Methods for training a predictive model for vascular deformation after stent implantation

Abstract

This disclosure presents a method for training a predictive model for vascular deformation after stent implantation. The method includes: performing 3D reconstruction processing on preoperative and postoperative cranial imaging data before and after stent implantation to obtain a preoperative and postoperative vascular model; mapping the stent pose information from the postoperative cranial imaging data to the preoperative vascular model to generate a stent model; constructing a target vascular map structure and a target stent map structure based on the preoperative and stent models; a first node on the target vascular map structure carrying first physical coding information; a second node on the target stent map structure carrying second physical coding information; and inputting the target vascular map structure, the target stent map structure, and the postoperative vascular model into a graph neural network model for training until a set training termination condition is met to obtain the predictive model. This method enables rapid and accurate prediction of vascular deformation after stent implantation.

Description

Technical Field

[0001] This disclosure generally relates to the field of artificial intelligence. More specifically, this disclosure relates to a method for training a predictive model for predicting vascular deformation after stent implantation. Background Technology

[0002] With the widespread use of interventional vascular therapy in vascular diseases (such as intracranial aneurysms and intracranial vascular stenosis), stent implantation has become a common clinical treatment method. After implantation, the stent causes local expansion and remodeling of the blood vessel, leading to deformation of the vessel wall. In actual stent implantation surgery, it is necessary to predict the deformation of the blood vessel after stent implantation before proceeding, which can assist the surgeon in choosing the optimal stent implantation strategy during the procedure.

[0003] Currently, the finite element method (FEM) or other numerical simulation methods are commonly used to predict vascular deformation. Although these methods offer high prediction accuracy, they involve large computational loads and are time-consuming, making them unsuitable for real-time surgical planning and intraoperative monitoring.

[0004] Therefore, there is an urgent need to provide a scheme for training a predictive model to predict vascular deformation after stent implantation, so as to use the trained predictive model to quickly and accurately predict vascular deformation after stent implantation. Summary of the Invention

[0005] In order to at least address one or more of the technical problems mentioned above, this disclosure proposes a scheme for training a predictive model for predicting vascular deformation after stent implantation in several aspects.

[0006] In a first aspect, this disclosure provides a method for training a predictive model for predicting vascular deformation after stent implantation. The method includes: acquiring a pair of cranial imaging training data; the cranial imaging training data pair includes preoperative cranial imaging data before stent implantation and postoperative cranial imaging data after stent implantation; performing three-dimensional reconstruction processing on the preoperative cranial imaging data and the postoperative cranial imaging data respectively to obtain a preoperative vascular model and a postoperative vascular model; mapping the pose information of the stent in the postoperative cranial imaging data to the preoperative vascular model to generate a stent model; constructing a target vascular map structure and a target stent map structure based on the preoperative vascular model and the stent model; wherein, a first node on the target vascular map structure is a sampling point on the vascular wall in the preoperative vascular model, and the first node carries first physical coding information; a second node on the target stent map structure is a sampling point on the stent in the stent model, and the second node carries second physical coding information; inputting the target vascular map structure, the target stent map structure, and the postoperative vascular model into a pre-constructed graph neural network model for training until a set training termination condition is reached to obtain the predictive model.

[0007] In some embodiments, constructing a target vascular map structure based on the preoperative vascular model and the stent model includes: marking the contact area between the stent and the blood vessel on the preoperative vascular model; sampling vertices on the blood vessel wall of the preoperative vascular model to obtain a first node; performing physical encoding on the first node in the contact area of ​​the preoperative vascular model based on the node information of the first node and the stent information to obtain the first physical encoding information of the first node; and performing physical encoding on the first node in the non-contact area of ​​the preoperative vascular model based on the node information of the first node to obtain the first physical encoding information of the first node, thereby forming the target vascular map structure.

[0008] In some embodiments, the graph neural network model includes: an encoder, an attention module, and a decoder; the attention module includes a multi-head attention unit; the step of inputting the target vascular map structure, the target stent map structure, and the postoperative vascular model into the pre-constructed graph neural network model for training includes: inputting the target vascular map structure and the target stent map structure into the encoder for feature extraction to obtain vascular features and stent features respectively; inputting the vascular features and stent features into the attention module for feature fusion to obtain fused features; inputting the fused features into the decoder for decoding to obtain a deformed vascular model; calculating a loss value based on the postoperative vascular model and the deformed vascular model, and adjusting the model parameters of the graph neural network model based on the loss value.

[0009] In some embodiments, the loss value includes: a first loss value and a second loss value; the first loss value is calculated using a mean squared error loss function, and the second loss value is calculated using... Figure 1 The loss function of the consistency regularization term is calculated.

[0010] In some embodiments, the graph neural network model includes the TAGConv architecture.

