Intelligent tumor image classification method based on three-dimensional medical images

Abstract

This invention relates to an intelligent tumor image classification method based on three-dimensional medical images, belonging to the interdisciplinary field of artificial intelligence-assisted medical image analysis and computer simulation technology. Addressing the problem that existing technologies heavily rely on physical elasticity measurement equipment when extracting tumor biomechanical properties, this method, based solely on static raw medical images, maps voxel grayscale to virtual tissue mass density and constructs a virtual extrapolation vector using the tumor surface mesh as a reference. Combining the local curvature of the mesh and the density spatial gradient, a three-dimensional virtual solid stress tensor field is generated through tensor analysis. This tensor is then transformed into a graph structure and input into a graph network to extract physical topological features, which are then used as query vectors input into a cross-attention mechanism to guide visual features to complete weight allocation and classification. This invention relies purely on computing power to inversely construct a virtual physical force field on static images, eliminating the limitations of peripheral hardware and significantly improving classification accuracy.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence-assisted medical image analysis and computer simulation technology, specifically, it relates to an intelligent classification method for tumor images based on three-dimensional medical images. Background Technology

[0002] With the rapid development of computer vision and deep learning technologies, intelligent assisted diagnosis technology for tumors based on 3D medical images (such as CT and MRI) has made significant progress in clinical applications. In the task of classifying benign and malignant tumors, in addition to traditional visual morphological and textural features, the biomechanical properties of tumor tissue and its surrounding microenvironment (such as tissue stiffness, stress distribution, and shear modulus) have been proven to be extremely crucial distinguishing indicators. Typically, due to the disordered proliferation of internal cells and invasive growth into surrounding healthy tissues, malignant tumors exhibit a distinctly different non-uniform distribution and high-stress state in their overall or local mechanical properties compared to benign tumors.

[0003] To acquire and utilize these biomechanical properties, existing technologies typically rely on specific elastography physics devices and techniques. For example, Chinese invention patent CN114758783B (Elastography Method, Device, Computer Equipment, and Storage Medium Based on Deep Learning) discloses a scheme that uses a deep learning network to replace the traditional finite element method to solve for tumor material parameters and stress distribution. Its actual implementation requires the use of MRI elastography (MRE) or OCT elastography equipment to apply mechanical waves to the human body using an external exciter, thereby physically measuring the true displacement field of the tissue surrounding the tumor. Another example is Chinese invention patent CN111681210A (Method for Identifying Benign and Malignant Breast Nodules Based on Shear Wave Elastography Using Deep Learning), which discloses the use of an ultrasonic shear wave device to acquire the physical elastography map of the lesion, and then uses this map as a feature input to a convolutional neural network for benign and malignant identification.

[0004] However, the aforementioned existing technologies face a specific technical problem in practical clinical application and the use of massive historical data: the effectiveness of these methods highly depends on the intervention of physical elastography hardware (such as vibrators, shear wave probes, etc.) to obtain real physical displacement or deformation data. Due to the high procurement cost and limited availability of elastography equipment, ordinary static basic 3D medical images (such as conventional plain or enhanced CT scans, and conventional MRI) cannot be directly applied to these advanced deep learning models with mechanical properties. Simultaneously, given the massive amount of existing static 3D medical image data accumulated in medical institutions, how to extract equivalent tumor biomechanical features through innovative data processing logic and use them for heterogeneous fusion classification without relying on any physical elastography measurement hardware and with only static 3D grayscale images is a specific technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent tumor image classification method based on three-dimensional medical images, which solves the problem in the prior art that it is extremely dependent on physical elasticity measurement equipment when acquiring tumor biomechanical characteristics.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A tumor image intelligent classification method based on 3D medical images includes the following parts:

[0008] The process involves acquiring an original 3D medical image, extracting a 3D surface mesh model containing the tumor region, inputting the original 3D medical image into a 3D convolutional neural network to extract a visual feature matrix, and inputting the fused features into a classifier to output the classification result.

[0009] The voxel gray values ​​of the original three-dimensional medical image are mapped to virtual tissue mass density;

[0010] Using the internal center point of the three-dimensional surface mesh model as a reference, a virtual thrust vector pointing towards the external mesh node is constructed;

[0011] The local Gaussian curvature at the grid nodes and the spatial gradient of the virtual tissue mass density in the external adjacent region are extracted. Based on the local Gaussian curvature and the spatial gradient of the virtual tissue mass density, the virtual thrust vector is analyzed by tensor analysis to generate a three-dimensional virtual solid stress tensor field covering the tumor edge.

[0012] The three-dimensional virtual solid stress tensor field is transformed into graph structure data containing nodes and edges, and the graph structure data is input into a graph convolutional neural network to extract physical topological features.

