A cerebral aneurysm stent deployment prediction method, device, and medium

Abstract

This invention discloses a method for predicting stent deployment in cerebral aneurysms, comprising: acquiring vascular images containing cerebral aneurysms and extracting multiple vascular structures related to the aneurysm; encoding each vascular structure as a structural latent variable using a structural encoder; establishing edges based on the anatomical connections of each vascular structure using its structural latent variables as nodes to generate a structural interaction graph, and aggregating information from the structural interaction graph using a graph neural network to obtain global latent variables representing global topological relationships; extracting the geometric features of the aneurysm neck and fusing the aneurysm neck geometric features with the global latent variables; and processing the fused features using a predictive decoder to predict the location point sequence of the aneurysm stent deployment path. This embodiment improves the stability and accuracy of the prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image analysis technology, and in particular to a method, device and medium for predicting the deployment of stents for cerebral aneurysms. Background Technology

[0002] Cerebral aneurysm is a high-risk cerebrovascular disease, and FDS (Flow-Diverting Stent) is the mainstream endovascular interventional treatment tool for it. The success or failure of FDS (also known as stent) treatment is closely related to its deployment location and morphology within the blood vessel; a precise deployment plan is the key to surgical success.

[0003] Currently, preoperative planning for FDS (Follicular Unit Surgery) relies primarily on the physician's personal experience, involving manual planning based on the observation of 3D images. This approach is not only highly subjective and inefficient, but also struggles to accurately assess complex vascular geometry, making it difficult to guarantee the consistency and optimality of the planning results.

[0004] To address this issue, computer-aided techniques, particularly deep learning methods, have begun to be used to predict stent deployment. However, existing deep learning models share a common drawback: they typically analyze the parent artery in isolation, resulting in severely inadequate accuracy and reliability in cases involving complex bifurcation regions. Summary of the Invention

[0005] This invention provides a method for predicting the deployment of stents for cerebral aneurysms to solve the above-mentioned technical problems.

[0006] In a first aspect, embodiments of the present invention provide a method for predicting the deployment of a cerebral aneurysm stent, comprising:

[0007] The vascular images containing cerebral aneurysms were acquired, and multiple vascular structures related to the aneurysms were extracted from them. The multiple vascular structures included the aneurysm, the carrier artery segment, and adjacent major bifurcation segments that affect the anchoring of the aneurysm stent.

[0008] Each vascular structure is encoded as a structural latent variable using a structural encoder.

[0009] Using the structural latent variables of each vascular structure as nodes, edges are established according to the anatomical connection relationship of each vascular structure to generate a structural interaction graph. Then, a graph neural network is used to aggregate information from the structural interaction graph to obtain global latent variables used to characterize global topological relationships.

[0010] Extract the geometric features of the aneurysm neck and fuse the geometric features of the aneurysm neck with the global latent variables;

[0011] The fused features are processed using a predictive decoder to predict the location sequence of the aneurysm stent deployment path.

[0012] In a second aspect, embodiments of the present invention provide an electronic device, the electronic device comprising:

[0013] One or more processors;

[0014] Memory, used to store one or more programs;

[0015] When the one or more programs are executed by the one or more processors, the one or more processors implement the cerebral aneurysm stent deployment prediction method of any embodiment.

[0016] Thirdly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the cerebral aneurysm stent deployment prediction method described in any embodiment.

[0017] In summary, this invention addresses the technical problems of existing technologies, such as reliance on human experience, high subjectivity, low efficiency, and insufficient utilization of complex three-dimensional vascular structural information, in order to provide a method for automatically and accurately predicting the three-dimensional deployment position of stents for cerebral aneurysms. This method constructs a unified three-dimensional point cloud integrating the aneurysm, the carrier artery segment, and adjacent major bifurcation segments that influence stent anchorage, achieving a complete geometric representation of the complex vascular anatomy and overcoming the information loss caused by structural simplification in traditional models. Simultaneously, the use of a variational autoencoder for nonlinear dimensionality reduction extracts lower-dimensional latent variables with more structural semantics from the high-dimensional point cloud, significantly improving the fitting and expression capabilities for complex vascular morphological variations. More importantly, this method creatively utilizes a graph neural network to fuse the features and topological relationships of different vascular structures into global structural variables, and integrates these with the aneurysm neck geometric features representing key local influencing factors, providing dual-channel input features. This allows the prediction model to simultaneously perceive the overall structure and local details, thereby greatly improving the accuracy and stability of predicting key deployment states such as stent adhesion and slippage trends. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1A flowchart illustrating a method for predicting the deployment of a cerebral aneurysm stent, provided in an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram of a statistical shape modeling process provided in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of an enhanced modeling flowchart that integrates tumor cyst and tumor neck features, provided as an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram illustrating the construction and application process of a stent deployment prediction model provided in an embodiment of the present invention.

[0023] Figure 5 This is a structural diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0025] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0026] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0027] As described in the background section, existing deep learning models typically analyze the parent artery of the aneurysm in isolation. However, in human anatomy, the parent artery, the aneurysm, and adjacent branch vessels constitute a complex, interconnected system. The presence of branch vessels directly affects the final morphology and apposition of the stent. This embodiment provides a method for predicting stent deployment in cerebral aneurysms by incorporating this "structural context" information into a deep learning model. This model learns the complex spatial geometric relationships between these components to achieve more accurate and reliable stent deployment prediction.

[0028] Figure 1 This is a flowchart illustrating a method for predicting the deployment of a cerebral aneurysm stent, as provided in an embodiment of the present invention. The method is executed by an electronic device. Figure 1 As shown, the method specifically includes:

[0029] S110. Acquire vascular images containing cerebral aneurysms and extract multiple vascular structures related to the aneurysm from them, wherein the multiple vascular structures include the aneurysm, the carrier artery segment, and adjacent major bifurcation segments that affect the anchoring of the aneurysm stent.

[0030] This step involves data acquisition and 3D model extraction. It obtains raw CTA (Computed Tomography Angiography) image data (DICOM format) of patients containing cerebral aneurysms, and extracts 3D models of multiple aneurysm-related vascular structures from the 2D image sequence for subsequent geometric analysis. Specifically, for ease of description, this embodiment treats the aneurysm itself as an aneurysm-related vascular structure, and refers to each extracted structure (including the aneurysm) collectively as a vascular structure.

