Information processing method, device and equipment for intracranial aneurysm hemodynamic prediction and medium

Abstract

The invention discloses an information processing method and device for intracranial aneurysm hemodynamic prediction, equipment and a medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: receiving a three-dimensional medical image containing an intracranial aneurysm, inputting the three-dimensional medical image into a target hemodynamic prediction model, and capturing the comprehensive morphological characteristics of the intracranial aneurysm from the three-dimensional medical image; inputting the blood flow boundary condition into the target hemodynamics prediction model, and performing low-dimensional mapping on the blood flow boundary condition to obtain a first low-dimensional feature vector; receiving a spatial point coordinate and a directional distance corresponding to each sampling point in the CFD model formed in the intracranial aneurysm area, and performing low-dimensional mapping on the spatial point coordinate and the directional distance to obtain a second low-dimensional feature vector; and performing feature fusion on the comprehensive morphological features, the first low-dimensional feature vectors and the second low-dimensional feature vectors through a feature fusion module to predict hemodynamic parameters. And end-to-end hemodynamic parameter prediction is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an information processing method, device, equipment and medium for predicting intracranial aneurysm hemodynamics. Background Art

[0002] In medicine, accurately predicting the distribution of physical fields (such as pressure and velocity fields) within complex geometric structures is crucial. For example, hemodynamic analysis of intracranial aneurysms (IAs) is crucial for assessing rupture risk and guiding clinical treatment plans. Early detection and risk assessment of unruptured intracranial aneurysms (UIAs) can help physicians implement timely interventions, potentially saving patients' lives.

[0003] Currently, the primary means of acquiring three-dimensional geometric information about a patient's blood vessels and aneurysms is through clinical medical imaging techniques, such as magnetic resonance imaging (MRI) and computed tomography (CT). These imaging data contain rich geometric morphological information. Computational fluid dynamics (CFD) simulations are currently the primary tool for studying the hemodynamic characteristics of aneurysms. CFD can be used to calculate key parameters such as pressure and velocity within aneurysms. Numerous clinical studies have demonstrated that traditional CFD methods are complex and time-consuming, requiring tedious steps such as geometry reconstruction and mesh generation. These methods require high computational resources and are difficult to apply to real-time clinical diagnosis. Deep learning-based physical field prediction methods lack generalization capabilities, have low prediction accuracy for novel geometries, and require case-specific adjustments, limiting their applicability. Parametric or point cloud-based neural operator methods have limited ability to represent complex geometries, and point cloud preprocessing increases process complexity, making them incompatible with clinical imaging workflows.

[0004] In summary, how to use three-dimensional medical images as input to achieve end-to-end rapid and accurate prediction of intracranial aneurysm hemodynamics is a technical problem to be solved in this field. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide an information processing method, device, equipment, and medium for predicting intracranial aneurysm hemodynamics, which can use three-dimensional medical images as input to achieve end-to-end rapid and accurate prediction of intracranial aneurysm hemodynamics. The specific scheme is as follows: In a first aspect, the present application discloses an information processing method for predicting intracranial aneurysm hemodynamics, applied to a computer device, comprising: receiving a three-dimensional medical image containing an intracranial aneurysm and inputting the three-dimensional medical image into a target hemodynamic prediction model so as to capture comprehensive morphological features of the intracranial aneurysm from the three-dimensional medical image through a geometric feature extraction module of the target hemodynamic prediction model; wherein the comprehensive morphological features include local morphological features and global anatomical features; receiving a blood flow boundary condition and inputting the blood flow boundary condition into the target hemodynamic prediction model so as to perform low-dimensional mapping on the blood flow boundary condition through a boundary condition encoding module of the target hemodynamic prediction model to obtain a first low-dimensional feature vector; wherein the first low-dimensional feature vector is a vector representation of the effect of the blood flow boundary condition on a physical field, wherein the physical field is a physical field related to hemodynamics; Receiving spatial point coordinates and orientation distances corresponding to each sampling point in a CFD model formed of an intracranial aneurysm region, and performing low-dimensional mapping on the spatial point coordinates and the orientation distances through a spatial query point processing module of the target hemodynamic prediction model to obtain a second low-dimensional feature vector; wherein the orientation distance is the minimum Euclidean distance between each sampling point and a wall surface of the CFD model of the intracranial aneurysm, and the second low-dimensional feature vector is a position attribute representation of the sampling point and a relationship between each sampling point and the wall surface of the CFD model of the intracranial aneurysm; The comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector are subjected to feature fusion by a feature fusion module of the target hemodynamic prediction model to predict the hemodynamic parameters of each sampling point based on the target fused features.

[0006] Optionally, before receiving the three-dimensional medical image containing the intracranial aneurysm, the method further includes: The original medical image is voxelized to convert the original medical image into a binary three-dimensional medical image; wherein the voxel value of the intracranial aneurysm region in the three-dimensional medical image is 1, and the voxel value of the non-intracranial aneurysm region is 0.

[0007] Optionally, the geometric feature extraction module of the target hemodynamic prediction model is a three-dimensional computer vision model; Accordingly, the geometric feature extraction module of the target hemodynamic prediction model is used to capture the comprehensive morphological features of the intracranial aneurysm from the three-dimensional medical image, including: The three-dimensional medical image is divided into blocks using the target hemodynamic prediction model to obtain a plurality of three-dimensional voxel blocks, and each of the three-dimensional voxel blocks is projected into a preset feature vector space to obtain an initial input sequence; Inputting the initial input sequence into the geometric feature extraction module, so that the geometric feature extraction module performs shallow feature extraction on the initial input sequence based on the local window self-attention and shift window mechanism to obtain local morphological features including blood vessel diameter, tumor neck width, and subcapsule morphology; The initial input sequence is downsampled by the geometric feature extraction module, and deep feature extraction is performed on the downsampled initial input sequence based on the local window self-attention and shift window mechanism to obtain global anatomical features including the position features of the intracranial aneurysm in the vascular tree and the spatial relationship between the intracranial aneurysm and the surrounding brain tissue.

[0008] Optionally, the boundary condition encoding module of the target hemodynamic prediction model is a multi-layer perceptron network structure; Accordingly, the boundary condition encoding module of the target hemodynamic prediction model performs low-dimensional mapping on the blood flow boundary condition to obtain a first low-dimensional feature vector, including: The nonlinear relationship between the inlet mass flow parameter and the outlet pressure parameter is captured by the multilayer perceptron network structure of the target hemodynamic prediction model to obtain a first low-dimensional feature vector.

[0009] Optionally, before receiving the spatial point coordinates and directional distances corresponding to each sampling point in the CFD model formed by the intracranial aneurysm region, the method further includes: Uniformly sample the region with a voxel value of 1 in the binarized three-dimensional medical image to obtain each sampling point; Determine the spatial point coordinates corresponding to each sampling point based on a preset spatial coordinate system; Obtaining the shortest distance between each sampling point and the wall of the CFD model of the intracranial aneurysm, and determining the directional distance from each sampling point to the wall of the CFD model based on the current position of each sampling point and the shortest distance; Wherein, if the current position is that the sampling point is inside the wall of the CFD model, the orientation distance is the negative shortest distance; if the current position is that the sampling point is on the wall of the CFD model, the orientation distance is 0.