[0013] The physical topology features are used as query vectors and input into the cross-attention mechanism module to perform weight allocation and feature fusion on the visual feature matrix.

[0014] Furthermore, the method also includes a step of extracting virtual hydrodynamic features in parallel, specifically including:

[0015] The inverse of the voxel grayscale gradient in the original three-dimensional medical image is mapped to the virtual permeability of the porous medium;

[0016] The interior of the three-dimensional surface mesh model is defined as a virtual fluid source, and the virtual fluid diffusion velocity field in the solution space is calculated based on the virtual permeability of the porous medium.

[0017] The gradient abrupt change value of the virtual fluid diffusion velocity field at the boundary of the three-dimensional surface mesh model is extracted to generate a virtual fluid shear force map. The virtual fluid shear force map and the three-dimensional virtual solid stress tensor field are then concatenated and input into the graph convolutional neural network.

[0018] Furthermore, the steps for calculating the virtual fluid diffusion velocity field within the solution space include:

[0019] A three-dimensional partial differential equation is established based on Darcy's law of fluid diffusion;

[0020] The discrete voxel coordinates of the original three-dimensional medical image are used as the boundary conditions for solving the three-dimensional spatial partial differential equations by using the finite difference method to obtain the velocity vector set of each voxel node.

[0021] Furthermore, the steps for generating virtual fluid shear force maps include:

[0022] Calculate the dot product matrix between the normal vector of the grid node and the adjacent velocity vector set;

[0023] Based on the dot product matrix, the tangential velocity components parallel to the tangential plane of the three-dimensional surface mesh model are extracted, and the virtual fluid shear force map is constructed based on the distribution of the tangential velocity components.

[0024] Furthermore, the method also includes a step of extracting virtual thermodynamic features in parallel, specifically including:

[0025] The voxel gray values ​​of the original three-dimensional medical image are mapped to a virtual thermal conductivity tensor;

[0026] The center of the three-dimensional surface mesh model is set as an isothermal energy source, and the spatial distribution field of the virtual temperature gradient in the solution domain is calculated based on the three-dimensional distribution law of the virtual thermal conductivity tensor.

[0027] The three-dimensional matrix of the virtual temperature gradient spatial distribution field is input as an auxiliary attention map into the cross-attention mechanism module.

[0028] Furthermore, the steps for calculating the spatial distribution field of the virtual temperature gradient within the solution domain include:

[0029] Extract the local maxima of the virtual temperature gradient spatial distribution field under the multi-scale Gaussian pyramid;

[0030] The three-dimensional coordinates corresponding to the local maxima are added to the graph structure data as thermodynamic anomaly nodes.

[0031] Furthermore, the steps for constructing the virtual thrust vector pointing to the external mesh node include:

[0032] The original three-dimensional medical image is subjected to tubular structure filtering to extract the three-dimensional blood vessel density distribution matrix in the area surrounding the tumor.

[0033] Calculate the chemotaxis weight parameters for each grid node pointing to regions with high blood vessel density;

[0034] The initial isotropic extrapolation vector is multiplied by the chemitropic weighting parameter to generate an anisotropic virtual extrapolation vector with spatial directional differences.

[0035] Furthermore, the step of extracting the three-dimensional blood vessel density distribution matrix of the region surrounding the tumor by tubular structure filtering of the original three-dimensional medical image includes:

[0036] Calculate the Hessian matrix for each voxel in the original three-dimensional medical image;

[0037] Extract the three eigenvalues ​​of the Hessian matrix;

[0038] A set of voxels is selected that satisfies the condition that two large eigenvalues ​​are less than zero and a third eigenvalue is close to zero. The distribution of the voxel set is then used to construct the three-dimensional blood vessel density distribution matrix.

[0039] Furthermore, the step of mapping the inverse of the voxel grayscale spatial gradient in the original three-dimensional medical image to the virtual permeability of the porous medium includes:

[0040] Extract the gray-level co-occurrence matrix of a single voxel in three-dimensional space;

[0041] Calculate the local texture anisotropy vector of the gray-level co-occurrence matrix;

[0042] The local texture anisotropy vector is used as a direction correction multiplier, and combined with the reciprocal of the voxel grayscale gradient, to generate an anisotropic virtual permeability tensor.

[0043] Furthermore, when calculating the virtual fluid diffusion velocity field in the solution domain based on the virtual permeability of the porous medium, the anisotropic virtual permeability tensor is substituted as the flow field conduction coefficient of the partial differential equation for iterative solution.