[0031] In one specific embodiment, the process may include the following steps:

[0032] Step 1: Obtain CTA images of the vessels containing the cerebral aneurysm. Combined with... Figure 1 and Figure 2 The original input data used in this embodiment is CTA images of the patient's cerebral aneurysm region. CTA has advantages such as fast acquisition speed, high resolution, and strong ability to separate bone and vascular soft tissues, and can clearly show the morphological relationship between intracranial aneurysms and surrounding vascular structures.

[0033] In the data acquisition phase, this embodiment receives raw CTA image sequences in DICOM format, covering complete three-dimensional volumetric data. To ensure the accuracy of structural identification and the consistency of feature representation, the acquisition process must cover the entire vascular segment, including the parent artery, aneurysm body, aneurysm neck region, and a certain surrounding area. Data sources may include publicly available image databases, real clinical data from hospitals, or datasets constructed from previous research. Ethical approval or data authorization is required, and data anonymization must comply with privacy protection regulations.

[0034] The acquired image data will serve as the foundation for subsequent structural reconstruction, morphological encoding, and stent deployment prediction model training. It requires clear image quality, a high signal-to-noise ratio, and a complete image sequence without missing frames. Optionally, to adapt to subsequent point cloud encoding and feature fusion processes, the image acquisition should meet the accuracy requirements of structural reconstruction and geometric calculation in terms of spatial resolution and slice thickness settings. For example, a slice thickness of 0.5–1.0 mm should be set, ensuring sufficient scanning coverage in the Z-axis direction to facilitate the subsequent construction of vascular centerline and cross-sectional structural information.

[0035] Furthermore, to ensure that the acquired CTA images fully meet the accuracy requirements for vascular structure reconstruction and geometric inference required by this invention, specific parameter settings and scanning range specifications can be followed during the acquisition process. Regarding scanning parameters, a tube voltage of 120kVp and a tube current of 300–400mA can be used, with the slice thickness controlled at 0.5–0.625mm. Thin-slice reconstruction with consistent reconstruction spacing should be enabled to achieve detailed reconstruction of the tumor and its surrounding complex vascular structures. The scanning speed can be controlled at 0.5–1 second / cycle to reduce interference from motion artifacts. Contrast agent can be injected simultaneously using a dual-barrel syringe at a flow rate of 4.0–5.0mL / s. The total dose is set according to the patient's weight, generally between 60–100mL, and supplemented with normal saline to enhance vascular enhancement.

[0036] In terms of scanning range, it can cover the entire length of the anterior, middle, and posterior cerebral arteries, from the internal carotid artery segment to the main trunks. Emphasis should be placed on ensuring at least two vascular bifurcation points upstream and downstream of the aneurysm to completely capture the blood flow pathway and potential stent anchorage area. Furthermore, to facilitate subsequent 3D modeling of the stent deployment path, the tortuous and bifurcation segments of the artery containing the aneurysm can be included to ensure geometrical continuity in centerline extraction and cross-sectional reconstruction. If the aneurysm is located in the proximal internal carotid artery or posterior communicating artery, or other small vessel branches, the acquisition window can be adjusted and the dose increased to improve the local signal-to-noise ratio and angiography enhancement, avoiding errors or omissions in the identification of important structures.

[0037] Step 2: Perform region segmentation on the CTA image to obtain multiple vascular structures related to the aneurysm. After acquiring the original CTA image, this embodiment first performs structured preprocessing on the original CTA image. The first step is to automatically segment the structural regions of the CTA image. Specifically, a region of interest segmentation method can be used, employing semi-automatic or fully automatic image segmentation algorithms (such as thresholding, region growing, or deep learning segmentation networks like U-Net) to accurately segment the vascular portion of interest from the CTA image. The segmented objects include not only the aneurysm itself, but also the parent artery directly connected to it, as well as adjacent major bifurcation vessels that have a significant impact on stent anchoring. In this embodiment, at least three structures are extracted: the aneurysm, the parent artery segment, and a major bifurcation vessel segment.

[0038] Optionally, considering the complex morphology of cerebral aneurysms and the relatively small density difference between blood vessels and bone or other soft tissues, a deep learning model based on a 3D U-Net structure can be used to perform voxel-level anatomical segmentation of CTA images and output vascular structure masks to improve segmentation accuracy. This model has been pre-trained on a publicly available vascular segmentation dataset and fine-tuned on local samples to adapt to the specific morphology of cerebral aneurysms. Post-processing of the segmentation results removes small-area error segments, retains the main vascular structures and aneurysm regions, and forms the initial volume of the anatomical structure. In cases of insufficient training samples or extreme image quality, a classic rule-based workflow based on multi-scale Hessian filtering and region growing can be switched. Both schemes output three types of voxel masks, and the connectivity consistency between the aneurysm and vascular segments is ensured during topological post-processing.

[0039] Step 3: Perform 3D surface reconstruction on the segmentation results to obtain 3D surface models of each vascular structure. This step constructs the corresponding 3D structural surface model based on the segmentation mask. Specifically, the MarchingCubes algorithm is used to convert the binary volume into a polygonal mesh model composed of triangular facets to achieve continuous reconstruction of the vascular surface. The output of this step is a 3D surface model containing the complete morphology of the tumor and the tumor-bearing vessel, laying the foundation for subsequent point cloud sampling and structural standardization. Optionally, to eliminate noise and staircase effects that may occur during reconstruction, post-processing such as Laplacian smoothing can be performed on the model to obtain a smooth 3D surface model, such as... Figure 2 As shown.

[0040] S120. Using a structural encoder, each vascular structure is encoded as a structural latent variable.

[0041] This step involves point cloud processing and structural latent variable generation. Point cloud sampling and standardization are performed on the 3D model of each vascular structure, including aneurysms. For the standardized point cloud of each vascular structure, the corresponding structural latent variables are generated through a structural encoder network.