[0010] Optionally, the spatial query point processing module of the target hemodynamic prediction model is a multilayer perceptron network structure; Accordingly, the spatial query point processing module of the target hemodynamic prediction model performs low-dimensional mapping on the spatial point coordinates and the directional distance to obtain a second low-dimensional feature vector, including: The spatial query point processing module of the target hemodynamic prediction model performs spatial position extraction and symbol mapping processing on the spatial point coordinates and the directional distance to obtain a second low-dimensional feature vector.

[0011] Optionally, the performing feature fusion on the comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector by a feature fusion module of the target hemodynamic prediction model includes: Performing feature dimension alignment and feature normalization on the comprehensive morphological feature, the first low-dimensional feature vector, and the second low-dimensional feature vector to obtain a target comprehensive morphological feature, a target first low-dimensional feature vector, and a target second low-dimensional feature vector; Performing feature splicing processing on the target comprehensive morphological feature, the target first low-dimensional feature vector, and the target second low-dimensional feature vector through a first feature fusion unit of the target hemodynamic prediction model to obtain a first fused feature; performing attention calculation on the target first low-dimensional feature vector and the target second low-dimensional feature vector through a second feature fusion unit of the target hemodynamic prediction model based on a cross-attention mechanism to obtain a second fused feature characterizing the blood flow response of the intracranial aneurysm; The first fused feature and the second fused feature are fused by a third feature fusion unit of the target hemodynamic prediction model and by bilinear transformation to obtain a target fused feature.

[0012] In a second aspect, the present application discloses an information processing device for predicting intracranial aneurysm hemodynamics, comprising: a feature extraction module, configured to receive a three-dimensional medical image containing an intracranial aneurysm and input the three-dimensional medical image into a target hemodynamic prediction model, so as to capture comprehensive morphological features of the intracranial aneurysm from the three-dimensional medical image via a geometric feature extraction module of the target hemodynamic prediction model; wherein the comprehensive morphological features include local morphological features and global anatomical features; a first mapping module, configured to receive a blood flow boundary condition and input the blood flow boundary condition into the target hemodynamic prediction model, so as to perform low-dimensional mapping on the blood flow boundary condition through the boundary condition encoding module of the target hemodynamic prediction model to obtain a first low-dimensional feature vector; wherein the first low-dimensional feature vector is a vector representation of the effect of the blood flow boundary condition on a physical field, wherein the physical field is a physical field related to hemodynamics; a second mapping module, configured to receive spatial point coordinates and orientation distances corresponding to each sampling point in the CFD model formed of the intracranial aneurysm region, and perform low-dimensional mapping on the spatial point coordinates and the orientation distances through the spatial query point processing module of the target hemodynamic prediction model to obtain a second low-dimensional feature vector; wherein the orientation distance is the minimum Euclidean distance between each sampling point and the wall surface of the CFD model of the intracranial aneurysm, and the second low-dimensional feature vector is a position attribute representation of the sampling point and a relationship between each sampling point and the wall surface of the CFD model of the intracranial aneurysm; A fusion prediction module is used to perform feature fusion on the comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector through the feature fusion module of the target hemodynamic prediction model, so as to predict the hemodynamic parameters of each sampling point based on the target fused features.

[0013] In a third aspect, the present application discloses an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the steps of the aforementioned disclosed information processing method for intracranial aneurysm hemodynamic prediction.

[0014] In a fourth aspect, the present application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, the steps of the aforementioned disclosed information processing method for intracranial aneurysm hemodynamic prediction are implemented.

[0015] It can be seen that the present application discloses an information processing method for hemodynamic prediction of intracranial aneurysms, which is applied to a computer device, including: receiving a three-dimensional medical image containing an intracranial aneurysm, and inputting the three-dimensional medical image into a target hemodynamic prediction model, so as to capture the comprehensive morphological characteristics of the intracranial aneurysm from the three-dimensional medical image through a geometric feature extraction module of the target hemodynamic prediction model; wherein, the comprehensive morphological characteristics include local morphological characteristics and global anatomical characteristics; receiving blood flow boundary conditions, and inputting the blood flow boundary conditions into the target hemodynamic prediction model, so as to perform low-dimensional mapping on the blood flow boundary conditions through a boundary condition encoding module of the target hemodynamic prediction model to obtain a first low-dimensional feature vector; wherein, the first low-dimensional feature vector is a vector of the influence of the blood flow boundary conditions on the physical field. The quantity represents that the physical field is a physical field related to hemodynamics; receiving the spatial point coordinates and orientation distance corresponding to each sampling point in the CFD model formed by the intracranial aneurysm area, and performing low-dimensional mapping on the spatial point coordinates and the orientation distance through the spatial query point processing module of the target hemodynamic prediction model to obtain a second low-dimensional feature vector; wherein the orientation distance is the minimum Euclidean distance between each sampling point and the wall of the CFD model of the intracranial aneurysm, and the second low-dimensional feature vector is a position attribute representation of the sampling point and the relationship between each sampling point and the wall of the CFD model of the intracranial aneurysm; performing feature fusion on the comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector through the feature fusion module of the target hemodynamic prediction model to predict the hemodynamic parameters of each sampling point based on the target fused features. It can be seen that by directly inputting the target hemodynamic prediction model into the three-dimensional medical image, bypassing the traditional CFD geometric reconstruction and mesh generation steps, the geometric feature extraction module is used to hierarchically capture the local morphology and global anatomical characteristics of the aneurysm, combined with the boundary condition encoding module for low-dimensional mapping of blood flow parameters and the spatial query point processing module for characterizing the relationship between the sampling point position and the vascular wall boundary. After feature fusion, end-to-end hemodynamic parameter prediction is achieved, which effectively solves the problems of complex prediction process, insufficient generalization ability and poor compatibility with clinical images. It has the advantages of high prediction efficiency, high accuracy and adaptability to clinical real-time diagnosis needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0017] Figure 1This is a flow chart of an information processing method for intracranial aneurysm hemodynamic prediction disclosed in this application; Figure 2 A flowchart of the neural operator model execution process and traditional CFD simulation method disclosed in this application; Figure 3 Schematic diagram of the overall architecture of a neural operator model based on the DeepONet framework disclosed in this application; Figure 4 This is a structural diagram of an information processing device for intracranial aneurysm hemodynamic prediction disclosed in this application; Figure 5 This is a structural diagram of an electronic device disclosed in this application. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] In medicine, accurately predicting the distribution of physical fields (such as pressure and velocity fields) within complex geometric structures is crucial. For example, hemodynamic analysis of intracranial aneurysms (IAs) is crucial for assessing rupture risk and guiding clinical treatment plans. Early detection and risk assessment of unruptured intracranial aneurysms (UIAs) can help physicians implement timely interventions, potentially saving patients' lives.

[0020] Currently, the primary means of obtaining three-dimensional geometric information about a patient's blood vessels and aneurysms is through clinical medical imaging techniques, such as magnetic resonance imaging (MRI) and computed tomography (CT). These imaging data contain rich geometric information. Computational fluid dynamics (CFD) simulations are currently the primary tool for studying the hemodynamic characteristics within aneurysms. CFD can be used to calculate key parameters such as pressure and velocity within the aneurysm.