[0011] In some embodiments, after obtaining the prediction model, the method further includes: acquiring a target vascular model for vascular deformation prediction and a candidate stent model of at least one candidate stent; for each candidate stent model, constructing a candidate vascular map structure based on the target vascular model and the candidate stent model, and constructing a candidate stent map structure based on the candidate stent model; for each candidate stent map structure, inputting the candidate stent map structure and the corresponding candidate vascular map structure into the prediction model for prediction to obtain a candidate vascular prediction model; and determining the target stent from at least one candidate stent based on each candidate vascular prediction model.

[0012] In some embodiments, the method further includes: obtaining an actual vascular model after the target stent is implanted into a blood vessel; and optimizing the prediction model based on the actual vascular model and a candidate vascular prediction model corresponding to the target stent.

[0013] In a second aspect, this disclosure provides an apparatus for training a predictive model for predicting vascular deformation after stent implantation. The apparatus includes: a brain image training data pair acquisition module for acquiring brain image training data pairs; the brain image training data pairs include preoperative brain image data before stent implantation and postoperative brain image data after stent implantation; a three-dimensional reconstruction module for performing three-dimensional reconstruction processing on the preoperative brain image data and the postoperative brain image data respectively to obtain a preoperative vascular model and a postoperative vascular model; a stent model generation module for mapping the stent pose information from the postoperative brain image data to the preoperative vascular model to generate a stent model; and a graph structure construction module. A block is used to construct a target vascular map structure and a target stent map structure based on the preoperative vascular model and the stent model; wherein, the first node on the target vascular map structure is a sampling point on the vascular wall in the preoperative vascular model, and the first node carries first physical coding information; the second node on the target stent map structure is a sampling point on the stent in the stent model, and the second node carries second physical coding information; a training module is used to input the target vascular map structure, the target stent map structure, and the postoperative vascular model into the pre-constructed graph neural network model for training until a set training termination condition is reached to obtain the prediction model.

[0014] In a third aspect, this disclosure provides an electronic device comprising: a processor configured to execute program instructions; and a memory configured to store the program instructions, which, when loaded and executed by the processor, cause the processor to perform the steps of a method for training a predictive model for predicting vascular deformation after stent implantation as described in the first aspect or any alternative embodiment of the first aspect.

[0015] In a fourth aspect, this disclosure provides a computer-readable storage medium storing program instructions, characterized in that, when the program instructions are loaded and executed by a processor, the processor performs the steps of a method for training a predictive model for predicting vascular deformation after stent implantation as described in the first aspect or any optional embodiment of the first aspect.

[0016] Using the method described above for training a predictive model for vascular deformation after stent implantation, this embodiment of the present disclosure obtains a preoperative vascular model and a postoperative vascular model by performing three-dimensional reconstruction processing on preoperative and postoperative cranial imaging data, respectively. Then, the pose information of the stent in the postoperative cranial imaging data is mapped to the preoperative vascular model to generate a stent model. The target vascular map structure and the target stent map structure are constructed to train the predictive model, so that the predictive model can predict vascular deformation. Compared with the finite element method (FEM) or other numerical simulation methods for deformation prediction, it can predict more quickly and accurately, thereby realizing real-time prediction of complex vascular deformation. Attached Figure Description

[0017] The above and other objects, features, and advantages of exemplary embodiments of this disclosure will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this disclosure are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein:

[0018] Figure 1 An exemplary flowchart of a method for training a predictive model for predicting vascular deformation after stent implantation, according to some embodiments of this disclosure, is shown.

[0019] Figure 2 A schematic diagram of the model architecture of some embodiments of this disclosure is shown;

[0020] Figure 3 An exemplary structural block diagram of an apparatus for training a predictive model for predicting vascular deformation after stent implantation, according to some embodiments of this disclosure, is shown.

[0021] Figure 4 An exemplary structural block diagram of an electronic device according to some embodiments of this disclosure is shown. Detailed Implementation

[0022] The technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, not all of them. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0023] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0024] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0025] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0026] The specific embodiments disclosed herein will now be described in detail with reference to the accompanying drawings.

[0027] Exemplary application scenarios

[0028] In the medical field, especially in the treatment of vascular diseases, stent implantation is a common surgical procedure. However, predicting vascular deformation after stent implantation is crucial for surgical planning and postoperative evaluation. Traditional prediction methods often employ the finite element method (FEM) or other numerical simulation methods for deformation prediction, but these methods are computationally intensive and time-consuming, making them unsuitable for real-time surgical planning and intraoperative monitoring.

[0029] In view of this, the present disclosure provides a method for training a prediction model for predicting vascular deformation after stent implantation, so as to use the trained prediction model to quickly and accurately predict vascular deformation after stent implantation.

[0030] Figure 1 An exemplary flowchart of a method 100 for training a predictive model for predicting vascular deformation after stent implantation, according to some embodiments of this disclosure, is shown. It is understood that the method 100 for training a predictive model for predicting vascular deformation after stent implantation can be performed by any suitable device with data processing capabilities, such as, but not limited to, terminal devices, processors, and servers.