[0042] In one specific embodiment, the process may include the following steps:

[0043] Step 1: Point cloud sampling is performed on the 3D surface models of each vascular structure to obtain the 3D point cloud of each vascular structure. To input the data into a deep learning network capable of processing point clouds, this step samples the point cloud of each 3D surface model, converting the surface model into a structural point cloud. Optionally, to achieve standardized modeling and cross-individual morphological alignment of cerebral aneurysm vascular structures, this embodiment provides a unified format 3D point cloud construction method for structural fusion modeling needs. Based on the statistical model of the vascular structure output in the previous stage, multi-source geometric data such as the vascular centerline, cross-sectional contour, aneurysm mesh, and aneurysm neck boundary are fused. Through a systematic sampling, normalization, and coordinate alignment process, a 3D point cloud representation with a unified topological structure and geometric scale is constructed, providing stable and high-quality input data for subsequent variational coding and graph neural network processing modules.

[0044] Specifically, firstly, based on the structure output by the Statistical Shape Model (SSM), geometric representations including the triangular mesh of the aneurysm sac surface and the aneurysm neck boundary curve are extracted. These original mesh data have diverse structures and uneven node densities, making them unsuitable for direct neural network modeling; therefore, geometric resampling and standardization are necessary. To ensure uniform distribution of sampling points and preserve surface geometric features, Faster Pythagorean Theorem (FPS) is used as the point set construction strategy. Let the input be the aneurysm surface mesh. ,in Represents a grid node. This represents the number of grid nodes; a dense and uniform point set is constructed by iteratively selecting the point farthest from the current point set. ,in This represents the sampling points in the point set, where N is a fixed number of sampling points (e.g., 1024 or 2048) to ensure the consistency of the point cloud structure among different cases.

[0045] Then, to further eliminate the influence of differences in head posture, imaging scale, and aneurysm location among different patients, the point cloud was normalized and aligned. In this process, the center of the point set was first translated to bring the geometric center of the entire point cloud back to the origin, i.e.:

[0046] in, This indicates the sampling point after translation.

[0047] Secondly, to unify scale information, a maximum bounding sphere normalization strategy is adopted to scale the point cloud as a whole to the interior of a unit sphere, giving it a fixed boundary scale. The formula is as follows:

[0048]

[0049] in, This represents the scaled sampling points.

[0050] After the above two steps, the position and size of the point cloud in space are standardized. However, to further align its orientation with the spatial structure and ensure geometric consistency in subsequent latent variable encoding, attitude standardization along the principal axis can be performed. Specifically, PCA (Principal Component Analysis) is performed on the normalized point cloud to obtain its first principal axis. And based on the normal vector of the neck plane Perform a rotation transformation to align the two elements to a preset spatial orientation. The rotation matrix R satisfies:

[0051] Where R is the first spindle Rotate to Rotation matrix; It is used to determine the reference direction of the rotation matrix R. In the derivation, it is used as a reference vector, and its direction information is implicit in the calculation of the rotation matrix R.

[0052] The point cloud with a uniform format is finally obtained by rotation. , where each point This standardized point cloud meets unified standards in terms of spatial scale, orientation, and topological density, and has good comparability and modeling compatibility.

[0053] Furthermore, geometric preservation of the tumor neck boundary points can be considered during point cloud construction. Specifically, to avoid insufficient coverage of key anatomical structures by point cloud sampling, a region-weighted sampling strategy can be adopted. A sampling weight enhancement factor can be set in the tumor neck boundary region to increase the sampling density of points adjacent to the tumor neck, which helps the downstream network to better capture the geometric transition features between the tumor body and the tumor neck.

[0054] It is important to emphasize that the unified format point cloud is not only a standardized form of data input, but also a fundamental support for realizing cross-case structural latent variable modeling. In the subsequent multi-structure fusion modeling framework, the unified format point cloud, as the geometric input of the tumor, provides rich three-dimensional spatial information and is composable with the neck structural parameters. This provides the conditions for the subsequent structural encoder to generate a stable structural latent variable vector z, further supporting the performance improvement and generalization capability expansion of downstream stent deployment prediction tasks.

[0055] Step 2: Using a structural encoder, the 3D point clouds of each vascular structure are encoded into structural latent variables for that vascular structure. After standardizing the structural point clouds, this step generates corresponding structural latent variables for each standardized point cloud using a structural encoder network. This enables efficient representation and cross-sample modeling of the complex anatomical structures of cerebral aneurysms. Each structural latent variable is a low-dimensional vector designed to compress and represent the macroscopic shape and morphological information of the vascular structure.

[0056] Optionally, the structural encoder can be the encoder portion of a VAE (Variational Autoencoder). This encoder can utilize the PointNet++ framework in its network architecture, effectively learning local and global geometric features of point cloud data through multi-level ensemble abstraction, thereby generating high-quality structural latent variables. The PointNet++ framework can include two levels of SetAbstraction layers and one level of Feature Propagation layer, embedding local graph convolutions (DGCNN EdgeConv) in each layer to enhance neighborhood morphological awareness; each MLP (Multi-Layer Perceptron) sub-block can use ReLU activation and Batch Normalization, with residual connections added to the top layer to mitigate deep gradient vanishing.

[0057] The encoder part consists of multiple MLPs and global aggregation operations (such as max-pooling or attention-based pooling). Its function is to progressively extract point-level features and synthesize them into a global geometric description vector. Specifically, the input data of the encoder part is the standardized point cloud set generated in the previous step, denoted as . , where each point This represents its coordinates in three-dimensional space. In this process, the local features of each point are mapped through a nonlinear mapping function f. θ Convert to high-dimensional feature vectors The formula is expressed as follows:

[0058]

[0059] Subsequently, a global representation of the entire point cloud is obtained through pooling operations. :

[0060]

[0061] In the VAE architecture, this global representation will be fed into two parallel fully connected networks to generate the mean vectors of the latent variable distributions. With the log-variance vector logσ 2 ,Right now:

[0062]

[0063] in, These represent the parameters of a fully connected network.

[0064] The encoder does not directly output the latent variable z, but instead generates it through sampling using the reparameterization trick, thus ensuring the network is differentiable during training.