[0021] Take the traditional CFD simulation method as an example. Its process is usually very complex and time-consuming. The first step is often to accurately reconstruct the three-dimensional geometric model based on medical imaging data. This requires professional software and operators. For aneurysms with complex morphology and large variability, this process is time-consuming and prone to errors. After reconstructing the geometry, a high-quality computational mesh needs to be generated. The quality of the mesh directly affects the accuracy and stability of the simulation results. Generating a mesh suitable for complex geometry is itself a technical challenge and also requires a lot of time and expertise. Next, boundary conditions are set based on clinical data or assumptions. Finally, numerical algorithms are used to solve the Navier-Stokes equations that describe fluid motion. The entire solution process is computationally intensive and needs to be run on high-performance computing resources for several hours or even longer to obtain results.

[0022] The main drawbacks of traditional CFD methods are their low efficiency, high cost, and complexity. Long computational times make them difficult to apply to rapid clinical diagnosis and real-time decision-making. The high requirements for specialized skills and computing resources also limit their widespread application.

[0023] Secondly, some physical field prediction methods based on deep learning have also been proposed. For example, some methods try to directly train neural networks to predict the physical fields on specific geometric shapes, or use PINNs (Physics-Informed Neural Networks) in combination with physical equations for solution. However, such methods often have the problem of insufficient generalization ability. They usually need to be trained for specific geometric shapes, and when encountering new geometric shapes that do not appear in the training set, the prediction accuracy will drop significantly. For patient-specific aneurysms with highly variable morphology, this means that the model is difficult to apply directly, and time-consuming data preparation or model fine-tuning may be required for each new case.

[0024] In addition, some neural operator methods based on parameterization or point cloud representation attempt to input geometric information into the neural operator model in the form of parameters or point clouds. The disadvantage of this type of method is that for geometric bodies such as aneurysms that are highly complex in morphology and difficult to accurately describe with a small number of parameters, parameter-based methods have limited expression capabilities. Although point clouds can represent complex shapes more finely, point cloud data is not the original output form of clinical medical images and requires additional preprocessing steps (extracting point clouds from images), which increases the complexity of the process and may lose some information. Therefore, these methods are inconvenient to connect with existing clinical imaging workflows.

[0025] In general, existing technical solutions have obvious limitations in quickly and accurately predicting complex geometric internal physical fields (such as aneurysm hemodynamics) from medical images, especially in terms of efficiency, generalization ability, and compatibility with clinical data processes.

[0026] The numerous clinical studies mentioned above have confirmed that traditional CFD methods are complex and time-consuming, requiring tedious steps such as geometry reconstruction and mesh generation. These methods require high computational resources and are difficult to apply to real-time clinical diagnosis. Deep learning-based physical field prediction methods lack generalization capabilities, have low prediction accuracy for new geometries, require individual adjustments for each case, and are therefore limited in their applicability. Parametric or point cloud-based neural operator methods have limited ability to express complex geometries, and point cloud preprocessing increases process complexity, making them incompatible with clinical imaging workflows.

[0027] To this end, the present invention provides an information processing solution for predicting the hemodynamics of intracranial aneurysms, which can use three-dimensional medical images as input to achieve end-to-end rapid and accurate prediction of the hemodynamics of intracranial aneurysms.

[0028] Reference Figure 1 As shown, an embodiment of the present invention discloses an information processing method for intracranial aneurysm hemodynamic prediction, which is applied to a computer device and includes:

[0029] Step S11: Receive a three-dimensional medical image containing an intracranial aneurysm, and input the three-dimensional medical image into a target hemodynamic prediction model so as to capture the comprehensive morphological features of the intracranial aneurysm from the three-dimensional medical image through a geometric feature extraction module of the target hemodynamic prediction model; wherein the comprehensive morphological features include local morphological features and global anatomical features.

[0030] In this embodiment, before receiving the three-dimensional medical image containing the intracranial aneurysm, the method further includes: performing voxelization processing on the original medical image to convert the original medical image into a binary three-dimensional medical image; wherein the voxel value of the intracranial aneurysm region in the three-dimensional medical image is 1, and the voxel value of the non-intracranial aneurysm region is 0. It is understandable that the binarization processing is performed on the original medical image (the region of interest containing the intracranial aneurysm region in the MRI image or CT image) to convert the voxel value of the intracranial aneurysm region to 1 and the voxel value of the non-intracranial aneurysm region to 0, thereby obtaining the binary three-dimensional medical image.

[0031] In this embodiment, the target hemodynamic prediction model is a prediction model built based on a deep neural operator (DeepONet framework), wherein the geometric feature extraction module of the target hemodynamic prediction model is a three-dimensional SwinTransformer (abbreviated as Swin-T, an image feature extraction network used in a geometric branch network) architecture; Accordingly, the geometric feature extraction module of the target hemodynamic prediction model is used to capture the comprehensive morphological features of the intracranial aneurysm from the three-dimensional medical image, including: The three-dimensional medical image is divided into blocks using the target hemodynamic prediction model to obtain a number of three-dimensional voxel blocks, and each of the three-dimensional voxel blocks is projected into a preset feature vector space to obtain an initial input sequence; the initial input sequence is input into the geometric feature extraction module, and the geometric feature extraction module performs shallow feature extraction on the initial input sequence based on the local window self-attention and shift window mechanism to obtain local morphological features including blood vessel diameter, aneurysm neck width, and subcapsular morphology; the initial input sequence is downsampled by the geometric feature extraction module, and deep feature extraction is performed on the downsampled initial input sequence based on the local window self-attention and shift window mechanism to obtain global anatomical features including the positional features of the intracranial aneurysm in the vascular tree and the spatial relationship between the intracranial aneurysm and the surrounding brain tissue. It can be understood that the geometric feature extraction module performs operations such as three-dimensional image segmentation, hierarchical feature representation, three-dimensional window division, and three-dimensional window shifting, and the specific execution operations are as follows: 3D Image Segmentation: The input 3D medical image (e.g., a CT or MRI image of size H×W×D) is segmented into a series of non-overlapping 3D voxel blocks (called voxel blocks or 3D patches), for example, each block is 4×4×4 voxels. The voxel values ​​of each 3D patch are flattened and projected into a high-dimensional feature vector through a linear layer (or 3D convolutional layer), forming a series of 3D patch tokens. These tokens constitute the initial input sequence of the SwinTransformer.

[0032] Hierarchical feature representation: The Swin Transformer maintains its hierarchical design, consisting of multiple stacked Swin Transformer blocks. Each stage contains multiple Swin Transformer blocks, and patch merging (image stitching) operations are introduced between stages. Patch merging specifically involves performing a three-dimensional patch merging operation at the end of each stage. For example, it merges adjacent 2×2×2 patch tokens into a larger patch token, while adjusting the feature dimension through a linear layer. This can gradually reduce the spatial resolution (number of voxel blocks) of the feature map while increasing the receptive field of each token, thereby extracting more abstract and global features.