[0044] The beneficial effects of this invention are:

[0045] 1. This invention overcomes the limitation of existing technologies that require physical excitation equipment or shear wave probes to obtain biomechanical characteristics. By performing specific mathematical combinations and tensor analysis of voxel grayscale gradients, local Gaussian curvature, and spatial distances, a three-dimensional virtual solid stress tensor field is inversely reconstructed in pure digital space. This approach enables ordinary static basic CT or MRI images to possess high-dimensional mechanical feature representations, greatly lowering the threshold for clinical application and making retrospective biomechanical analysis of massive amounts of existing static images possible.

[0046] 2. For complex malignant tumors exhibiting diffuse or permeable growth patterns that do not conform to the expansion laws of a single solid, this invention innovatively introduces a virtual fluid dynamics model of porous media in parallel. By utilizing the reciprocal of the gray-scale gradient to map virtual permeability and solving the virtual fluid shear force spectrum based on Darcy's law, it can accurately capture extremely subtle texture disturbances and local tissue friction effects at the tumor edge, effectively compensating for the feature blind spots of a single solid dynamics model and improving the generalization ability of the classifier when facing different subtypes of tumors.

[0047] 3. This invention abandons the conventional fusion method of simply stitching together channels after pixelating the mechanical field. By transforming the virtual physical tensor into graph-structured data, it uses a graph convolutional neural network to extract physical topological features and uses them as query vectors for the cross-attention mechanism to guide visual features. This scheme forces the classifier to assign higher evaluation weights to tissue boundaries that exhibit drastic shear stress abrupt changes or are affected by microvascular chemotaxis weights when extracting visual texture, fundamentally overcoming the technical defect of conventional convolutional networks that are prone to misjudgment due to interference from the edges of healthy organs. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of the virtual solid stress tensor extraction and fusion process of the present invention;

[0050] Figure 2 This is a schematic diagram illustrating the calculation principle of the virtual fluid shear force spectrum of this invention.

[0051] Figure 3 This is a flowchart of the blood vessel density weighted anisotropy correction process of the present invention. Detailed Implementation

[0052] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] A tumor image intelligent classification method based on 3D medical images includes the following parts: acquiring the original 3D medical image; extracting a 3D surface mesh model containing the tumor region; inputting the original 3D medical image into a 3D convolutional neural network to extract a visual feature matrix; and inputting the fused features into a classifier to output the classification result; mapping the voxel gray values ​​of the original 3D medical image to virtual tissue mass density; constructing a virtual thrust vector pointing to the external mesh nodes based on the internal center point of the 3D surface mesh model; extracting the local Gaussian curvature at the mesh nodes and the spatial gradient of the virtual tissue mass density in the external adjacent region; performing tensor analysis calculation on the virtual thrust vector based on the local Gaussian curvature and the spatial gradient of the virtual tissue mass density to generate a 3D virtual solid stress tensor field covering the tumor edge; converting the 3D virtual solid stress tensor field into graph structure data containing nodes and edges; inputting the graph structure data into a graph convolutional neural network to extract physical topological features; and using the physical topological features as query vectors to input into a cross-attention mechanism module to perform weight allocation and feature fusion on the visual feature matrix.

[0055] The overall technical concept of this embodiment lies in transforming static medical image data into dynamic biomechanical simulation data by relying on the central processing unit and graphics processing unit inside the computing device, completely freeing it from the constraints of external physical elasticity measurement hardware, such as... Figure 1 The virtual solid stress tensor extraction and fusion process shown begins with static 3D array data transmitted from the medical image acquisition device to the computing device's memory. The central processing unit (CPU) first reads this set of discrete voxels. To simulate the infiltration process of tumors in healthy tissue, the purely visual grayscale information must be converted into physical quantities reflecting tissue stiffness and resistance. At this stage, the system introduces a nonlinear mapping mechanism to convert voxel grayscale values ​​into virtual tissue mass density. Because the imaging responses of different tissues exhibit nonlinear differences, simple linear scaling can lead to confusion between the mechanical properties of dense and loose tissues. Therefore, a specific mapping function is needed to correct noise in low-density areas and compensate for the characteristic responses of high-density areas. Let the grayscale value of any voxel in the input 3D medical image in the spatial coordinate system be... The central processing unit presets a background noise grayscale threshold. When the gray value is below the critical value, it is considered as a completely unresisted liquid or cavity region; when the gray value is above the critical value, an exponential density growth model is used to simulate the density of the tissue.

