressure and velocity. The features (except t and α) are normalized, and the statistics calculated from the data set are used to ensure that the independent features follow the Gaussian distribution N(0,1). Note that V in The value of is obtained from the normalized data. For the outlet boundary condition, only the non-zero elements of the data set are considered during the normalization process.
[0038] The edge features include the actual distance between two points on the center line of the blood vessel and the edge type. ij is the true distance of the arc between the two points on the centerline of the blood vessel, D ij ∈R + , edge type β ij Specifically include: the edge from the internal node to the exit node, the edge from the internal node to the entry node, the edge from the branch start node to the inside of the branch, β ij ∈R 4 .
[0039] One-hot encoding is used to represent different edge types. In summary, the edge features in the vascular map data of the embodiment of the present application are:
[0040] E ij =[D ij ,β ij ]∈R 5 Formula 2
[0041] In the embodiment of the present invention, all edges are bidirectional, and the edge feature vector D is normalized by nodes. ij Normalize.
[0042] For subsequent use, use the matrix The set of all nodes representing the vessel graph data.
[0043] Furthermore, the deep graph neural network model in the present application can be constructed based on the graph convolutional neural network architecture, and by performing convolution operations on vascular graph data, it can capture the complex relationships between nodes and the deep structural features of the graph.
[0044] refer to Figure 3 As shown, Figure 3 Schematic diagram of the architecture of the deep graph neural network model provided by the embodiment of the present invention. In this embodiment, by inputting the vascular graph data constructed by the vascular geometry and boundary conditions, the corresponding vascular fluid dynamics parameters are output. In the deep graph neural network model provided by the embodiment of the present invention, specific reference is made to Figure 4 As shown, it at least includes a convolution layer, an edge residual connection layer and a node residual connection layer. The convolution layer is used to integrate adjacent nodes and edge features. The edge residual connection layer is used to realize the edge features of the edge area and the output of this layer, and perform residual connection processing; the node residual connection layer is used to realize the node features outside the edge area and the output of this layer, and perform residual connection processing.
[0045] For the deep graph neural network model provided by the embodiment of the present invention, simply increasing the depth of the network cannot guarantee better results. In fact, too much depth may lead to the occurrence of over-smoothing, thereby reducing the performance of the network and causing the results to become worse. To this end, a residual connection layer is added in the present application, and the residual connection layer includes a node residual connection layer and an edge residual connection layer. When performing residual connection, the node residual connection layer is used to perform residual connection on the nodes, and the edge residual connection layer is used to perform residual connection on the edges, and the node residual connection layer and the edge residual connection layer are integrated as the recognition result of the deep graph neural network model, and the vascular fluid dynamics parameters are output in the above manner.
[0046] First, calculate the updated edge features within the edge region:
[0047]
[0048] Among them, E is the set of edges, E ij represents the edge between source node i and target node j, W 0 Represents a learnable weight matrix, and information aggregation is performed after the edge features and the two end nodes are multiplied by the corresponding weight matrix. LeakyReLU represents an activation function, which has a small slope in the negative region to allow a certain degree of negative value transmission.
[0049] Furthermore, the edge residual connection layer is also used to update the edge features when the edge features are updated.
[0050] The edge is updated multiple times. After the first edge update, graph normalization and nonlinear activation function are used for processing, wherein the graph normalization is used to normalize the neighborhood features of the nodes in the graph, and the nonlinear activation function is used to retain more marginal features.
[0051] The new node features are calculated using the information of the upper layer and the updated edge features. The node features of the upper layer are normalized and nonlinearly activated to obtain X l′ =LeakyReLU(GN(X l )), and then perform aggregation operations on adjacent nodes:
[0052]
[0053] For node i, there are adjacent nodes and the corresponding edge All adjacent eigenvalues are obtained by combining Furthermore, the SoftMaxAgg aggregation function is used for aggregation, where t is a continuous variable, called the inverse temperature. When t tends to 0, SoftMax(t·)=Mean(·), and when t tends to ∞, SoftMax(t·)=Max(·).
[0054]
[0055] In the embodiment of the present invention, adding residual connections between the input and output of each layer can effectively alleviate the problem of gradient vanishing, so that the graph neural network has a deeper number of layers, and significantly improves the performance of the deep graph neural network model. 1 It can be composed of one or more weight matrices, and can be considered as a multi-layer perceptron. In addition, residual connections are also used for edge layers to retain more residual edge features.
[0056] It should also be noted that after updating the edge, the updated edge needs to be involved in the node aggregation and update operation, that is, the edge will also be added as a parameter to the edge result of the next layer, so as to effectively retain the gradient of the edge.
[0057]
[0058] Except for edge updates, which use graph normalization (BatchNormal), all other normalization layers use normalization (GraphNormal). α, γ, and β represent learnable parameters. Since there is a lot of batch noise in the vascular graph data, the batch statistics cannot be well concentrated near the data set statistics. Using GraphNormal can make the deep graph neural network model converge faster and obtain better generalization performance on the data set.
[0059] It should also be noted that the vascular fluid dynamics parameter includes any one of a fluid velocity parameter and a pressure parameter. Figure 5 Schematic diagram of vascular fluid dynamics parameters provided by an embodiment of the present invention. Figure 5 As shown, two groups of comparison graphs are provided, the upper part is a comparison graph of fluid velocity parameters, the left side is the identification value of the fluid velocity parameter identified by the vascular fluid dynamics parameter identification method, and the right side is the true value of the fluid velocity parameter in the vascular fluid; the lower part is a comparison graph of pressure parameters, the left side is the identification value of the pressure parameter identified by the vascular fluid dynamics parameter identification method, and the right side is the true value of the pressure parameter in the vascular fluid. It can be seen that the difference between the identification and the true value of the two groups of parameters is small.