[0033] 3D Window Partitioning: In each Swin Transformer Block, the feature map (composed of patch tokens) is partitioned into a series of non-overlapping 3D windows of fixed size, for example, each window is 7×7×7 patch tokens. The attention mechanism is only computed within these local 3D windows, which significantly reduces computational complexity and allows the model to process high-resolution 3D inputs.

[0034] 3D Shifting Window Mechanism: To enable information exchange between different 3D windows without increasing computational complexity, the Swin Transformer introduces a shifting window mechanism. In two consecutive Swin Transformer blocks, one block uses conventional 3D windowing, while the next block simultaneously shifts the feature map by half the window size in all three dimensions (x, y, z) (for example, if the window size is M×M×M, the shift is M / 2×M / 2×M / 2). After the shift, the 3D windowing is performed again. This allows patch tokens originally in different windows to be grouped into the same window after the shift, enabling cross-window information exchange and enhancing the model's receptive field and modeling capabilities.

[0035] Swin Transformer (Swin-T) can adapt to input images of different sizes and demonstrates its superiority in processing data of different resolutions, mainly due to its hierarchical architecture, core patch merging operation, and fixed window-based self-attention mechanism.

[0036] Specifically, the Swin Transformer first divides the original input image into a series of small, non-overlapping image patches (Patches). These image patches are then flattened and projected into a fixed-dimensional feature vector (called PatchToken). In this way, no matter how large the original image size is, this operation simply divides the image into more or fewer fixed-size patches. For example, a 224×224 image will produce 56×56 4×4 patches, while a 448×448 image will produce 112×112 4×4 patches. The model processes sequences of these Patch Tokens rather than raw pixels.

[0037] Furthermore, the Swin Transformer's network structure is designed to be divided into multiple stages. Between each stage, a patch merging operation is introduced. The patch merging operation merges adjacent 2×2×2 (3D) patch tokens into a larger patch token. During the merging process, a linear layer is used to adjust the feature dimension and effectively downsample the spatial resolution by a factor of 2. This layered and patch merging mechanism enables the Swin Transformer to gradually construct feature representations with different receptive fields, just like a convolutional neural network. For input images of different sizes, the patch merging operation merges the corresponding number of initial patches until the final feature map size is reached. This means that no matter how large the input image is, the model can gradually extract multi-scale features in this way.

[0038] Another core feature of the Swin Transformer is its windowed self-attention. Self-attention calculations are performed only within local, fixed-size windows (local windows) (for example, each window contains 7×7 patch tokens). It is important to note that the size of the local window is fixed and does not change with the size of the input image. This means that whether the input image is 224×224 or 1024×1024, the amount of self-attention calculations within each window is constant. When the input image size changes, only the total number of local windows changes, not the calculations within each local window. This makes the Swin Transformer more computationally efficient and able to flexibly handle inputs of different sizes because its core attention operation is local and fixed-size.

[0039] The Swin Transformer's layered design enables it to capture information at different scales, specifically extracting both shallow and deep features. The shallow layer primarily extracts local, fine-scale details from 3D images, including vascular morphology (vessel diameter, wall thickness, branching angles, and minor irregularities (such as wall bulges in early-stage aneurysms)), local lesion features (texture, density, and edge clarity of early-stage minor lesions), tissue texture (grayscale characteristics and texture details of local brain tissue), and microstructures (local orientation of small brain vascular networks and neural bundles). As patch merging progresses, the resolution of the feature map decreases, while the receptive field represented by each patch token increases. Therefore, the deep layer primarily extracts global contextual information and high-level semantic features. Specifically, this includes overall structural information (the anatomy of the entire brain, hemispheric symmetry, and the size and morphology of the ventricular system) and the global relationship between the lesion and surrounding tissues (the aneurysm's location within the vascular tree (e.g., within the Circle of Willis), its compressive effect on surrounding brain tissue, and its proximity to important neural structures). Multilesion association information (if multiple aneurysms or other lesions are present, their spatial distribution and interrelationships) and pathological patterns (identification of more complex pathological patterns, such as diffuse lesions, the overall distribution of multiple infarcts).

[0040] Step S12: Receive blood flow boundary conditions and input the blood flow boundary conditions into the target hemodynamic prediction model so as to perform low-dimensional mapping on the blood flow boundary conditions through the boundary condition encoding module of the target hemodynamic prediction model to obtain a first low-dimensional feature vector; wherein the first low-dimensional feature vector is a vector representation of the influence of the blood flow boundary conditions on the physical field, and the physical field is a physical field related to hemodynamics.

[0041] In this embodiment, the boundary condition encoding module of the target hemodynamic prediction model is a multilayer perceptron network structure; accordingly, the low-dimensional mapping of the blood flow boundary condition by the boundary condition encoding module of the target hemodynamic prediction model to obtain a first low-dimensional feature vector includes: capturing the nonlinear relationship between the inlet mass flow parameter and the outlet pressure parameter by the multilayer perceptron network structure of the target hemodynamic prediction model to obtain the first low-dimensional feature vector. It is understandable that the blood flow boundary condition is received, wherein the blood flow boundary condition specifically includes: inlet mass flow, outlet pressure, etc., wherein the inlet mass flow determines the overall amplitude of the blood flow velocity field and affects the formation of vortices within the tumor. Therefore, the inlet mass flow can be measured by phase contrast MRI (PC-MRI) to measure the real-time flow of the tumor-bearing artery (such as the internal carotid artery), or determined based on the standard flow range of healthy people, and this is not specifically limited. The outlet pressure constrains the pressure gradient of blood flow and affects the pressure distribution within the tumor. Therefore, it can be determined by the physiological pressure of the distal blood vessels (such as the diastolic pressure of the external carotid artery and the basilar artery, approximately 60-80 mmHg) or by the central pressure waveform (systolic / diastolic pressure curve) measured clinically, and there is no specific limitation on this. The above-mentioned blood flow boundary conditions are information that defines "what" the physical problem is, such as: temperature distribution, fluid velocity, displacement, etc. on a certain boundary. They are inputs in the form of functions that describe the behavior of the physical field at the boundary. In this way, by receiving the above-mentioned blood flow boundary conditions, the blood flow boundary conditions are then mapped to a low dimension using an MLP (Multilayer Perceptron) to capture the essential characteristics of how the blood flow boundary conditions affect the entire physical field solution, that is, to obtain the first low-dimensional eigenvector.

[0042] Step S13: Receive the spatial point coordinates and orientation distance corresponding to each sampling point in the CFD model formed by the intracranial aneurysm area, and perform low-dimensional mapping on the spatial point coordinates and the orientation distance through the spatial query point processing module of the target hemodynamic prediction model to obtain a second low-dimensional feature vector; wherein the orientation distance is the minimum Euclidean distance between each sampling point and the wall surface of the CFD model of the intracranial aneurysm, and the second low-dimensional feature vector is a position attribute representation of the sampling point and the relationship between each sampling point and the wall surface of the CFD model of the intracranial aneurysm.