[0031] like Figure 1 As shown, the method 100 for training a prediction model for predicting vascular deformation after stent implantation provided in this disclosure includes: Step S110: acquiring a pair of cranial imaging training data; Step S120: performing three-dimensional reconstruction processing on the preoperative cranial imaging data and the postoperative cranial imaging data respectively to obtain a preoperative vascular model and a postoperative vascular model; Step S130: mapping the pose information of the stent in the postoperative cranial imaging data to the preoperative vascular model to generate a stent model; Step S140: constructing a target vascular map structure and a target stent map structure based on the preoperative vascular model and the stent model; wherein, the first node on the target vascular map structure is a sampling point on the vascular wall in the preoperative vascular model, and the first node carries first physical coding information; the second node on the target stent map structure is a sampling point on the stent in the stent model, and the second node carries second physical coding information; Step S150: inputting the target vascular map structure, the target stent map structure, and the postoperative vascular model into the constructed graph neural network model for training until the set training termination condition is reached to obtain a prediction model.

[0032] For example, in this disclosed embodiment, the brain imaging training data pair in step S110 refers to a set of brain imaging data obtained from clinical practice for training the prediction model. Specifically, the brain imaging training data pair includes preoperative brain imaging data before stent implantation and postoperative brain imaging data after stent implantation.

[0033] In this disclosed embodiment, the aforementioned preoperative and postoperative cranial imaging data are image data of the internal structure of the brain obtained through medical imaging technology, such as any one of CT angiography (CTA), magnetic resonance angiography (MRA), or digital subtraction angiography (DSA).

[0034] Therefore, the aforementioned preoperative and postoperative cranial imaging data were acquired using specialized medical imaging equipment, such as computed tomography (CT), magnetic resonance imaging (MRI), ultrasound imaging, positron emission tomography (PET), and so on.

[0035] In practice, before stent implantation surgery of the cerebral blood vessels, specialized imaging equipment (CT, MRI, or DSA) is used to scan the patient's brain, collecting preoperative brain imaging data that reflects the original morphology, location, and size of the blood vessels before stent implantation. After the cerebral blood vessel stent implantation surgery, the same imaging equipment is used to scan the patient's brain again, collecting postoperative brain imaging data that reflects changes in blood vessels and other tissues after stent implantation, including vascular dilation and tortuosity. In this embodiment, preoperative and postoperative brain imaging data for each stent implantation surgery can be stored for direct retrieval when needed.

[0036] It should be noted that the aforementioned preoperative and postoperative cranial imaging data can only be used for model training with the patient's consent or after necessary privacy processing.

[0037] For example, in this disclosed embodiment, the three-dimensional reconstruction processing in step S120 refers to a method for converting two-dimensional medical image data into a three-dimensional model. This three-dimensional model can intuitively display the spatial structure of blood vessels, including the branches and curvature of the vessels. In this disclosed embodiment, during the three-dimensional reconstruction of preoperative and postoperative cranial image data, specialized medical image processing software can be used to convert the two-dimensional preoperative and postoperative cranial image data into three-dimensional preoperative and postoperative vascular models through a series of algorithms (e.g., surface reconstruction algorithms, voxel reconstruction algorithms, etc.).

[0038] Specifically, taking preoperative cranial imaging data as an example, the vascular contours are first extracted from the preoperative cranial imaging data using image segmentation algorithms. Then, the extracted vascular contours are reconstructed based on 3D reconstruction technology to obtain the preoperative vascular model. The process of performing 3D reconstruction processing on postoperative cranial imaging data to obtain the postoperative vascular model is the same as the process of obtaining the preoperative vascular model, and will not be repeated here.

[0039] For example, in this disclosed embodiment, the stent pose information in step S130 may include stent location information, stent orientation, angle, shape, size, degree of curvature, etc., which can characterize the specific placement position of the stent within the blood vessel. In this disclosed embodiment, the stent pose information can be extracted from postoperative cranial imaging data using methods such as image recognition and image registration.

[0040] Mapping the stent's pose information to the preoperative vascular model refers to spatially aligning the extracted stent pose information with the preoperative vascular model. Specifically, a specified algorithm (e.g., rigid registration algorithm, non-rigid registration algorithm, etc.) can be used to map the stent's pose information into the preoperative vascular model. Based on the mapped stent pose information, a virtual three-dimensional stent model is then constructed within the preoperative vascular model. Of course, the embodiments disclosed herein can also simulate the target stent model using other methods, such as the finite element method (FEA) simulation; this disclosure does not specifically limit this approach.