[0060] In an embodiment of the present invention, further, it also includes: using a ten-fold cross validation method to verify the deep graph neural network model. Figure 6 A schematic diagram of the results of the ten-fold cross validation provided by an embodiment of the present invention. During the validation process, the new vascular map data is evenly divided into ten parts, one for validation and the rest for training. All data need to be shuffled before division to ensure that each sample is independent and identically distributed.
[0061] like Figure 6 As shown, it shows the comparison results of the recognition method with edge residual (as shown in the icon with diamond squares in the figure) and the recognition method without edge residual (as shown in the icon with circular squares in the figure) shown in the present application. From the horizontal point of view, the first row uses the vascular blood flow simulation data set as the new vascular map data to compare the overall loss, pressure error, and velocity error respectively. The second row uses the human aorta blood flow simulation data set as the new vascular map data to compare the overall loss, pressure error, and velocity error respectively. It can be seen that the error of the recognition method with edge residual provided in the present application is smaller than the error of the recognition method without edge residual.
[0062] An embodiment of the present invention also provides a vascular fluid dynamics parameter recognition model, such as the deep graph neural network model described above, wherein the deep graph neural network model identifies vascular fluid dynamics parameters based on vascular map data, wherein the vascular map data includes node features and edge features, and the vascular map data is a feature map that maps three-dimensional blood vessels to one dimension through vascular centerlines; the deep graph neural network model includes at least a convolution layer, an edge residual connection layer, and a node residual connection layer, wherein the convolution layer is used to integrate adjacent nodes and edge features, and the edge residual connection layer is used to realize edge features of edge areas and outputs of this layer, and perform residual connection processing; the node residual connection layer is used to realize node features outside the edge area and outputs of this layer, and perform residual connection processing.
[0063] In an optional embodiment, the node features include a cross-sectional area of the blood vessel and a node type, and the node type includes an inlet node, an outlet node, a bifurcation node, and a node located inside a branch.
[0064] In an optional embodiment, the edge features include the actual distance between two points on the centerline of the blood vessel and the edge type.
[0065] In an optional embodiment, the edge residual connection layer is further used to update the edge features when the edge features are updated.
[0066] In an optional embodiment, the edge is updated multiple times, and after the first edge update, graph normalization and nonlinear activation function are used for processing, wherein the graph normalization is used to normalize the neighborhood features of the nodes in the graph, and the nonlinear activation function is used to retain more marginal features.
[0067] In an optional embodiment, it also includes: using a ten-fold cross-validation method to validate the deep graph neural network model.
[0068] In an optional embodiment, the vascular fluid dynamics parameter includes any one of a fluid velocity parameter and a pressure parameter.
[0069] The above describes multiple implementation schemes provided by the embodiments of the present invention. The various optional methods introduced in the implementation schemes can be combined and cross-referenced with each other without conflict, thereby extending a variety of possible implementation schemes, which can all be considered as implementation schemes disclosed and open in the embodiments of the present invention.
[0070] Although the embodiments of the present invention are disclosed above, the present application is not limited thereto. Any person skilled in the art may make various changes and modifications without departing from the spirit and scope of the present application. Therefore, the protection scope of the present application shall be subject to the scope defined by the claims.
Claims
1. A method for identifying vascular fluid dynamics parameters, characterized in that: include: Constructing blood vessel map data, wherein the blood vessel map data includes node features and edge features, and the blood vessel map data is a feature map that maps three-dimensional blood vessels to one-dimensional through blood vessel centerlines; A deep graph neural network model is used to identify vascular fluid dynamics parameters. The deep graph neural network model includes at least a convolution layer, an edge residual connection layer and a node residual connection layer. The convolution layer is used to integrate adjacent nodes and edge features. The edge residual connection layer is used to realize edge features of edge areas and outputs of this layer, and perform residual connection processing; the node residual connection layer is used to realize node features outside the edge area and outputs of this layer, and perform residual connection processing.
2. The method according to claim 1, characterized in that The node features include the cross-sectional area of the blood vessel and the node type, and the node type includes an entrance node, an exit node, a bifurcation node, and a node located inside a branch.
3. The method according to claim 1, characterized in that The edge features include the actual distance between two points on the center line of the blood vessel and the edge type.
4. The method according to claim 1, characterized in that: The edge residual connection layer is also used to update the edge features when the edge features are updated.
5. The method according to claim 4, characterized in that The edge is updated multiple times. After the first edge update, graph normalization and nonlinear activation function are used for processing, wherein the graph normalization is used to normalize the neighborhood features of the nodes in the graph, and the nonlinear activation function is used to retain more marginal features.
6. The method according to any one of claims 1 to 5, characterized in that Also includes: The deep graph neural network model was validated using the ten-fold cross validation method.
7. The method according to claim 1, characterized in that The vascular fluid dynamics parameter includes any one of a fluid velocity parameter and a pressure parameter.
8. A vascular fluid dynamics parameter identification model, characterized in that: The deep graph neural network model according to any one of claims 1 to 7, wherein the deep graph neural network model identifies vascular fluid dynamic parameters based on vascular map data, wherein the vascular map data includes node features and edge features, and the vascular map data is a feature map that maps three-dimensional blood vessels to one dimension through vascular centerlines; the deep graph neural network model includes at least a convolution layer, an edge residual connection layer, and a node residual connection layer, wherein the convolution layer is used to integrate adjacent nodes and edge features, the edge residual connection layer is used to realize edge features of edge areas and outputs of this layer, and perform residual connection processing; the node residual connection layer is used to realize node features outside the edge area and outputs of this layer, and perform residual connection processing.