[0065]

[0066] This latent variable The low-dimensional semantic embedding representing the point cloud structure contains both the morphological information of the tumor structure and the spatial distribution features of the tumor neck, becoming the key input for the subsequent decoder and structure fusion prediction module.

[0067] In addition to the encoder part mentioned above, the VAE also includes a decoder part. During the training phase, both parts are jointly trained using reconstruction loss (such as Chamfer Distance) and KL divergence loss. This ensures that the latent variables generated by the encoder part can be effectively used to reconstruct the original point cloud and also follow a standard normal distribution, thus forming a structured and meaningful feature space. The decoder part is removed after training and does not participate in inference and prediction during the model's usage phase.

[0068] In one specific implementation, to ensure the interpretability and generative capability of the encoded latent variables, during training, the encoder encodes the 3D point clouds of each vascular structure into structural latent variables, and the decoder reconstructs each vascular structure based on these structural latent variables. The encoder and decoder are jointly trained using the reconstruction error between the reconstructed and original structures, as well as the KL divergence of each structural latent variable. Optionally, the loss function term of the above VAE architecture... Expressed as:

[0069]

[0070] in, Indicates reconstruction error, To reconstruct the point cloud, This represents the points in the reconstructed point cloud. Used to reflect whether structural information is effectively preserved; Indicates the Kullback–Leibler divergence. d represents the dimension of the structural latent variables. Used to approximate the latent variable distribution of the encoder output to a standard normal distribution. This enhances the model's generative and generalization capabilities; the constant β is a balancing term used to adjust the weights of the two loss components. Initially, β can be set to 0.5 to emphasize reconstruction, and then increased to 4.0 using cosine annealing to obtain a more discriminative distribution of latent variables.

[0071] The trained encoder portion serves as a structural encoder, capable of automatically extracting low-dimensional latent representations with discriminativeness, reconstructability, and structural semantic consistency from a uniformly formatted aneurysm point cloud structure. These representations are then used for structural condition modeling and policy generalization learning in subsequent stent deployment prediction tasks. This latent variable representation not only preserves the geometric and topological information of the original point cloud but also effectively compresses the data dimensionality, reduces the number of model parameters, and improves training efficiency, making it one of the core components for achieving structural fusion modeling.

[0072] Optionally, to further enhance the model's ability to represent the neck and body regions of the tumor, a region label mask can be introduced into the original point cloud input. This involves adding dimensional labels to the point set of the neck region, guiding the network to learn important geometric patterns of the region during training. This structural prior avoids the problem of insufficient discrimination ability in key regions inherent in purely data-driven models, enhancing the accuracy of subsequent location prediction tasks. Specifically, in the structural latent variable encoding of this step, region label embedding and dimensional labels can be used as additional input channel features, fed together into the encoder network to guide the model in distinguishing different spatial anatomical regions during feature extraction. The region label embedding is embedded into the feature dimensions of the point cloud, influencing the feature weight distribution within the local receptive field through feature concatenation or positional encoding, enabling the network to identify the boundary differences between the neck and body of the tumor when learning point cloud geometry. The dimensional labels are used to assist attention allocation in the global aggregation stage, guiding the model to preserve the topological hierarchy between regions when generating latent variables. By introducing these two labeling mechanisms, explicit guidance on the geometric patterns of key regions is achieved during network training, thereby improving the expressive power of spatial geometric features at the aneurysm neck and the accuracy of stent deployment location prediction.

[0073] S130. Using the structural latent variables of each vascular structure as nodes, establish edges according to the anatomical connection relationship of each vascular structure to generate a structural interaction graph, and use a graph neural network to aggregate information from the structural interaction graph to obtain global latent variables used to characterize global topological relationships.

[0074] This step, based on a GNN (Graph Neural Network), performs global information fusion on the structural latent variables extracted in S120. This step is one of the core innovations of this invention. It constructs a structural interaction graph with each structural latent variable as a node based on the anatomical connectivity of each vascular structure, and uses a graph neural network to aggregate information from the structural interaction graph to obtain a globally fused latent variable representing the global topological relationship.

[0075] In one specific embodiment, the process may include the following steps:

[0076] Step 1: Construct a structural interaction graph. Optionally, the structural latent variables of each vascular structure are used as nodes, and edges are established between vascular structure nodes with anatomical connections. Weights are assigned to each edge based on the distance between the vascular structures connected by each edge. If the point cloud overlap rate of the vascular structures connected by each edge exceeds a set threshold, the weights of each edge are increased.

[0077] For example, based on the anatomical physical connections of various vascular structures, a structural interaction graph G=(V,E) is constructed. The node set V of the graph consists of the latent variables of each structure generated in the previous step, for example, V={z_aneurysm,z_parent,z_branch}, where z_aneurysm, z_parent, and z_branch represent the structural latent variables of the aneurysm, the parent artery segment, and the bifurcation segment, respectively. The edge set E of the graph represents the connections between structures. If two vascular structures are physically connected (e.g., an aneurysm grows on the parent artery, and the parent artery is connected to the bifurcation segment), then an edge is established between their corresponding latent variable nodes.

[0078] Furthermore, weighted edges can be established based on anatomical connections or overlapping point cloud regions. If the distance between the closest points of two blood vessels in three-dimensional space is less than 1 mm and they are topologically connected along the CTA reconstruction centerline, then edges are established between the nodes and weighted accordingly.

[0079]

[0080] in, This represents the weight of the edge between the i-th and j-th vascular structures. , Let be the geometric centroid coordinates of the i-th and j-th vascular structures, respectively, and σ be an empirical constant (e.g., 2 mm). If the point clouds of two structures overlap by more than 15 mm², an additional weight of 0.2 is applied. This edge weight design balances geometric proximity and anatomical connectivity, ensuring that graph convolution propagates effective information during training.

[0081] Step 2: Utilize a graph neural network (GNN) to aggregate information from the structural interaction graph, obtaining a global latent variable representing the global topological relationships. Optionally, the GNN can employ a graph convolutional network (GCN) or a graph attention network (GAT). By performing information transfer and aggregation between neighboring nodes on the graph, the representation of each node incorporates information from its neighbors. For example, the node representation of the aneurysm-bearing artery, after GNN processing, will include morphological information from nearby bifurcation vessels. After multi-layer GNN computation, a global fusion latent variable z, incorporating all relevant structural information, is finally obtained. fused This latent variable captures the topological relationships and systemic characteristics of the entire local vascular network.