[0041] For example, in this disclosed embodiment, the target vascular graph structure in step S140 is a graph structure constructed based on the preoperative vascular model. The first node on the target vascular graph structure is a sampling point on the vascular wall in the preoperative vascular model. The first node carries first physical encoding information, used to describe the physical characteristics and other related information at the node. In this disclosed embodiment, the first physical encoding information may include the node's geometric information and mechanical information. Of course, the first physical encoding information may vary depending on the region where the node is located (e.g., the contact area or non-contact area between the stent and the blood vessel). For example, when the first node is a node in the contact area between the stent and the blood vessel, the first physical encoding information may also include stent information. This disclosed embodiment enriches the information of the nodes by encoding stent information into the nodes of the graph structure during the construction of the graph structure.

[0042] In this disclosed embodiment, the geometric information of the nodes may include the node's coordinates, connection relationships, normal vector, Gaussian curvature, etc.; the mechanical information of the nodes may include the node's mass, stiffness, and damping coefficient, etc.; the stent information may include stent type (e.g., drug-eluting type, bare metal type, etc.), stent dimensions (e.g., diameter, length), etc. This disclosed embodiment does not specifically limit the geometric information, mechanical information, and stent information of the nodes, and these can be determined according to actual circumstances.

[0043] In this disclosed embodiment, the target support structure is a graph structure constructed based on the support model. The second node on this graph is a sampling point on the support in the support model, and the second node carries second physical encoding information. Here, the second physical encoding information may include the geometric information and mechanical information of the node. For details on the geometric and mechanical information of the node, please refer to the description in the above embodiment; it will not be repeated here.

[0044] As for how to construct the target vascular map structure based on the preoperative vascular model and the stent model, and how to construct the target stent map structure based on the stent model, the following examples provide examples, which will not be elaborated here.

[0045] Exemplarily, in this disclosed embodiment, the graph neural network model in step S150 is a neural network model specifically designed for processing graph-structured data. It can learn the features of nodes and edges in a graph structure and the relationships between them. Specifically, it can employ a lightweight and efficient GNN architecture, such as convolutional mechanisms like TAGConv and GCN. As a specific example of this disclosure, such as... Figure 2 As shown, the graph neural network model may include an encoder and a decoder, both of which may include multiple layers (e.g., 3 layers) of graph convolutional layers. In this disclosed embodiment, in addition to the encoder and decoder, the graph neural network model may also introduce a cross-region multi-head attention mechanism (e.g., Figure 2 The cross-attention module in the model enables it to capture the complex interaction between the stent and the vascular region.

[0046] In this disclosed embodiment, the above-mentioned training termination condition can be of many kinds, such as reaching a set number of iterations (e.g., 500 times), or the loss value converging, etc. This disclosed embodiment does not specifically limit the above-mentioned training termination condition.

[0047] In this disclosed embodiment, the input of the target vascular map structure, the target stent map structure, and the postoperative vascular model into the pre-constructed graph neural network model for training can be specifically as follows: based on the graph neural network model, feature extraction, feature fusion, and feature decoding are performed on the target vascular map structure and the target stent map structure respectively to obtain the deformed vascular model. Then, the postoperative vascular model is used as the target data to calculate the loss value. The model parameters of the graph neural network model are then adjusted based on the calculated loss value until the above-mentioned training termination condition is met, and the prediction model can be obtained.

[0048] This disclosed embodiment obtains a preoperative vascular model and a postoperative vascular model by performing three-dimensional reconstruction processing on preoperative and postoperative cranial imaging data, respectively. Then, the pose information of the stent in the postoperative cranial imaging data is mapped to the preoperative vascular model to generate a stent model. The target vascular map structure and the target stent map structure are constructed and trained to generate a prediction model, so that the prediction model can predict vascular deformation. Compared with the finite element method (FEM) or other numerical simulation methods for deformation prediction, it can make predictions faster and more accurate, thereby realizing real-time prediction of complex vascular deformation.

[0049] The following describes the specific implementation method of constructing the target vascular map structure based on the preoperative vascular model and stent model in step S140. Specifically: the contact area between the stent and the blood vessel is marked on the preoperative vascular model; vertices on the blood vessel wall of the preoperative vascular model are sampled to obtain the first node; for the first node in the contact area on the preoperative vascular model, physical encoding is performed based on the node information of the first node and the stent information to obtain the first physical encoding information of the first node; and for the first node in the non-contact area on the preoperative vascular model, physical encoding is performed based on the node information of the first node to obtain the first physical encoding information of the first node, so as to form the target vascular map structure.

[0050] For example, in this disclosed embodiment, after constructing a stent model in a preoperative vascular model, the contact area between the stent and the blood vessel can be directly marked on the preoperative vascular model. In this disclosed embodiment, when specifically marking the contact area, it can be based on manual marking, or it can be done by calculating the spatial distance between each vertex on the preoperative vascular model and the vascular model. If the spatial distance is less than a set threshold (e.g., 0.1 mm), the vertex is considered to belong to the contact area. This disclosed embodiment does not specifically limit the method for determining the contact area.