[0056] Virtual organization quality density The calculation formula is defined as follows:

[0057] in Based on the density scaling factor, The parameter represents the tissue hardening rate. Using this mapping formula, the central processing unit (CPU) generates a 3D virtual density matrix in memory that is identical in size to the original image, providing fundamental environmental resistance data for subsequent mechanical simulations. After constructing the virtual tissue environment, the computing device needs to determine the source of the mechanics, namely, the expansion force simulating tumor growth. The graphics processing unit (GPU) uses the moving cube algorithm to process the preprocessed 3D medical image, extracting a 3D surface mesh model surrounding the tumor lesion. This model consists of a large number of discrete mesh nodes and triangular facets connecting the nodes. The GPU calculates the coordinates of the geometric center point of the tumor as the virtual expansion source by integrating and averaging the 3D spatial coordinates of all mesh nodes. Tumor growth is equivalently assumed to be an isotropic thrust radiating outward from this center point.

[0058] To quantify this thrust, the system constructs a virtual thrust vector for any node on the 3D surface mesh model. Since the thrust of tumor growth decreases with distance as it propagates outwards, and the decrease is inversely proportional to the square of the distance and the relative distance from the node to the center, the magnitude of the virtual thrust vector is set to follow this inverse attenuation rule. Let the coordinates of the tumor's geometric center be... The first on the surface mesh The coordinates of each node are The spatial distance vector between the two points is Then the initial virtual thrust vector received at that node The calculation formula is ,in The pre-defined intrinsic expansion kinetic energy coefficient of the tumor is used. This formula ensures that the surface nodes closer to the tumor center experience greater outward compressive force, perfectly fitting the solid mechanical expansion law within the tumor entity. After obtaining the initial thrust, the scheme enters the core stage of resistance and stress calculation, a process completed collaboratively by the tensor calculation units of the central processing unit and the graphics processing unit. The true biomechanical response is the result of the interplay between action and tissue reaction forces. When a tumor expands outward, it is not only affected by its own thrust but also by the uneven resistance of external healthy tissue, which manifests as spatial abrupt changes in density; simultaneously, the smoothness of the tumor surface directly determines the degree of force concentration. Therefore, the system extracts the first... The spatial gradient of virtual tissue mass density within a preset range adjacent to each grid node It is used to characterize the intensity of tissue resistance in that direction.

[0059] Simultaneously, the graphics processor calculates the local Gaussian curvature at this node. The curvature reflects the unevenness of the tumor surface; a larger curvature indicates a more pronounced tendency for spiky, invasive growth. Based on the local Gaussian curvature and the spatial gradient of virtual tissue mass density, the system performs tensor analysis on the initial virtual extrapolation vector to construct a three-dimensional virtual solid stress tensor matrix. .

[0060] This tensor is a third-order square matrix containing normal stress components along the main diagonal and shear stress components along the off-diagonal. For benign tumors, the surface is smooth and the surrounding tissue density is uniform, resulting in primarily uniform normal stress. However, malignant tumors, due to their irregular, crab-like infiltration at the edges, exhibit extremely high local Gaussian curvature and chaotic density gradients in the surrounding tissue, thus generating extremely high shear stress at specific nodes. The formula for calculating the stress tensor is constructed as follows:

[0061] in For virtual shear modulus correction parameters, For virtual bulk modulus parameters, This represents the outer product operation of vectors. Represents the trace of a matrix. This is a third-order identity matrix. This formula perfectly integrates geometric parameters, the spatial rate of change of virtual material density, and the central thrust into a physical tensor with nine components. The central processing unit (CPU) traverses all grid nodes, ultimately generating a three-dimensional virtual solid stress tensor field covering the entire tumor margin in space. To match this continuous physical tensor field with the discrete deep learning model, the system performs heterogeneous feature processing and transformation. Since the stress tensor field essentially expresses the topological relationships of forces in various regions of the tumor surface, rather than a simple image pixel matrix, the graphics processor transforms it into graph-structured data.

[0062] Using each node of the 3D surface mesh model as a vertex of the graph structure, the corresponding 3D virtual solid stress tensor matrix of the node is... The flattened vector serves as the initial feature vector for that vertex. Simultaneously, an adjacency matrix is ​​constructed based on the Euclidean distances of the nodes in 3D space. To avoid the enormous computational burden of fully connected graphs and to filter out the influence of irrelevant distant nodes, the system sets a truncation threshold distance. When the spatial distance between two nodes is less than the truncation threshold, a connecting edge is established between the two vertices in the graph, and the weight of the edge is inversely proportional to the spatial distance between the nodes; otherwise, the weight is set to zero. Let the nodes... and nodes The distance between them is Then the adjacent matrix elements The formula is ,in The Gaussian kernel width parameter is used to control the smoothness of the connection strength.