[0043] In this embodiment, before receiving the spatial point coordinates and orientation distances corresponding to each sampling point in the CFD model formed in the intracranial aneurysm area, the method further includes: uniformly sampling the area with a voxel value of 1 in the binarized three-dimensional medical image to obtain each sampling point; determining the spatial point coordinates corresponding to each sampling point based on a preset spatial coordinate system; obtaining the shortest distance between each sampling point and the wall of the CFD model of the intracranial aneurysm, and determining the orientation distance from each sampling point to the wall of the CFD model based on the current position of each sampling point and the shortest distance; wherein, if the current position is that the sampling point is inside the wall of the CFD model, the orientation distance is the negative shortest distance, and if the current position is that the sampling point is on the wall of the CFD model, the orientation distance is 0. It can be understood that the coordinates (x, y, z) of the specific spatial points specified by the user or the system for predicting the physical field parameters are obtained, where the preset spatial coordinate system is the Cartesian coordinate system, and the sampling points are obtained by uniformly sampling at fixed intervals in the intracranial aneurysm area (aneurysm and surrounding vascular area), and then the signed distance function (SDF, Signed Distance Function, indicating the distance from the spatial point to the geometric boundary) value of the sampling point relative to the wall of the geometric body is obtained to obtain the directional distance. The SDF value is responsible for processing these spatial position information.

[0044] In this embodiment, the spatial query point processing module of the target hemodynamic prediction model is a multilayer perceptron network structure; accordingly, the spatial query point processing module of the target hemodynamic prediction model performs low-dimensional mapping on the spatial point coordinates and the directional distance to obtain a second low-dimensional feature vector, including: performing spatial position extraction and symbol mapping processing on the spatial point coordinates and the directional distance by the spatial query point processing module of the target hemodynamic prediction model to obtain the second low-dimensional feature vector. It can be understood that the spatial point coordinates and SDF values ​​are information that defines the "where" of the physical field, that is, the spatial location of the solution to be queried. In this way, by receiving the spatial point coordinates and SDF values, the spatial query point processing module constructed using the MLP learns a set of spatial basis functions and maps them to another low-dimensional space to obtain the second low-dimensional vector, which is an abstract representation of the spatial query point position attributes. The spatial query point processing module captures information such as the position of the spatial point in the physical domain and its relative relationship to the geometric boundary.

[0045] Step S14: performing feature fusion on the comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector through the feature fusion module of the target hemodynamic prediction model to predict the hemodynamic parameters of each sampling point based on the target fused features.

[0046] In this embodiment, the comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector are subjected to feature dimension alignment processing and feature normalization processing to obtain target comprehensive morphological features, target first low-dimensional feature vector, and target second low-dimensional feature vector; the target comprehensive morphological features, the target first low-dimensional feature vector, and the target second low-dimensional feature vector are subjected to feature splicing processing by the first feature fusion unit of the target hemodynamic prediction model to obtain first fused features; the target first low-dimensional feature vector and the target second low-dimensional feature vector are subjected to attention calculation based on the cross-attention mechanism by the second feature fusion unit of the target hemodynamic prediction model to obtain second fused features characterizing the blood flow response of intracranial aneurysms; the first fused features and the second fused features are fused by the third feature fusion unit of the target hemodynamic prediction model and through bilinear transformation to obtain target fused features. It is understandable that since each feature is extracted by different modules, it is easy to have dimensional mismatch. Therefore, before feature fusion, it is normalized and dimensionally aligned to obtain the target comprehensive morphological features, the target first low-dimensional feature vector, and the target second low-dimensional feature vector, and then enter the primary fusion step of feature splicing, as follows: Feature concatenation: The 32-dimensional target comprehensive morphological features, the 32-dimensional target first-low-dimensional feature vector, and the 32-dimensional target second-low-dimensional feature vector are concatenated channel by channel to form a 96-dimensional fused feature. The features are then compressed into a 96-dimensional vector through global average pooling. Channel weights are generated through two layers of MLP (96 dimensions to 16 dimensions to 96 dimensions), and an importance coefficient is assigned to each channel of the concatenated features (for example, the channel weight of the tumor neck geometry feature is increased by 2.3 times).

[0047] After the primary fusion, we proceed to the intermediate fusion step of space-boundary interaction modeling, as follows: Using the SDF value in the spatial feature (the second lowest-dimensional feature vector of the target) as the query and the boundary condition feature (the first lowest-dimensional feature vector of the target) as the key / value, the dependence of the boundary distance on the blood flow condition (such as the weight of the point-enhanced pressure feature with SDF < 0.5 mm) is calculated; the 32-dimensional interactive feature (the second fused feature) is output to characterize the blood flow response near the geometric boundary.

[0048] After the intermediate fusion step, we enter the advanced fusion step of geometric-physical joint coding, as follows: Input primary fusion features (96 dimensions) + intermediate interaction features (32 dimensions); A 48-dimensional joint feature (target fusion feature) is generated through bilinear transformation, of which the first 24 dimensions encode the constraints of the geometric shape on the physical field (such as the increase in flow velocity caused by tumor neck stenosis), and the second 24 dimensions encode the reaction of the physical field to the geometric boundary (such as the deformation trend of the tumor wall in the high-pressure area).

[0049] In this embodiment, the fused MLP processing flow involves passing the 48-dimensional joint features through two layers of MLP (from 48 to 64 to 32 dimensions), using LeakyReLU activation to capture the nonlinear coupling of pressure, velocity, and geometry. In the first MLP layer, the neck width from the geometric features is element-wise multiplied by the flow rate from the boundary conditions to generate a velocity amplification factor for the stenosis. A learnable linear scaling matrix is ​​then introduced to map the 32-dimensional features to the dimensional space of physical field parameters. The output physical field parameters related to hemodynamics include pressure field, velocity field components, and wall shear force. For the velocity field, a 3D vector (u, v, w) is output in m / s; for the pressure field, a scalar is output in Pa; and for the wall shear force, a scalar is output in Pa. The blood viscosity parameter is embedded via the MLP bias term.

[0050] The velocity field is the distribution of blood flow velocity vectors at various spatial points within the vessel. Its clinical significance lies in the fact that high-velocity areas (such as the aneurysm neck entrance) may increase shear stress on the vessel wall and is associated with the risk of aneurysm rupture. The pressure field is the distribution of fluid pressure at various spatial points within the vessel. Its clinical significance lies in the fact that pressure fluctuations within the aneurysm may cause aneurysm wall fatigue. The wall shear force field is the distribution of the tangential force exerted by blood flow on the vessel wall. Its clinical significance lies in the fact that low WSS areas (such as the aneurysm base) have been shown to be associated with aneurysm growth.

[0051] This model, which directly uses 3D medical images as input, can predict the internal physical fields of complex geometric objects directly and in one step from their 3D images. This completely bypasses the extremely time-consuming and complex intermediate steps required by traditional methods, such as mesh generation after 3D geometry reconstruction and traditional numerical solutions (such as CFD). This "end-to-end" approach greatly simplifies the process and is the key to achieving highly efficient predictions.