[0051] In this disclosed embodiment, when calibrating the contact area, it can be calibrated by numbers, for example, marking the vertices of the contact area as "1"; it can also be calibrated by color, for example, marking the vertices of the contact area as "red"; this disclosed embodiment does not specifically limit the specific calibration method.

[0052] In this disclosed embodiment, if every vertex on the preoperative vascular model is physically encoded when constructing the target vascular map structure, the computational load would be extremely large. Therefore, in this disclosed embodiment, before constructing the target vascular map structure, the vertices on the preoperative vascular model are sampled to obtain the first node (i.e., the sampling point). Specific sampling methods can be, for example, uniform sampling or non-uniform sampling based on the geometric features of the vessel wall (e.g., curvature, thickness, etc.). This disclosed embodiment does not specifically limit the sampling method.

[0053] After obtaining the first node, physical encoding is performed on each first node to form the target vascular map structure. Specifically, if the first node is a vertex of the contact area, physical encoding is performed based on the node information of the first node (i.e., the geometric information and mechanical information of the node) and the stent information (i.e., the stent model, stent size, etc.) to obtain the first physical encoding information of the first node; if the first node is a vertex of the non-contact area, physical encoding is performed based on the node information of the first node to obtain the first physical encoding information of the first node.

[0054] In this disclosed embodiment, in order to obtain physical feature encoding with higher discriminative ability, when physically encoding each first node, the above-mentioned node information, scaffold information, etc. can be adaptively fused through a multilayer perceptual mechanism (MLP) or a self-attention mechanism.

[0055] The following describes the specific implementation method of constructing the target scaffold graph structure based on the scaffold model in step S140. Specifically: the vertices on the scaffold model are sampled to obtain the second node; for the second node, physical encoding is performed based on the node information of the second node to obtain the second physical encoding information of the second node, so as to form the target scaffold graph structure.

[0056] For example, in this disclosed embodiment, if each vertex on the scaffolding model is physically encoded when constructing the target scaffolding graph structure, the computational load would be very large. Therefore, in this disclosed embodiment, before constructing the target scaffolding graph structure, the vertices on the scaffolding model are sampled to obtain the second node (i.e., the sampling point). The specific sampling method is the same as the sampling method for the first node described above, and will not be repeated here.

[0057] In this disclosed embodiment, after obtaining the second node, physical encoding is performed on each second node to form the target scaffold graph structure. Specifically, for each second node, physical encoding is performed based on the node information of the second node (i.e., node coordinates, node connection relationships, node normal vector, node Gaussian curvature, scaffold material, scaffold shape, etc.) to obtain the second physical encoding information of that second node. Similarly, in this disclosed embodiment, in order to obtain physical feature encoding with higher discriminative power, when physically encoding each second node, the adaptive fusion of the above-mentioned node information, scaffold information, etc., can be achieved through a multilayer perceptual mechanism (MLP) or a self-attention mechanism.

[0058] The training method for the prediction model is described below with an example. Specifically, the graph neural network model includes an encoder, an attention module, and a decoder; the attention module includes multi-head attention units; the target vascular map structure, the target stent map structure, and the postoperative vascular model are input into the pre-constructed graph neural network model for training, including: inputting the target vascular map structure and the target stent map structure into the encoder for feature extraction to obtain vascular features and stent features respectively; inputting the vascular features and stent features into the attention module for feature fusion to obtain fused features; inputting the fused features into the decoder for decoding to obtain the deformed vascular model; calculating the loss value based on the postoperative vascular model and the deformed vascular model, and adjusting the model parameters of the graph neural network model based on the loss value.

[0059] For example, the encoder is used to extract features from the input graph structure data. In this disclosed embodiment, the encoder encodes the node features in the target vascular graph structure and the target stent graph structure to extract vascular features and stent features. The encoder is typically composed of a multi-layer neural network, where each layer transforms and extracts the input features, ultimately obtaining a high-dimensional feature vector, for example, Figure 2 In this model, the encoder consists of three graph convolutional layers.

[0060] For example, the attention module's function is to fuse the vascular features and stent features output by the encoder. In this disclosed embodiment, it includes a multi-head attention unit capable of weighted fusion of features from different perspectives, enabling the model to capture the complex interaction effects between the stent and the blood vessel, thereby obtaining a more comprehensive fused feature set. For example, Figure 2 In this module, the attention module includes three multi-head attention units.

[0061] For example, the decoder is used to decode the fused features and reconstruct the deformed pattern to obtain a deformed vascular model. Specifically, the decoder predicts the deformation of the blood vessel after stent implantation based on the fused features, including displacement of the vessel wall and changes in the size of the vessel lumen. For example, Figure 2 In the decoder, there are 3 graph convolutional layers.