[0082] This step integrates deep representations of multiple vascular structures (such as tumor-bearing arteries, bifurcation vessels, and distal pathways) to achieve cross-structural semantic interaction in a unified geometric space. This addresses the problems of large stent deployment errors and poor adaptability caused by traditional single anatomical structure modeling. The update expression of the structural latent vector of each node in the GNN is as follows:

[0083]

[0084] in, Represents a node The set of adjacent nodes, These are the graph attention weight coefficients. For the first The weight matrix of the layer, For node i to pass through the first The updated feature vector after layer fusion Similarly, σ is a non-linear activation function. After multi-layer fusion, the resulting z... fused This refers to a high-order abstract feature that integrates all key structural representations.

[0085] Furthermore, to maintain the topological coherence of the point cloud after resampling, an adjacency matrix can be constructed after the point cloud is generated. If two points share an edge in the original mesh or the angle difference between them and the centerline on the same tangent plane is less than 5° (a geometric adjacency condition used to determine whether two points are on the same cross-section and have approximately the same direction, used to maintain the spatial continuity of the support point cloud), they are considered adjacent and set as adjacent. Additionally, structural labels can be attached to each sampling point. These correspond to the tumor sac, tumor neck, and mother blood vessel regions, respectively. This adjacency-label pair, during the encoding stage of the graph neural network, can recover the local topology and cross-regional geometric transition relationships, avoiding the loss of topological information caused by pure coordinate representation. Specifically, when computing message passing between nodes, the GNN can use the label l iAs an additional input or mask term, this allows the model to assign higher attention weights to the aneurysm neck region when updating node representations, thereby enhancing the learning of geometric features in key regions. This ensures that the model fully maintains the topological continuity and structural semantic distinction between the aneurysm neck and the parent vessel during the latent variable fusion stage.

[0086] In one specific implementation, to maintain the consistency of representation of anatomically connected or geometrically proximate structures in the latent space, during training, a graph neural network generates latent vectors for each vascular structure in the structural interaction graph, and uses the weights of each edge in the graph to perform a weighted average of the differences in the latent vectors of the vascular structures connected by each edge; the graph neural network is trained by minimizing the weighted average result. Optionally, the loss function term of the graph neural network... This can be expressed as:

[0087]

[0088] in, For the node's hidden vector, These are the edge weight coefficients. This represents the number of elements in set E. Also known as the consistency error in graph feature propagation, it encourages nodes with similar structures to generate similar feature vectors.

[0089] S140. Extract the geometric features of the aneurysm neck and fuse the geometric features of the aneurysm neck with the global latent variables.

[0090] The aneurysm neck is a critical area that the stent must precisely cover. Considering the importance of the aneurysm's morphological specificity for deployment prediction, this embodiment uses a geometric calculation algorithm to automatically or semi-automatically identify the aneurysm neck plane on a 3D model of the aneurysm and extract a set of key aneurysm neck geometric features. Optionally, the aneurysm neck geometric features include the opening angle, aneurysm neck area, and edge curvature. Let the normal vector of the aneurysm neck plane be n, then the aneurysm neck opening angle can be defined as the angle between the principal axis of the aneurysm and n, the area is estimated using the convex hull of a point set, and the curvature is derived based on a second-order difference approximation. These features together constitute the aneurysm neck geometric feature vector. It is integrated with the global latent variables constructed by S130 to provide high-quality, low-bias structural prior support for the scaffold deployment model.

[0091] In one specific implementation, the surface mesh of the aneurysm cyst is first extracted from the three-dimensional structural model of S110; then, using a seed point growth algorithm, triangular segments on the surface of the aneurysm are traced from the bifurcation point of the centerline to generate a mesh set M that encloses the aneurysm region. aneurysm Further extract the neck boundary point set (M1 is the number of boundary points), used to calculate the geometric constraints of the neck.

[0092] Then, based on the set of boundary points of the tumor neck, the geometric features of the tumor neck are calculated, including at least one of the following: tumor neck opening angle, tumor neck edge curvature, tumor neck area, center offset angle, axial offset amplitude of the tumor body, coordinates of the center point of the tumor neck plane, normal vector of the tumor neck plane, maximum diameter of the tumor neck, and minimum diameter of the tumor neck. Among these, the tumor neck opening angle... It is a core parameter reflecting the angular relationship between the aneurysm and the mother vessel, and is constructed by constructing the normal vector of the aneurysm neck plane. Vector of the principal axis of the mother blood vessel The angle between them is defined, and its calculation formula is:

[0093]

[0094] This angle reflects the degree of alignment between the stent's unfolding surface and the geometry of the aneurysm neck. Generally, the larger the angle, the more difficult it is to adhere to the aneurysm, and the more problems will occur after deployment. The greater the risk in a region.

[0095] Secondly, the curvature of the tumor neck margin From the set of points at the neck of the tumor The average curvature of the fitted two-dimensional curve is estimated by calculating the local curvature at each point using the circle fitting method and then representing it as the global average.

[0096]

[0097] Neck area This value, obtained by triangulation and summation of the closed curve at the aneurysm neck boundary, reflects the absolute magnitude of the area to be covered by stent deployment. This value is closely related to indicators such as grid density and stress state after stent deployment, and is crucial for the model to determine whether the coverage is sufficient.

[0098] When considering the impact of tumor spatial offset on deployment, this embodiment also defines the angle between the tumor center vector and the parent blood vessel axis as... That is, the vector formed in space by the center of gravity of the tumor relative to the center of gravity of the tumor neck. The angle between the vector and the direction vector of the principal axis of the mother blood vessel is expressed as:

[0099]

[0100] If the tumor location deviates significantly from the plane of the tumor neck, the stent may deploy asymmetrically along the misalignment direction. Furthermore, axial offset... This represents the absolute projected distance of the tumor's center of gravity along the direction of the parent vessel, calculated using vector projection difference. The calculation of other geometric features is relatively simple and will not be elaborated further.