[0052] like Figure 2 As shown, Figure 2The first section describes the information processing method used for intracranial aneurysm hemodynamic prediction using CFD to obtain realistic hemodynamic labels for 3D images. The dataset construction and preprocessing involved two representative baseline aneurysm geometric models from a real intracranial aneurysm morphology database. To construct diverse training and testing datasets, these two baseline geometric models were subjected to various parameterized continuous deformation operations. These deformations included "deflation" and "inflation" operations to simulate aneurysm growth or shrinkage, with a total of 20 different deformations performed. Through these operations, 20 synthetic aneurysm geometric models with significantly different morphological features were generated. These models exhibited a rich diversity in both parent vessel morphology and aneurysm morphology (particularly aspect ratio), effectively representing the morphological variability of aneurysms encountered in clinical practice.

[0053] The 21 generated aneurysm geometric models were meshed with high quality to ensure that the mesh quality met the requirements of CFD simulation. Specifically, the minimum mesh size was set to 0.15 mm, and 10 layers of prismatic mesh were generated near the vessel wall to accurately capture the boundary layer flow characteristics. The three-dimensional steady-state Navier-Stokes equations were solved for each aneurysm geometric model using the solver for incompressible flow in the open source computational fluid dynamics software. The simulated fluid medium properties were set to simulate blood properties: density of 1050 kg / m³, dynamic viscosity of .

[0054] Boundary Conditions: A constant mass flow boundary condition was established at the vessel inlet. To simulate different physiological flow conditions, six different mass flow rates were selected, ranging from 0.001 kg / s to 0.004 kg / s. These flow rates were chosen to ensure laminar flow, consistent with the blood flow characteristics of most intracranial aneurysms. A zero pressure boundary condition was established at the vessel outlet.

[0055] CDF solution process: transient solver is used for simulation, and the virtual time step is set to , the total simulation time is set to 1 second, and a total of Iterations are performed to ensure that the solution process reaches steady-state convergence. The flow field results at 1 second (including the values ​​of the pressure and velocity vectors at all grid points in the calculation domain) are extracted as the true value label data for training and testing the model of the present invention.

[0056] like Figure 2As shown in the lower part, the generated aneurysm geometric model is voxelized using medical image processing and visualization software. This step converts the continuous geometric model into a binary image on a discrete 3D voxel grid (i.e., voxel values ​​are 1 within the aneurysm region and 0 outside), generating a high-resolution 3D binary image. These 3D binary images serve directly as input to the geometry branch (geometry feature extraction module) of the proposed neural operator model. This processing mimics the process of extracting aneurysm regions of interest (ROIs) from clinical MRI / CT scans and performing a simple binarization process, which is a key input data form for end-to-end prediction.

[0057] Training set: 40 of the 42 geometric shapes are selected and combined with 4 of the 6 inlet flow rates (0.001, 0.002, 0.003, and 0.004 kg / s), resulting in a total of 40 × 4 = 160 training samples.

[0058] Test set: The remaining two geometries were combined with the remaining two inlet flow rates (0.0015 and 0.0035 kg / s), resulting in a total of 2 × 2 = 4 test samples. This partitioning ensured that the geometries and inlet flow rates in the test set were not present in the training set, thus enabling a rigorous evaluation of the generalization ability of the proposed model.

[0059] Construction and training of neural operator model (target hemodynamic prediction model): Model architecture based on Figure 2 The overall functional block diagram shown here constructs the neural operator model proposed in this invention. The geometry branch (Geo-Branch) uses a 3D Swin Transformer architecture to process the input 3D binary image. The boundary condition branch (BC-Branch, boundary condition encoding module) and the backbone (Trunk, spatial query point processing module) both utilize MLP networks. The specific model hyperparameter settings, including the number of network layers, number of neurons, patch size, window size, and embedding dimension, are shown in Table 1: Table 1 Hyperparameter settings

[0060] Loss function definition: The sum of squared errors (SSE) including pressure and three velocity components (u, v, w) is used as the total loss function for model training In order to balance the magnitude differences of different physical quantities, different weight coefficients are set for pressure and velocity, among which the pressure weight coefficient is , speed weight coefficient ( , , ) are respectively represented as follows: .

[0061] Sum of Squared Error (SSE): ; in, Indicates the total number of sampling points under a single category, represents the true value of the pressure or velocity component, Represents the predicted value of the pressure or velocity component.

[0062] Average relative Standard (Mean Relative L2 Norm): ; in, Indicates the number of categories of hemodynamic parameters, such as flow rate, pressure and other different parameter types, representing the total number of parameter types for which errors need to be calculated; Represents the parameter category index, which is used to traverse different hemodynamic parameters; Indicates the Among the hemodynamic parameters, The true value of the sampling points, The target hemodynamic prediction model is Parameters, The predicted value output by each sampling point.

[0063] MNAE Mean Normalized Absolute Error: ; in, Indicates the The minimum value of the hemodynamic parameter among all sampling points, Indicates the The maximum value of the hemodynamic parameter at all sampling points, Indicates the Among the hemodynamic parameters, The true value of the sampling points; Indicates the Among the hemodynamic parameters, The predicted value of each sampling point.

[0064] Total loss function: ; Optimization algorithm: Use Adam optimizer to iteratively update model parameters.

[0065] Learning rate setting: During the entire training process, a fixed learning rate is set to .

[0066] Training process execution: The model is trained on a computing platform consisting of four high-performance A100-80G GPUs for a total of 40,000 epochs.

[0067] Model evaluation and result analysis: Evaluation metric selection: The sum of squared errors (SSE), mean relative L2 standard, and mean normalized absolute error (MNAE) are used as the main quantitative evaluation metrics to calculate the errors between the pressure and velocity components (u, v, w) predicted by the model and the true CFD values, respectively.

[0068] Comparative Verification of Geometric Representation Methods: In order to verify the effectiveness of image-based representation (IBR), a comparison model based on point cloud representation (PBR) extracted using PointNet++ was constructed and trained. The loss curves of the two models during training and on the test set were compared. The results show that the IBR model is slightly better than the PBR model in terms of training convergence smoothness and pressure prediction accuracy. These comparative results strongly support the superiority of the geometric feature extraction scheme based on images and Swin Transformer in the present invention.

[0069] Comparison with CFD Results: The pressure fields predicted by the trained neural operator model (as an example implementation) on the training and test sets are compared with the corresponding CFD simulation results to qualitatively evaluate the model's prediction accuracy. Furthermore, the summed MNAE of pressure and velocity on the test set is calculated to quantitatively evaluate the model's prediction accuracy and generalization ability to unseen data.

[0070] Computational efficiency evaluation: The inference time of the neural operator model on the A100-80G GPU or CPU was recorded and compared with the time required for traditional CFD simulation, highlighting the significant computational speed advantage of the present invention. (In the implementation of the present invention using information processing for intracranial aneurysm hemodynamic prediction as an example, the inference time was approximately 0.3 seconds, much faster than CFD.)