[0062] For example, after obtaining the deformed vascular model, a loss value is calculated based on the postoperative vascular model and the deformed vascular model. The loss value can reflect the difference between the predicted result and the actual result. In this disclosed embodiment, the loss value is calculated using a loss function. The loss value in this disclosed embodiment may include a first loss value and a second loss value; wherein, the first loss value is calculated using the Mean Squared Error (MSE) loss function, which can be used to measure the error between the predicted deformation and the actual deformation; the second loss value is calculated using... Figure 1The consistency regularization (GCR) loss function is calculated, which can ensure the continuity and smoothness of vascular deformation by constraining the local gradient of the prediction results.

[0063] In this disclosed embodiment, the aforementioned loss value can be the sum of the first loss value and the second loss value, or it can be a weighted average of the first loss value and the second loss value. This disclosed embodiment does not specifically limit the aforementioned loss value. After obtaining the loss value, the model parameters of the graph neural network model (e.g., the model's weights and biases, etc.) are adjusted based on the loss value.

[0064] As an optional embodiment of this disclosure, after obtaining the prediction model, prediction can be performed during surgery based on the trained prediction model. Specifically, this includes: obtaining a target vessel model for vessel deformation prediction and candidate stent models of at least one candidate stent; for each candidate stent model, constructing a candidate vessel map structure based on the target vessel model and the candidate stent model, and constructing a candidate stent map structure based on the candidate stent model; for each candidate stent map structure, inputting the candidate stent map structure and the corresponding candidate vessel map structure into the prediction model for prediction to obtain a candidate vessel prediction model; and determining the target stent from at least one candidate stent based on each candidate vessel prediction model.

[0065] For example, in this disclosed embodiment, the target vascular model for vascular deformation prediction refers to the specific vascular model whose vascular deformation needs to be predicted after stent implantation, which can be obtained by three-dimensional reconstruction of cranial imaging data of a specific patient. A candidate stent refers to a series of stents of different types, sizes, structures, etc., which serve as potential candidates for implantation into the target vascular vessel. A candidate stent model refers to a three-dimensional model of the candidate stent; in this disclosed embodiment, the aforementioned candidate stent model can be obtained from a pre-built stent model database or other sources.

[0066] In this disclosed embodiment, the method for constructing the candidate vascular map structure and the candidate stent map structure is the same as the method for constructing the target vascular map structure and the target stent map structure described above, and will not be repeated here.

[0067] In this disclosed embodiment, for each candidate stent image structure, the candidate stent image structure and the corresponding candidate vessel image structure are input into a prediction model for prediction, resulting in a candidate vessel prediction model. Then, based on each candidate vessel prediction model, a target stent is determined from at least one candidate stent. Here, the target stent refers to the stent most suitable for implantation into the target vessel. During selection, the target stent can be determined from at least one candidate stent using certain evaluation indicators (e.g., the rationality of vessel deformation, stent-vessel compatibility, etc.).

[0068] As an optional embodiment of this disclosure, the method 100 for training the prediction model for predicting vascular deformation after stent implantation further includes: obtaining the actual vascular model after the target stent is implanted into the vascular body; and optimizing the prediction model based on the actual vascular model and the candidate vascular prediction model corresponding to the target stent.

[0069] For example, in this disclosed embodiment, the actual vascular model refers to a model that realistically reflects the vascular morphology, obtained through medical imaging equipment and three-dimensional reconstruction after the target stent is implanted into the blood vessel. Once the actual vascular model is obtained, it is compared and analyzed with the candidate vascular prediction model corresponding to the target stent. Based on the differences between the two, appropriate optimization algorithms are used to perform parameter adjustments and other optimization operations on the prediction model, further improving the accuracy and reliability of the prediction model.

[0070] The trained prediction model in this embodiment can be deployed on a surgical planning platform to enable preoperative simulation and real-time intraoperative monitoring, assisting doctors in selecting the best implantation plan based on different stent parameters.

[0071] This disclosed embodiment organically integrates physical modeling with data-driven deep learning methods. In addition to the original vascular diagram structure and physical parameters, stent parameters are introduced, encoding the stent type (such as drug-eluting type, bare metal type, etc.), size (diameter, length), and placement position (relative to the geometric center of the blood vessel and local feature areas). This approach combines the accuracy of traditional physical simulation methods with the advantages of high computational efficiency and strong generalization ability of deep learning models, thereby enabling real-time prediction of complex vascular deformations. This technical solution can be widely applied to clinical interventional therapy, surgical planning, and risk assessment.

[0072] Figure 3 An exemplary structural block diagram of an apparatus 300 for training a predictive model for predicting vascular deformation after stent implantation, according to some embodiments of this disclosure, is shown.