[0101] The above geometric features together constitute the neck feature vector. This will serve as important supplementary information to the structural encoding and participate in the model input. Therefore, the global fusion latent variables obtained in the previous step will be used... The geometric feature vector of the tumor neck obtained in this step The simplest fusion method is vector concatenation, followed by non-linear mapping through one or more fully connected layers (MLP) to obtain the final fused features. This feature It also contains global vascular topology information and precise geometric information of the local aneurysm neck.

[0102] S150. The fused features are processed using a predictive decoder to predict the location sequence of the aneurysm stent deployment path.

[0103] This step predicts the deployment location of the support structure. The final fused features obtained in S150 are input into another decoder, which outputs the predicted sequence of coordinate points for the support structure deployment location. For ease of distinction and description, this embodiment refers to this other decoder as the prediction decoder.

[0104] Optionally, the predictive decoder uses an MLP architecture, with the input being... The output is a list of T three-dimensional coordinate points. The resulting sequence of coordinate points defines the centerline path of the stent. Each coordinate point serves as a control point in the stent deployment trajectory, enabling the mapping of a well-fitting path curve onto the actual vascular space.

[0105] The stent deployment prediction method in this embodiment utilizes a fusion of structural latent variables and aneurysm neck geometric features to model the final deployment location of the stent within the cerebral aneurysm vascular structure with high precision. This deployment location can essentially be expressed as a spatially continuous sequence of three-dimensional coordinate points, matching the segments of the mainstream mesh stent. For example, the number of nodes T is set to 20 (6×3+2 endpoints), or adaptively increased at 3mm intervals when the vessel centerline length exceeds 60mm, ensuring consistent output scale across different cases. Its geometry must conform to the local vascular lumen, reflecting the stent's adaptability to the vessel wall during clinical use.

[0106] In one specific implementation, the reasoning for the stent deployment location is based on fusion features. As input, the positions of several consecutive spatial nodes are mapped by the predictive decoder. The number of nodes T defaults to 20, and when the centerline length exceeds 60mm, it adaptively expands at 3mm intervals to ensure consistent output scale across cases. To ensure the generated sequence has vascular adaptability and geometric feasibility, the predictive decoder is constrained during the training phase to output a node sequence that is basically consistent with the vascular distribution morphology and has reasonable local curvature. This process can be formally represented as:

[0107]

[0108] Among them, node combination each This represents the three-dimensional position of the i-th node of the support frame in the reconstructed coordinate space. These are the network weight parameters. This indicates the specific operation corresponding to the predictive decoder.

[0109] Furthermore, to enhance the model's individual adaptability, the predictive decoder can also utilize conditional graph structural information to automatically adjust the node layout to adapt to complex arterial tortuous regions or morphologically abrupt changes near aneurysm segments. This adaptive adjustment is achieved through an attention mechanism, enabling the network to dynamically focus on the activation of latent features in key structural regions.

[0110] Ultimately, the coordinates output by the network can be physical spatial values ​​in millimeters, which can be obtained from data recorded during the preprocessing stage. The inverse transformation automatically maps back to DICOM voxel coordinates (row, column, slice), maintaining a one-to-one correspondence with the original CTA sequence. The resulting node sequence can then be directly used as input for subsequent visualization, virtual deployment, or mechanical simulation modules, maintaining the same coordinate system as the original CTA image without requiring further coordinate transformation. This entire method achieves fully automated reasoning for stent deployment paths, significantly reducing the need for manual planning and providing a structural-level prior for preoperative clinical assessment and intraoperative path correction.

[0111] In one specific implementation, to constrain the accuracy and anatomical rationality of the inference results, the predictive decoder updates its network parameters during training by minimizing the difference between the predicted and actual location point sequences of the aneurysm stent deployment path, ensuring the stent deployment morphology most closely resembles the actual vessel morphology, maximizing path smoothness, and ensuring the stent ends adhere to the vessel wall in the anchoring area. Optionally, the loss function term of the above predictive decoder... This can be expressed as:

[0112]

[0113] in, and These are the dot product of the predicted path and the set of true path points, respectively. This represents the Euclidean distance error between the predicted point and the actual deployment point, used for regression accuracy constraints. This is a path smoothness regularization term to avoid discontinuities or sharp turns in the predicted path; This is used to measure the spatial overlap between the predicted path and the actual vessel morphology (using ChamferDistance or Hausdorff distance), ensuring the deployment result has anatomical fit. For example, ChamferDistance in the above formula ensures that the predicted overall path fits the vessel lumen centerline. Weights of each loss term. , Adjust the settings empirically or using an adaptive learning strategy based on the model's convergence performance, for example... , The experience was set to 1.0, 0.1, and 0.5, and cosine annealing was used for automatic fine-tuning 20 epochs after training.

[0114] Optionally, to ensure the spatial continuity and morphological rationality of the output path, Furthermore, a node spacing balance constraint can be introduced. For example, to limit drastic fluctuations in the distance between adjacent nodes, a uniformity loss function term can be introduced as follows. :

[0115]

[0116] in, The target average spacing.

[0117] At the same time, in order to fix the position of the first and last nodes and the nearest point on the center line of the mother blood vessel, The following loss function term can also be introduced. To ensure that the path begins and ends within the anchorable section, adhering to the wall:

[0118]

[0119] in, These are the coordinates of the first and last nodes of the predicted path, respectively. and This refers to the coordinates of the nearest points adjacent to the aneurysm at the proximal and distal ends of the vessel centerline. The anchor point field ensures that the start and end positions of the stent path are attached to the inner wall of the implantable segment, preventing inference results from drifting at the beginning and end. Specifically, "proximal" refers to the parent vessel segment upstream of the aneurysm neck plane, closer to the inflow end / heart direction, representing the starting point of the landing zone where the interventional device first reaches; "distal" refers to the parent vessel segment downstream of the aneurysm neck plane, farther from the inflow end / heart direction, representing the ending point of the landing zone after the stent crosses the aneurysm neck. "Adjacent" in "adjacent to the aneurysm" refers to the closest relationship between the parent vessel centerline and the aneurysm neck boundary: cprox is the nearest point between the proximal end of the centerline and the aneurysm neck boundary, and cdist is the nearest point between the distal end of the centerline and the aneurysm neck boundary. Put the first point Pull to Tail point Pull to Used to fix the start and end points of the predicted path close to the nearest point proximal / distal to the centerline of the maternal vessel. .