[0063] After constructing the graph structure data, it is input into a graph convolutional neural network deployed on a graphics processing unit (GPU). The graph convolutional neural network aggregates and updates the stress features of neighboring nodes using the Laplacian matrix. Its single-layer propagation formula is:

[0064] in For the first The node feature matrix of the layer To add a self-loop to the adjacency matrix, This is the corresponding degree matrix. This is the learnable weight matrix for this layer. The activation function is nonlinear. After multi-layer graph convolution feature extraction, the system performs global average pooling on all node features, outputting a dimensionality-reduced high-order physical topology feature vector. This vector highly condenses the distribution patterns of mechanical anomalies across the entire tumor surface. In the final feature fusion and classification stage, to achieve optimal guidance of visual features by physical features, this embodiment abandons the traditional channel stitching method and instead uses a high-performance tensor core within the computing device to execute a cross-attention mechanism module. Traditional parallel structures input the original 3D medical image into a 3D convolutional neural network, extracting a visual feature matrix containing rich texture and grayscale distribution information through the sliding operation of multiple 3D convolutional kernels.

[0065] However, visual features alone are prone to misjudgment when faced with malignant tumors with blurred boundaries. Therefore, the system uses the physical topological structure features output by the graph convolutional neural network as the query vector, and performs a linear transformation on the visual feature matrix output by the 3D convolutional neural network to serve as the key vector and value vector, respectively. The core logic of the cross-attention mechanism is to use the topological features that exhibit drastic anomalies in the mechanical space to find the most matching key lesion regions in the visual feature space, and assign higher evaluation coefficients to these regions. (Cross-attention feature output matrix) The mathematical expression is ,in The query matrix is ​​obtained by linear mapping of physical topological features. and These are the key matrix and value matrix obtained by mapping the visual feature matrix, respectively. is the scaling factor for the dimension of the key vector. The normalized exponential function is used to transform the scoring results into a weighted probability distribution with a sum of one. Using this formula, the normalized exponential function quickly identifies the intersection regions in the tensor computation unit that simultaneously exhibit visual texture disorder and physical virtual shear stress abrupt changes, effectively suppressing interference from artifacts and the edges of healthy tissues and organs. The fused global feature matrix, after being flattened, is input into a classifier constructed from multiple fully connected neurons. The classification function at the end of the classifier outputs the final intelligent classification result of the tumor image, including benign, malignant, and specific invasive levels.

[0066] In summary, this embodiment relies on the computing power allocation of the processor in the computing device to establish a closed-loop mapping system from the grayscale space of medical images to the stress space of solid mechanics. The entire scheme, by substituting voxel grayscale gradients, local surface mesh curvature, and distance attenuation thrust into the tensor analytical formula, precisely generates a three-dimensional virtual solid stress tensor field characterizing tumor growth and infiltration behavior under extreme conditions where data input from physical elasticity measurement equipment is completely lacking. Subsequently, leveraging the topological aggregation capability of graph convolutional neural networks and feature weighting achieved through a normalized exponential function in the cross-attention mechanism, the system forces the final image classification network to focus on evaluating tissue boundaries exhibiting drastic shear stress abrupt changes in the virtual mechanical physical field. This complete processing flow, utilizing rigorous mathematical parameterization and structured hardware information flow scheduling, fundamentally overcomes the high dependence of existing technologies on physical excitation equipment. It achieves high-dimensional physical and visual heterogeneous feature coupling analysis with low-cost static image input, ultimately achieving high-precision tumor classification and discrimination through an unexpected technical approach.

[0067] Example 2

[0068] A tumor image intelligent classification method based on 3D medical images includes a step of parallel extraction of virtual fluid dynamics features. Specifically, this includes: mapping the reciprocal of the voxel grayscale gradient in the original 3D medical image to the virtual permeability of a porous medium; defining the interior of a 3D surface mesh model as a virtual fluid source; calculating the virtual fluid diffusion velocity field within the spatial solution domain based on the virtual permeability of the porous medium; extracting the gradient abrupt change values ​​of the virtual fluid diffusion velocity field at the boundary of the 3D surface mesh model to generate a virtual fluid shear force map; and inputting the virtual fluid shear force map and the 3D virtual solid stress tensor field into a graph convolutional neural network after channel concatenation. The step of calculating the virtual fluid diffusion velocity field within the spatial solution domain includes: establishing a 3D spatial partial differential equation based on Darcy's law of fluid diffusion; using the discrete voxel coordinates of the original 3D medical image as the solution boundary condition, solving the 3D spatial partial differential equation using the finite difference method to obtain the velocity vector set for each voxel node. The steps for generating a virtual fluid shear force map include: calculating the dot product matrix of the normal vector of the mesh node and the adjacent velocity vector set; extracting the tangential velocity components parallel to the tangential plane of the 3D surface mesh model based on the dot product matrix; and constructing the virtual fluid shear force map based on the distribution of the tangential velocity components. The overall system architecture of this embodiment relies on a high-performance computing cluster of computing devices, aiming to replace traditional physical elasticity measurement hardware with pure computational fluid dynamics simulation.