[0073] like Figure 3As shown in the present disclosure, an apparatus 300 for training a predictive model for predicting vascular deformation after stent implantation includes: a brain image training data pair acquisition module 310, used to acquire brain image training data pairs; the brain image training data pairs include preoperative brain image data before stent implantation and postoperative brain image data after stent implantation; a three-dimensional reconstruction module 320, used to perform three-dimensional reconstruction processing on the preoperative brain image data and the postoperative brain image data respectively to obtain a preoperative vascular model and a postoperative vascular model; and a stent model generation module 330, used to map the pose information of the stent in the postoperative brain image data to the preoperative vascular model to generate a stent model. The system includes a graph structure construction module 340, which is used to construct a target vascular graph structure based on the preoperative vascular model and the stent model, and a target stent graph structure based on the stent model. The first node on the target vascular graph structure is a sampling point on the vascular wall in the preoperative vascular model, and the first node carries first physical coding information. The second node on the target stent graph structure is a sampling point on the stent in the stent model, and the second node carries second physical coding information. A training module 350 is used to input the target vascular graph structure, the target stent graph structure, and the postoperative vascular model into the pre-constructed graph neural network model for training until a set training termination condition is met, thus obtaining a prediction model.

[0074] As an optional embodiment of this disclosure, the aforementioned three-dimensional reconstruction module 320 is specifically used for: marking the contact area between the stent and the blood vessel on the preoperative vascular model; sampling the vertices on the blood vessel wall of the preoperative vascular model to obtain a first node; performing physical encoding on the first node in the contact area of ​​the preoperative vascular model based on the node information of the first node and the stent information to obtain the first physical encoding information of the first node; and performing physical encoding on the first node in the non-contact area of ​​the preoperative vascular model based on the node information of the first node to obtain the first physical encoding information of the first node, so as to form a target vascular map structure.

[0075] As an optional embodiment of this disclosure, the graph neural network model includes: an encoder, an attention module, and a decoder; the attention module includes a multi-head attention unit; the training module 350 is specifically used for: inputting the target vascular image structure and the target stent image structure into the encoder for feature extraction to obtain vascular features and stent features; inputting the vascular features and stent features into the attention module for feature fusion to obtain fused features; inputting the fused features into the decoder for decoding to obtain the deformed vascular model; calculating the loss value based on the postoperative vascular model and the deformed vascular model, and adjusting the model parameters of the graph neural network model based on the loss value.

[0076] As an optional embodiment of this disclosure, the aforementioned loss value includes: a first loss value and a second loss value; the first loss value is calculated using a mean squared error loss function, and the second loss value is calculated using... Figure 1 The loss function of the consistency regularization term is calculated.

[0077] As an optional embodiment of this disclosure, the graph neural network model includes the TAGConv architecture.

[0078] As an optional embodiment of this disclosure, the apparatus 300 for training a prediction model for predicting vascular deformation after stent implantation further includes: a model acquisition module for acquiring a target vascular model for vascular deformation prediction and candidate stent models of at least one candidate stent; a candidate graph structure construction module for constructing a candidate vascular graph structure based on the target vascular model and the candidate stent model for each candidate stent model, and constructing a candidate stent graph structure based on the candidate stent model; a prediction module for inputting the candidate stent graph structure and the corresponding candidate vascular graph structure into the prediction model for prediction to obtain a candidate vascular prediction model; and a target stent determination module for determining a target stent from at least one candidate stent based on each candidate vascular prediction model.

[0079] As an optional embodiment of this disclosure, the apparatus 300 for training a prediction model for predicting vascular deformation after stent implantation further includes: an actual vascular model acquisition module for acquiring an actual vascular model after the target stent is implanted into the vascular body; and an optimization module for optimizing the prediction model based on the actual vascular model and the candidate vascular prediction model corresponding to the target stent.

[0080] Correspondingly, this disclosure also provides Figure 3 The hardware structure diagram of the device shown is as follows: Figure 4 As shown, the electronic device 400 can be a device for implementing the method 100 described above. For example... Figure 4 As shown, the electronic device 400 includes a processor 410 and a memory 420. The memory 420 is configured to store program instructions; the processor 410 is configured to load and execute the program instructions stored in the memory 420 to implement an embodiment of the method for training a predictive model for predicting vascular deformation after stent implantation, as shown above.

[0081] As one embodiment, memory 420 can be any electronic, magnetic, optical, or other physical storage device that can contain or store information such as program instructions, data, etc. For example, memory 420 can be volatile memory, non-volatile memory, or similar storage media. Specifically, memory 420 can be RAM (Random Access Memory), flash memory, storage drives (such as hard disk drives), solid-state drives, any type of storage disk (such as optical discs, DVDs, etc.), or similar storage media, or combinations thereof.

[0082] While numerous embodiments of this disclosure have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and intent of this disclosure. It should be understood that various alternatives to the embodiments of this disclosure described herein may be employed in the practice of this disclosure. The appended claims are intended to define the scope of this disclosure and therefore cover equivalents or alternatives within the scope of these claims.