[0120] As described above, the cerebral aneurysm stent deployment method of this embodiment includes several parts in its network structure: VAE, GNN, and a prediction decoder. These components together achieve efficient encoding and accurate inference of various complex geometric relationships of the cerebral aneurysm vascular structure during stent deployment. Furthermore, the loss function terms for each part during training are described in the above embodiments. Based on these loss function terms, this embodiment provides a VAE-GNN joint training mechanism, which achieves collaborative optimization of structural latent variable learning and stent deployment path prediction through an end-to-end training process. Specifically, latent variable encoding, graph structure information fusion, and path inference decoding are executed simultaneously within each batch, and parameters are updated jointly through a backpropagation mechanism. The joint loss function is as follows:

[0121]

[0122] Weighting coefficients of each loss term The settings are dynamically adjusted according to the training phase.

[0123] Optionally, to improve convergence stability, the training can be divided into two phases:

[0124] Phase 1: Freeze the GNN and predictive decoder, training only the encoder and decoder parts of the VAE. First, the encoder part encodes the unified structured point cloud data X, extracts spatial geometric features through a convolutional neural network, and then maps them to the structural latent variable space to obtain the latent variable distribution parameters. Corresponding latent variables are generated. The decoder then reconstructs the point cloud based on the latent variables, and then... Update the parameters for the encoder and decoder sections.

[0125] The second stage involves unlocking the GNN and predictive decoder. New samples are input into the encoder of the VAE to obtain the latent variables z of each vascular structure, which are then passed to the GNN. Each structural subgraph node input into the GNN contains a corresponding latent variable vector and anatomical attribute labels. Information is transferred in the structural graph through graph convolution operations, achieving semantic alignment and geometric relationship modeling between multiple structures. The graph structure is composed of an edge weight matrix. With node feature matrix The structure is defined by M2, where M2 represents the number of structures and d is the dimension of the latent variables. The update formula for graph convolution is as follows:

[0126]

[0127] in, I represents the identity matrix. For the first Layer weight parameters, This is a non-linear activation function. After multiple layers of graph convolution operations, the feature representations of each structure are fused into a new latent semantic vector. , used to represent the fused local and global information. The input is fed into the prediction decoder to obtain the predicted scaffold deployment path.

[0128] This phase adopts an end-to-end approach, simultaneously executing three steps within each batch: latent variable encoding, graph structure information fusion, and path reasoning decoding. The parameters are updated jointly through the backpropagation mechanism.

[0129] For example, first freeze the GNN and prediction decoder, and train only the VAE subnet for 20 epochs to obtain low-error reconstruction; then unlock the entire network and introduce... and The process continues for 50 epochs using an alternating strategy of "5-batch VAE update → 5-batch GNN + prediction decoder update". This scheduling improves the overall loss convergence speed by approximately 30% and reduces the average Euclidean error of the deployment path by approximately 18% on the internal validation set. The optimizer is Adam, configured with a warm-up strategy and a periodically decaying learning rate to improve model convergence stability.

[0130] After training, the network parameters for each part are fixed. During the inference phase, only preprocessed, unified structural point cloud data needs to be input. The encoder generates structural latent variables, which are then fused using a fixed-weight GNN module. Finally, the decoder generates the stent deployment path prediction result. The inference path is represented as a three-dimensional coordinate sequence, supporting interface with real vascular meshes. The entire inference process exhibits high stability and strong generalization ability, adapting to structural variations in different patients while maintaining the continuity and interpretability of the prediction results.

[0131] In summary, this embodiment addresses the technical problems of existing technologies, such as reliance on human experience, high subjectivity, low efficiency, and insufficient utilization of complex three-dimensional vascular structural information, in order to provide a method for automatically and accurately predicting the three-dimensional deployment position of stents for cerebral aneurysms. This method constructs a unified three-dimensional point cloud integrating the aneurysm, the carrier artery segment, and adjacent major bifurcation segments that influence stent anchorage, achieving a complete geometric representation of the complex vascular anatomy and overcoming the information loss caused by structural simplification in traditional models. Simultaneously, the use of a variational autoencoder for nonlinear dimensionality reduction extracts lower-dimensional latent variables with more structural semantics from the high-dimensional point cloud, significantly improving the fitting and expression capabilities for complex vascular morphological variations. More importantly, this method creatively utilizes a graph neural network to fuse the features and topological relationships of different vascular structures into global structural variables, and integrates these with the aneurysm neck geometric features representing key local influencing factors, providing dual-channel input features. This allows the prediction model to simultaneously perceive the overall structure and local details, thereby greatly improving the accuracy and stability of predicting key deployment states such as stent adhesion and slippage trends.

[0132] Compared with the prior art, this embodiment has the following advantages:

[0133] High prediction accuracy: By integrating the global, abstract "structural latent variables" extracted by the deep learning model with the manually defined "aneurysm neck geometric features" with clear anatomical significance, the model can simultaneously understand the macroscopic morphology of blood vessels and the microscopic constraints of key areas, significantly improving the accuracy of stent deployment location prediction.

[0134] High degree of automation: This invention realizes an end-to-end automated processing flow from raw CTA images to the final stent deployment three-dimensional coordinate sequence, without the need for manual intervention, which greatly improves the efficiency of clinical preoperative planning and shortens preparation time.

[0135] High objectivity and repeatability: Based on a unified data processing flow and a defined algorithm model, this invention can generate unique and repeatable prediction results for the same image data, overcoming the subjectivity and inconsistency of manual planning.

[0136] Good robustness: By performing strict standardization processing (centering, normalization, pose alignment) on the input point cloud, the model can effectively cope with the differences in scale, position and pose caused by different patients and different scanning devices, and has good generalization ability and robustness.