[0069] The entire technical solution abandons the physical concept of treating tumors as solid entities with fixed hardness. Instead, it innovatively introduces the osmosis model from computational biology, transforming the tumor lesion into a virtual fluid source that continuously secretes pathological substances into surrounding healthy tissue. Figure 2 The diagram shown illustrates the calculation principle of the virtual fluid shear force map. In the initial processing stage of the information flow, the central processing unit (CPU) of the computing device is responsible for reading the preprocessed discrete original three-dimensional medical image matrix. To construct the spatial medium environment required for fluid diffusion, the CPU needs to assess the density of the surrounding tissue, as malignant tumors often disrupt the normal tissue texture during invasive growth, leading to drastic fluctuations in local grayscale distribution. The CPU calculates the grayscale gradient magnitude of each voxel in three-dimensional space. To overcome the noise amplification problem caused by simple linear mapping, the system introduces a permeability mapping function based on nonlinear changes. Let the grayscale value of the three-dimensional voxel at the coordinate point of the original image be... Its corresponding spatial gradient magnitude is The central processing unit presets a gradient threshold value representing the boundary between healthy, smooth tissue and pathological, disordered tissue. .

[0070] Virtual permeability of porous media The calculation formula is defined as follows:

[0071] in The system's preset maximum basic penetration rate coefficient for the organization. This is a smoothing factor to control the nonlinear decay rate. The physical meaning of this mathematical formula is that when the grayscale gradient of a voxel region is much lower than a critical value, it indicates that the region is homogeneous healthy tissue, and fluid can easily penetrate it, with its virtual permeability approaching its maximum value. Conversely, when the grayscale gradient exceeds the critical value, it indicates the encounter with a dense or structurally disordered lesion boundary, and its permeability exhibits an exponentially nonlinear decreasing trend. This generates a three-dimensional porous media space matrix in the computing memory that accurately reflects the tissue's resistance to fluid penetration. After completing the modeling of the three-dimensional porous media environment, the computing device calls the massively parallel computing core of the graphics processing unit (GPU) to execute the virtual flow field construction process based on Darcy's law. The GPU extracts the three-dimensional surface mesh model surrounding the tumor and uniformly defines all voxels within it as virtual fluid sources with constant secretory pressure. The fluid dynamics model assumes that the flow velocity of the virtual fluid in the porous media space is proportional to the pressure gradient and inversely proportional to the drag coefficient of the medium on the fluid. The GPU constructs a three-dimensional spatial partial differential equation describing the conservation of fluid mass and pressure transmission in the space.

[0072] Let the dynamic pressure field of the virtual fluid be... The hydrodynamic viscosity coefficient is The tumor secretion source item is Based on the virtual permeability of the porous medium calculated in the aforementioned steps, a system of equations is established. Since the partial differential equation cannot be solved analytically in the complex medical image solution domain, the graphics processor uses the discrete voxel coordinate array of the original 3D medical image as the solution space and sets Dirichlet boundary conditions with zero pressure decay at the image edges. By calling the finite difference algorithm library, the graphics processor discretizes the continuous partial differential operators into a system of difference equations between adjacent voxel nodes, and uses an iterative solver to continuously update the pressure value at each node until the overall pressure residual decreases to the convergence threshold. After the pressure field converges, the graphics processor further uses Darcy's law to calculate the velocity vector at every point in the entire field, and its velocity vector set... The calculation formula is Through this complex numerical algebraic operation, the system perfectly reproduces the three-dimensional velocity field distribution of virtual fluid diffusion and infiltration into surrounding tissues in digital space, relying solely on the topological relationship of static grayscale, without requiring any dynamic imaging or external excitation hardware intervention.

[0073] After obtaining the full-field velocity vector set, the system enters the core stage of feature extraction, namely, fluid shear force calculation. From a fluid dynamics perspective, the irregular, tentacle-like infiltration of malignant tumors causes extremely complex deflections and frictions in the flow field at the boundaries; this friction is represented by tangential shear force in fluid dynamics. The central processing unit of the computing device, in conjunction with the tensor acceleration unit, re-traverses the three-dimensional surface mesh model of the tumor. For any x-th ... For each node, extract its outward normal unit vector. And the virtual fluid velocity vector at the spatial location of the node is obtained by interpolation from the three-dimensional flow field. To isolate the simple outward expansion effect and focus on the frictional effect of the fluid along the surface, the central processing unit calculates the dot product matrix of the normal vector of the grid node and the adjacent velocity vector set, thereby decomposing the total velocity into normal velocity and tangential velocity.