Claims

1. A method for training a predictive model for vascular deformation after stent implantation, characterized in that, The method includes: Acquire brain imaging training data pairs; the brain imaging training data pairs include preoperative brain imaging data before stent implantation and postoperative brain imaging data after stent implantation. The preoperative and postoperative cranial imaging data were respectively processed into three-dimensional reconstruction to obtain the preoperative and postoperative vascular models. The pose information of the stent in the postoperative cranial imaging data is mapped to the preoperative vascular model to generate a stent model; A target vascular map structure is constructed based on the preoperative vascular model and the stent model, and a target stent map structure is constructed based on the stent model; wherein, the first node on the target vascular map structure is a sampling point on the vascular wall in the preoperative vascular model, and the first node carries first physical coding information; the second node on the target stent map structure is a sampling point on the stent in the stent model, and the second node carries second physical coding information; The target vascular map structure, the target stent map structure, and the postoperative vascular model are input into the pre-constructed graph neural network model for training until the set training termination condition is met, thus obtaining the prediction model.

2. The method according to claim 1, characterized in that, The construction of the target vascular map structure based on the preoperative vascular model and the stent model includes: The contact area between the stent and the blood vessel is marked on the preoperative vascular model; The first node is obtained by sampling the vertices on the vessel wall of the preoperative vascular model. For the first node in the contact area of ​​the preoperative vascular model, physical encoding is performed based on the node information and stent information of the first node in the contact area to obtain the first physical encoding information of the first node in the contact area; and for the first node in the non-contact area of ​​the preoperative vascular model, physical encoding is performed based on the node information of the first node in the non-contact area to obtain the first physical encoding information of the first node in the non-contact area, so as to form the target vascular map structure.

3. The method according to claim 1, characterized in that, The graph neural network model includes an encoder, an attention module, and a decoder; the attention module includes a multi-head attention unit; the step of inputting the target vascular map structure, the target stent map structure, and the postoperative vascular model into the pre-constructed graph neural network model for training includes: The target vascular image structure and the target stent image structure are respectively input into the encoder for feature extraction to obtain vascular features and stent features; The vascular features and the stent features are input into the attention module for feature fusion to obtain fused features; The fused features are input into the decoder for decoding to obtain the deformed blood vessel model; The loss value is calculated based on the postoperative vascular model and the deformed vascular model, and the model parameters of the graph neural network model are adjusted based on the loss value.

4. The method according to claim 3, characterized in that, The loss values ​​include: a first loss value and a second loss value; the first loss value is calculated using the mean squared error loss function, and the second loss value is calculated using the graph consistency regularization term loss function.

5. The method according to claim 1, characterized in that, The graph neural network model includes the TAGConv architecture.

6. The method according to claim 1, characterized in that, After obtaining the prediction model, the method further includes: Obtain the target vascular model for vascular deformation prediction and at least one candidate stent model. For each candidate stent model, a candidate vessel map structure is constructed based on the target vessel model and the candidate stent model, and a candidate stent map structure is constructed based on the candidate stent model. For each candidate stent image structure, the candidate stent image structure and the corresponding candidate vessel image structure are input into the prediction model for prediction, and a candidate vessel prediction model is obtained. The target stent is determined from at least one candidate stent based on the prediction model of each candidate vessel.

7. The method according to claim 6, characterized in that, The method further includes: Obtain an actual vascular model after the target stent is implanted into a blood vessel; The prediction model is optimized based on the actual blood vessel model and the candidate blood vessel prediction model corresponding to the target stent.

8. An apparatus for training a predictive model for predicting vascular deformation after stent implantation, characterized in that, The device includes: The brain imaging training data pair acquisition module is used to acquire brain imaging training data pairs; the brain imaging training data pairs include preoperative brain imaging data before stent implantation and postoperative brain imaging data after stent implantation. The three-dimensional reconstruction module is used to perform three-dimensional reconstruction processing on the preoperative cranial imaging data and the postoperative cranial imaging data to obtain the preoperative vascular model and the postoperative vascular model. The stent model generation module is used to map the pose information of the stent in the postoperative cranial imaging data to the preoperative vascular model to generate a stent model. The graph structure construction module is used to construct a target vascular graph structure and a target stent graph structure based on the preoperative vascular model and the stent model; wherein, the first node on the target vascular graph structure is a sampling point on the vascular wall in the preoperative vascular model, and the first node carries first physical coding information; the second node on the target stent graph structure is a sampling point on the stent in the stent model, and the second node carries second physical coding information. The training module is used to input the target vascular map structure, the target stent map structure, and the postoperative vascular model into the pre-constructed graph neural network model for training until the set training termination condition is met, thereby obtaining the prediction model.

9. An electronic device, characterized in that, include: A processor, configured to execute program instructions; as well as A memory configured to store the program instructions, which, when loaded and executed by the processor, cause the processor to perform the steps of the method for training a predictive model for predicting vascular deformation after stent implantation as described in any one of claims 1-7.

10. A computer-readable storage medium storing program instructions, characterized in that, When the program instructions are loaded and executed by the processor, the processor performs the steps of the method for training a predictive model for predicting vascular deformation after stent implantation as described in any one of claims 1-7.

Images