[0137] The output mechanism is reasonable: The three-dimensional deployment location output mechanism of this embodiment fully relies on the deep encoding expression of fused structural features. By optimizing the decoder mapping path, it generates a stent node sequence with spatial rationality and anatomical consistency, providing a highly reliable auxiliary prediction basis for personalized interventional strategies in complex aneurysm environments.

[0138] This embodiment employs a unified three-dimensional geometric representation mechanism to perform unified point cloud modeling of aneurysms and related vascular structures, constructing a high-dimensional joint description space of vascular morphology and local structures. It then utilizes nonlinear latent variable learning, graph neural network fusion, and end-to-end regression networks to predict the deployment morphology of stents in complex vascular anatomy. Unlike traditional methods that rely on linear statistical modeling and complex finite element simulations, this embodiment emphasizes the nonlinear nature of structural distribution and the geometric dominance of the aneurysm neck region in its modeling philosophy, systematically modeling this structure-state mapping relationship through deep neural networks. The entire method is based on fused morphological modeling, with latent variable compression expression and topological structure fusion as its backbone, combined with local feature extraction, three-dimensional state prediction, and structural sampling mechanisms, forming an integrated algorithm path from input structure to output deployment state. This approach possesses technical advantages such as strong structural adaptability, high inference efficiency, and stable prediction accuracy.

[0139] It should be noted that all user data involved in this application is information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0140] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 5 As shown, the device includes a processor 60, a memory 61, an input device 62, and an output device 63; the number of processors 60 in the device can be one or more. Figure 5 Taking a processor 60 as an example; the processor 60, memory 61, input device 62, and output device 63 in the device can be connected via a bus or other means. Figure 5 Taking the example of a connection between China and Israel via a bus.

[0141] The memory 61, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the cerebral aneurysm stent deployment prediction method in this embodiment of the invention. The processor 60 executes various functional applications and data processing of the device by running the software programs, instructions, and modules stored in the memory 61, thereby realizing the aforementioned cerebral aneurysm stent deployment prediction method.

[0142] The memory 61 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory 61 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory 61 may further include memory remotely located relative to the processor 60, which can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0143] Input device 62 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Output device 63 may include display devices such as a display screen.

[0144] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the cerebral aneurysm stent deployment prediction method of any embodiment.

[0145] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0146] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0147] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0148] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as C or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A cerebral aneurysm stent deployment prediction method, characterized by, include: Acquire vascular images containing cerebral aneurysms and extract multiple vascular structures related to the aneurysms, wherein the multiple vascular structures include the aneurysm, the carrier artery segment, and adjacent major bifurcation segments that affect the anchoring of the aneurysm stent. Each vascular structure is encoded as a structural latent variable using a structural encoder. Using the structural latent variables of each vascular structure as nodes, edges are established according to the anatomical connection relationship of each vascular structure to generate a structural interaction graph. Then, a graph neural network is used to aggregate information from the structural interaction graph to obtain global latent variables that characterize global topological relationships. Extract the geometric features of the aneurysm neck and fuse the geometric features of the aneurysm neck with the global latent variables; The fused features are processed using a predictive decoder to predict the location sequence of the aneurysm stent deployment path.

2. The method of claim 1, wherein, The process of acquiring vascular images containing cerebral aneurysms and extracting multiple aneurysm-related vascular structures from them includes: Obtain CTA images of blood vessels containing cerebral aneurysms; The CTA images were segmented to obtain multiple vascular structures related to the aneurysm; The segmentation results are reconstructed into three-dimensional surfaces to obtain three-dimensional surface models of each blood vessel structure.

3. The method according to claim 1, characterized in that, The process of encoding each vascular structure into a structural latent variable using a structural encoder includes: Point cloud sampling was performed on the three-dimensional surface models of each blood vessel structure to obtain the three-dimensional point cloud of each blood vessel structure. Using a variational autoencoder, the three-dimensional point cloud of each vascular structure is encoded into the structural latent variables of each vascular structure.

4. The method according to claim 3, characterized in that, Before using a variational autoencoder to encode the three-dimensional point clouds of each vascular structure into structural latent variables of each vascular structure, the method further includes: During training, the encoder part of the variational autoencoder encodes the three-dimensional point cloud of each vascular structure into structural latent variables, and the decoder part of the variational autoencoder reconstructs each vascular structure based on each structural latent variable. The encoder and decoder parts are trained by using the reconstruction error between the reconstructed structure and the original structure, as well as the KL divergence of each latent variable of the structure. The trained encoder part is used as the structure encoder.

5. The method according to claim 1, characterized in that, The establishment of edges based on the anatomical connections of various vascular structures includes: Establish edges between vascular structural nodes that have anatomically connected relationships; Weights are assigned to each edge based on the distance to the vascular structures connected to each edge; If the point cloud overlap rate of the blood vessel structures connected by each side exceeds the set threshold, the weight of each side will be increased.

6. The method according to claim 1, characterized in that, Before aggregating information from the structural interaction graph using a graph neural network, the method further includes: During training, the graph neural network generates the latent vectors of each blood vessel structure in the structural interaction graph, and uses the weights of each edge to perform a weighted average of the differences in the latent vectors of the blood vessels connected by each edge. The graph neural network is trained by minimizing the weighted average result.

7. The method according to claim 1, characterized in that, The extraction of the geometric features of the aneurysm neck includes: Extract at least one of the following: the aneurysm neck opening angle, the aneurysm neck edge curvature, the aneurysm neck area, the center offset angle, and the aneurysm body axial offset amplitude.

8. The method according to claim 1, characterized in that, Before processing the fused features using a predictive decoder to predict the location sequence of the aneurysm stent deployment path, the method further includes: During training, the predictive decoder is trained by minimizing the difference between the predicted location sequence and the actual location sequence of the aneurysm stent deployment path, ensuring that the stent deployment morphology is most similar to the real blood vessel morphology, maximizing the path smoothness, and ensuring that the stent ends adhere to the vessel wall in the anchoring area.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the cerebral aneurysm stent deployment prediction method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the cerebral aneurysm stent deployment prediction method according to any one of claims 1-8.

Images