[0074] Tangential velocity components parallel to the tangential plane of the 3D surface mesh model The calculation formula is Subsequently, the system quantifies the shear effect based on the drastic rate of change of the tangential velocity component along the normal direction. Let the minimum offset distance of the mesh boundary be... The virtual fluid shear force at this node The calculation formula is The tensor acceleration unit computes the shear force values ​​at all surface mesh nodes in parallel and maps these values ​​to the node feature attributes, ultimately constructing a high-resolution virtual fluid shear force map covering the entire tumor surface. This map, in the form of a quantitative and intuitive force field matrix, profoundly reveals the hidden infiltration features of tumor edges that are difficult for conventional pure visual convolutional networks to capture. Finally, the computing device needs to heterogeneously fuse and classify this unique fluid dynamic feature with traditional solid mechanical features or visual features. This embodiment abandons the simple numerical addition mechanism and instead introduces a dynamic feature balancing strategy within the deep learning framework.

[0075] The graphics processor performs a deep concatenation of the generated virtual fluid shear force map matrix with the three-dimensional virtual solid stress tensor field matrix obtained based on other physical models along the channel dimension. To enable the classification model to automatically find the optimal matching relationship between different physical fields during training and inference, a learnable weighted evaluation network is configured after the concatenated fusion layer. Let the extracted fluid dynamics feature vector be... The topological eigenvector of solid stress is The fused global representation matrix The calculation formula is:

[0076] in This indicates a tensor splicing operation. and This is a weighted evaluation coefficient matrix that is continuously optimized by the network during backpropagation. For bias vectors, This is a non-linear activation function. During network training, to mitigate the risk of gradient explosion caused by differences in the dimensions of different physical fields, the graphics processor introduces a variance penalty term in the loss function. This forces the network to increase its weights dependent on fluid shear force features when visual edges are blurred, thus achieving an optimal feature balance between the solid compression hypothesis and the fluid permeation hypothesis. After multiple fully connected mappings, the scorer outputs the final classification probability vector, accurately identifying the benign or malignant nature and invasive subtype of tumors. In summary, this embodiment, by deeply embedding porous media computational fluid dynamics theory in the underlying data processing chain, completely overturns the traditional paradigm in the field of medical image-assisted diagnosis that relies solely on grayscale visual texture.

[0077] The system utilizes rigorous partial differential equations and tensor dot product algebraic operations to perfectly equate the static medical grayscale distribution to fluid permeation resistance, successfully simulating the complex fluid shear stress field induced by tumors as abnormal secretion sources in digital space. This alternative not only forms a robust technical complement to solid mechanics expansion models, effectively filling the gap in feature extraction for tumors exhibiting diffuse or permeable growth patterns, but also completely eliminates hardware dependence on expensive external elasticity measurement equipment. Through a highly integrated equation solver and heterogeneous tensor fusion network within the computing hardware, this solution achieves a deep integration of prior knowledge of physical laws and data-driven deep learning, unexpectedly solving the classifier generalization bottleneck caused by cross-device imaging differences, and providing a highly robust and innovative intelligent tumor classification solution for clinical medicine.

[0078] Example 3

[0079] A tumor image intelligent classification method based on 3D medical images includes the following steps for constructing a virtual extrapolation vector pointing to external grid nodes: performing tubular structure filtering on the original 3D medical image to extract the 3D blood vessel density distribution matrix of the region surrounding the tumor; calculating the chemotaxis weight parameter for each grid node pointing to a high blood vessel density region; and performing a Hadamard product operation on the initial isotropic extrapolation vector and the chemotaxis weight parameter to generate an anisotropic virtual extrapolation vector with spatial directional differences. The steps for extracting the 3D blood vessel density distribution matrix of the region surrounding the tumor by performing tubular structure filtering on the original 3D medical image include: calculating the Hessian matrix of each voxel in the original 3D medical image; extracting three eigenvalues ​​of the Hessian matrix; selecting a set of voxels that satisfies the condition that two larger eigenvalues ​​are less than zero and the third eigenvalue is close to zero, and constructing the distribution of the voxel set as the 3D blood vessel density distribution matrix.