[0071] Figure 3 FIG is a schematic diagram of the overall architecture of the neural operator model based on the DeepONet framework proposed in the present invention. Figure 3As shown in the figure, the model is built on a stacked DeepONet framework and consists of the following modules: a geometry branch network (Geo-Branch), which receives 3D image inputs of complex geometries; a boundary condition branch network (BC-Branch), which receives boundary condition information that affects the distribution of physical fields; a backbone network (Trunk), which receives spatial point coordinates (xyz) and signed distance functions (SDFs) used to query physical field parameters; and a bypass linear scaler module (Linear Scaler), which scales the final prediction output based on the boundary conditions. The model enables end-to-end prediction from 3D image input to physical field output (such as pressure and velocity), without requiring intermediate geometry reconstruction, mesh generation, and traditional numerical solution steps. The model's final output is the physical field parameters corresponding to the query point under different boundary conditions and geometries.

[0072] It can be seen that by directly inputting the target hemodynamic prediction model into the three-dimensional medical image, bypassing the traditional CFD geometric reconstruction and mesh generation steps, the geometric feature extraction module is used to hierarchically capture the local morphology and global anatomical characteristics of the aneurysm, combined with the boundary condition encoding module for low-dimensional mapping of blood flow parameters and the spatial query point processing module for characterizing the relationship between the sampling point position and the vascular wall boundary. After feature fusion, end-to-end hemodynamic parameter prediction is achieved, which effectively solves the problems of complex prediction process, insufficient generalization ability and poor compatibility with clinical images. It has the advantages of high prediction efficiency, high accuracy and adaptability to clinical real-time diagnosis needs.

[0073] Reference Figure 4 As shown, the present invention also discloses an information processing device for intracranial aneurysm hemodynamic prediction, comprising: a feature extraction module 11 for receiving a three-dimensional medical image containing an intracranial aneurysm and inputting the three-dimensional medical image into a target hemodynamic prediction model so as to capture comprehensive morphological features of the intracranial aneurysm from the three-dimensional medical image via a geometric feature extraction module of the target hemodynamic prediction model; wherein the comprehensive morphological features include local morphological features and global anatomical features; a first mapping module 12, configured to receive a blood flow boundary condition and input the blood flow boundary condition into the target hemodynamic prediction model, so as to perform low-dimensional mapping on the blood flow boundary condition through the boundary condition encoding module of the target hemodynamic prediction model to obtain a first low-dimensional feature vector; wherein the first low-dimensional feature vector is a vector representation of the effect of the blood flow boundary condition on a physical field, wherein the physical field is a physical field related to hemodynamics; A second mapping module 13 is configured to receive the spatial point coordinates and orientation distances corresponding to each sampling point in the CFD model formed of the intracranial aneurysm region, and perform low-dimensional mapping on the spatial point coordinates and the orientation distances through the spatial query point processing module of the target hemodynamic prediction model to obtain a second low-dimensional feature vector; wherein the orientation distance is the minimum Euclidean distance between each sampling point and the wall surface of the CFD model of the intracranial aneurysm, and the second low-dimensional feature vector is a position attribute representation of the sampling point and the relationship between each sampling point and the wall boundary of the CFD model of the intracranial aneurysm; The fusion prediction module 14 is used to perform feature fusion on the comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector through the feature fusion module of the target hemodynamic prediction model, so as to predict the hemodynamic parameters of each sampling point based on the target fused features.

[0074] It can be seen that by directly inputting the target hemodynamic prediction model into the three-dimensional medical image, bypassing the traditional CFD geometric reconstruction and mesh generation steps, the geometric feature extraction module is used to hierarchically capture the local morphology and global anatomical characteristics of the aneurysm, combined with the boundary condition encoding module for low-dimensional mapping of blood flow parameters and the spatial query point processing module for characterizing the relationship between the sampling point position and the vascular wall boundary. After feature fusion, end-to-end hemodynamic parameter prediction is achieved, which effectively solves the problems of complex prediction process, insufficient generalization ability and poor compatibility with clinical images. It has the advantages of high prediction efficiency, high accuracy and adaptability to clinical real-time diagnosis needs.

[0075] Furthermore, the embodiment of the present application also discloses an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content in the diagram should not be considered as any limitation to the scope of application of the present application.

[0076] Figure 5 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of the present application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 is used to store a computer program, which is loaded and executed by the processor 21 to implement the relevant steps of the information processing method for intracranial aneurysm hemodynamic prediction disclosed in any of the aforementioned embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0077] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device. The communication protocol it follows is any communication protocol that can be applied to the technical solution of this application and is not specifically limited here; the input and output interface 25 is used to obtain external input data or output data to the outside world. Its specific interface type can be selected according to specific application needs and is not specifically limited here.

[0078] Among them, the processor 21 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 21 can be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 21 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0079] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or CD, etc. The resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0080] The operating system 221 is used to manage and control the various hardware devices and computer programs 222 on the electronic device 20, so as to enable the processor 21 to calculate and process the massive amount of data 223 in the memory 22. The operating system 221 can be Windows Server, Netware, Unix, Linux, etc. In addition to including computer programs capable of implementing the information processing method for intracranial aneurysm hemodynamic prediction performed by the electronic device 20 as disclosed in any of the aforementioned embodiments, the computer program 222 can further include computer programs capable of performing other specific tasks. In addition to data received by the electronic device and transmitted from external devices, the data 223 can also include data collected by its own input and output interface 25.

[0081] Furthermore, this application discloses a computer-readable storage medium for storing a computer program; wherein, when executed by a processor, the computer program implements the aforementioned information processing method for intracranial aneurysm hemodynamic prediction. The specific steps of this method can be found in the corresponding contents disclosed in the aforementioned embodiments and will not be further described here.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from the other embodiments. Reference can be made to the descriptions of the identical or similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and the relevant parts can be referred to the descriptions of the methods.

[0083] Professionals may further appreciate that the units and algorithmic steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. The steps of the method or algorithm described in conjunction with the embodiments disclosed herein can be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module can be placed in a random access memory RAM (Random Access Memory), memory, read-only memory ROM (Read Only Memory), electrically programmable EPROM (Electrically Programmable Read Only Memory), electrically erasable programmable EEPROM (Electric Erasable Programmable Read Only Memory), registers, hard disk, removable disk, CD-ROM (Compact Disc-Read Only Memory), or any other form of storage medium known in the technical field.

[0084] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0085] The above is a detailed introduction to the solution provided by the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. An information processing method for predicting intracranial aneurysm hemodynamics, characterized in that: Applicable to computer devices, including: receiving a three-dimensional medical image containing an intracranial aneurysm and inputting the three-dimensional medical image into a target hemodynamic prediction model so as to capture comprehensive morphological features of the intracranial aneurysm from the three-dimensional medical image through a geometric feature extraction module of the target hemodynamic prediction model; wherein the comprehensive morphological features include local morphological features and global anatomical features; receiving a blood flow boundary condition and inputting the blood flow boundary condition into the target hemodynamic prediction model so as to perform low-dimensional mapping on the blood flow boundary condition through a boundary condition encoding module of the target hemodynamic prediction model to obtain a first low-dimensional feature vector; wherein the first low-dimensional feature vector is a vector representation of the effect of the blood flow boundary condition on a physical field, wherein the physical field is a physical field related to hemodynamics; Receiving spatial point coordinates and orientation distances corresponding to each sampling point in a CFD model formed of an intracranial aneurysm region, and performing low-dimensional mapping on the spatial point coordinates and the orientation distances through a spatial query point processing module of the target hemodynamic prediction model to obtain a second low-dimensional feature vector; wherein the orientation distance is the minimum Euclidean distance between each sampling point and a wall surface of the CFD model of the intracranial aneurysm, and the second low-dimensional feature vector is a position attribute representation of the sampling point and a relationship between each sampling point and the wall surface of the CFD model of the intracranial aneurysm; The comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector are subjected to feature fusion by a feature fusion module of the target hemodynamic prediction model to predict the hemodynamic parameters of each sampling point based on the target fused features.

2. The information processing method for intracranial aneurysm hemodynamic prediction according to claim 1, characterized in that: Before receiving the three-dimensional medical image containing the intracranial aneurysm, the method further includes: The original medical image is voxelized to convert the original medical image into a binary three-dimensional medical image; wherein the voxel value of the intracranial aneurysm region in the three-dimensional medical image is 1, and the voxel value of the non-intracranial aneurysm region is 0.

3. The information processing method for intracranial aneurysm hemodynamic prediction according to claim 1, characterized in that: The geometric feature extraction module of the target hemodynamic prediction model is a three-dimensional computer vision model; Accordingly, the geometric feature extraction module of the target hemodynamic prediction model is used to capture the comprehensive morphological features of the intracranial aneurysm from the three-dimensional medical image, including: The three-dimensional medical image is divided into blocks using the target hemodynamic prediction model to obtain a plurality of three-dimensional voxel blocks, and each of the three-dimensional voxel blocks is projected into a preset feature vector space to obtain an initial input sequence; Inputting the initial input sequence into the geometric feature extraction module, so that the geometric feature extraction module performs shallow feature extraction on the initial input sequence based on the local window self-attention and shift window mechanism to obtain local morphological features including blood vessel diameter, tumor neck width, and subcapsule morphology; The initial input sequence is downsampled by the geometric feature extraction module, and deep feature extraction is performed on the downsampled initial input sequence based on the local window self-attention and shift window mechanism to obtain global anatomical features including the position features of the intracranial aneurysm in the vascular tree and the spatial relationship between the intracranial aneurysm and the surrounding brain tissue.

4. The information processing method for intracranial aneurysm hemodynamic prediction according to claim 1, characterized in that: The boundary condition encoding module of the target hemodynamic prediction model is a multi-layer perceptron network structure; Accordingly, the boundary condition encoding module of the target hemodynamic prediction model performs low-dimensional mapping on the blood flow boundary condition to obtain a first low-dimensional feature vector, including: The nonlinear relationship between the inlet mass flow parameter and the outlet pressure parameter is captured by the multilayer perceptron network structure of the target hemodynamic prediction model to obtain a first low-dimensional feature vector.

5. The information processing method for intracranial aneurysm hemodynamic prediction according to claim 2, characterized in that: Before receiving the spatial point coordinates and directional distances corresponding to each sampling point in the CFD model formed by the intracranial aneurysm region, the method further includes: Uniformly sample the region with a voxel value of 1 in the binarized three-dimensional medical image to obtain each sampling point; Determine the spatial point coordinates corresponding to each sampling point based on a preset spatial coordinate system; Obtaining the shortest distance between each sampling point and the wall of the CFD model of the intracranial aneurysm, and determining the directional distance from each sampling point to the wall of the CFD model based on the current position of each sampling point and the shortest distance; Wherein, if the current position is that the sampling point is inside the wall of the CFD model, the orientation distance is the negative shortest distance; if the current position is that the sampling point is on the wall of the CFD model, the orientation distance is 0.

6. The information processing method for intracranial aneurysm hemodynamic prediction according to claim 5, characterized in that: The spatial query point processing module of the target hemodynamic prediction model is a multi-layer perceptron network structure; Accordingly, the spatial query point processing module of the target hemodynamic prediction model performs low-dimensional mapping on the spatial point coordinates and the directional distance to obtain a second low-dimensional feature vector, including: The spatial query point processing module of the target hemodynamic prediction model performs spatial position extraction and symbol mapping processing on the spatial point coordinates and the directional distance to obtain a second low-dimensional feature vector.

7. The information processing method for intracranial aneurysm hemodynamic prediction according to claim 5, characterized in that: The step of fusing the comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector using a feature fusion module of the target hemodynamic prediction model includes: Performing feature dimension alignment and feature normalization on the comprehensive morphological feature, the first low-dimensional feature vector, and the second low-dimensional feature vector to obtain a target comprehensive morphological feature, a target first low-dimensional feature vector, and a target second low-dimensional feature vector; Performing feature splicing processing on the target comprehensive morphological feature, the target first low-dimensional feature vector, and the target second low-dimensional feature vector through a first feature fusion unit of the target hemodynamic prediction model to obtain a first fused feature; performing attention calculation on the target first low-dimensional feature vector and the target second low-dimensional feature vector through a second feature fusion unit of the target hemodynamic prediction model based on a cross-attention mechanism to obtain a second fused feature characterizing the blood flow response of the intracranial aneurysm; The first fused feature and the second fused feature are fused by a third feature fusion unit of the target hemodynamic prediction model and by bilinear transformation to obtain a target fused feature.

8. An information processing device for predicting intracranial aneurysm hemodynamics, characterized in that: include: a feature extraction module, configured to receive a three-dimensional medical image containing an intracranial aneurysm and input the three-dimensional medical image into a target hemodynamic prediction model, so as to capture comprehensive morphological features of the intracranial aneurysm from the three-dimensional medical image via a geometric feature extraction module of the target hemodynamic prediction model; wherein the comprehensive morphological features include local morphological features and global anatomical features; a first mapping module, configured to receive a blood flow boundary condition and input the blood flow boundary condition into the target hemodynamic prediction model, so as to perform low-dimensional mapping on the blood flow boundary condition through the boundary condition encoding module of the target hemodynamic prediction model to obtain a first low-dimensional feature vector; wherein the first low-dimensional feature vector is a vector representation of the effect of the blood flow boundary condition on a physical field, wherein the physical field is a physical field related to hemodynamics; a second mapping module, configured to receive spatial point coordinates and orientation distances corresponding to each sampling point in the CFD model formed of the intracranial aneurysm region, and perform low-dimensional mapping on the spatial point coordinates and the orientation distances through the spatial query point processing module of the target hemodynamic prediction model to obtain a second low-dimensional feature vector; wherein the orientation distance is the minimum Euclidean distance between each sampling point and the wall surface of the CFD model of the intracranial aneurysm, and the second low-dimensional feature vector is a position attribute representation of the sampling point and a relationship between each sampling point and the wall surface of the CFD model of the intracranial aneurysm; A fusion prediction module is used to perform feature fusion on the comprehensive morphological features, the first low-dimensional feature vector, and the second low-dimensional feature vector through the feature fusion module of the target hemodynamic prediction model, so as to predict the hemodynamic parameters of each sampling point based on the target fused features.

9. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the steps of the information processing method for intracranial aneurysm hemodynamic prediction according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that Used to store a computer program; wherein, when the computer program is executed by a processor, the steps of the information processing method for intracranial aneurysm hemodynamic prediction according to any one of claims 1 to 7 are implemented.

Images