[0080] This embodiment, as a vertical extension of the virtual solid mechanics stress tensor scheme, aims to completely solve the technical problem of mechanical simulation distortion in specific complex physiological environments caused by the isotropic expansion assumption of tumors in the basic scheme. In the actual clinical pathological evolution, the infiltration and growth of malignant tumors are highly dependent on the blood and nutrient supply provided by the neovascular network. Therefore, tumors not only compress surrounding healthy tissues but also exhibit strong chemotaxis, spontaneously infiltrating and invading cancer cells more intensely and deeply along directions with dense blood vessels and rich blood supply. To capture this extremely subtle biophysical coupling phenomenon without adding any additional hardware detection equipment, utilizing only the data flow scheduling within the computing device, this embodiment constructs a blood vessel density-weighted anisotropic virtual solid stress solution model, such as... Figure 3 The illustrated process for weighted anisotropy correction of blood vessel density relies on high-performance computing devices equipped with high-throughput graphics processors and tensor acceleration cores. The system first needs to perform high-precision stripping and reconstruction of the capillary network hidden in the complex background of medical images in the spatial dimension. Since 3D medical images are presented as a large array of discrete voxels, blood vessels appear visually as continuously extending 3D tubular topologies in space. The central processing unit of the computing device reads the grayscale matrix data of the original 3D medical image according to a preset sliding window step size and calculates the second-order partial derivative of the image matrix at each discrete voxel coordinate point in 3D space, thereby constructing the Hessian matrix describing the curvature changes in the local grayscale space. The Hessian matrix is ​​a symmetric third-order square matrix whose internal elements precisely record the rate of change of grayscale values ​​in different directions.

[0081] To transform the abstract Hessian matrix into a geometrically meaningful metric, the tensor computation unit of the graphics processing unit (GPU) performs eigenvalue decomposition on the Hessian matrix of every voxel in the entire field. Through parallel algebraic computation, the tensor computation unit extracts three eigenvalues ​​of the Hessian matrix and arranges them in ascending order of absolute value, denoted as the first, second, and third eigenvalues, with the first eigenvalue having the smallest absolute value. In the physical mapping of medical images, the magnitude of the eigenvalue reflects the rate of change of grayscale along its corresponding eigenvector. For a typical three-dimensional tubular blood vessel structure, the grayscale along the direction of the vessel's centerline should remain smooth and continuous; therefore, the corresponding first eigenvalue must be close to zero. However, in any cross-sectional direction perpendicular to the vessel's direction, due to the contrast difference between the vessel and surrounding tissues, the grayscale will exhibit dramatic peak-like convexities or deep valley-like depressions, resulting in extremely large absolute values ​​for the corresponding second and third eigenvalues. Furthermore, if the blood vessel appears as a bright signal in the image, both of these larger eigenvalues ​​must be less than zero. Based on the aforementioned structural discrimination logic, the system constructs a nonlinearly varying tubular structure metric function model within the graphics processor. Let the cross-sectional asymmetry evaluation coefficient be the ratio of the absolute value of the second eigenvalue to the absolute value of the third eigenvalue; the speckle morphology suppression coefficient be the ratio of the absolute value of the first eigenvalue to the square root of the product of the absolute values ​​of the second and third eigenvalues; and the structural background noise measure be the square root of the sum of the squares of the three eigenvalues. The graphics processor combines these three parameters through exponential operations. The output value of the tubular structure metric function is proportional to the nonlinear combination attenuation of the cross-sectional asymmetry evaluation coefficient and the speckle morphology suppression coefficient. When a local voxel and its neighborhood completely satisfy the critical condition that two larger eigenvalues ​​are less than zero and the third eigenvalue is close to zero, the metric function model will output a response weight close to the maximum saturation threshold; conversely, if the local structure presents as a spherical nodule or a planar film, the metric function output value will exhibit a drastic exponential decay and approach zero. By performing the aforementioned high-dimensional filtering on the entire set of voxels, the computing device generated a high-precision three-dimensional blood vessel density distribution matrix in memory that was exactly the same size as the original medical image, thus completing the digital spatial preparation of the biological prior data.

[0082] After successfully extracting the three-dimensional vascular density distribution matrix, the system enters the core information processing stage for anisotropic expansion force correction. The virtual normal force generation module in the basic scheme calculates the uniform extrapolation force based solely on the distance from the mesh node to the geometric center of the tumor, neglecting the attraction of environmental microvessels. To transform this biological chemotaxis phenomenon into a computable physical tensor, the system needs to quantify the non-uniform pulling effect of blood vessels on the tumor surface. The computing device again extracts the three-dimensional surface mesh model surrounding the tumor region and performs neighborhood space analysis on each mesh node on the surface. The graphics processor extracts the three-dimensional vascular density distribution matrix data of the spatial coordinates of the mesh node and solves for the local three-dimensional vascular density spatial gradient vector of the node by calculating the density difference at that point along the three orthogonal spatial coordinate axes. The physical meaning of this gradient vector is that it indicates the absolute direction of the fastest increase in vascular density in the local space and the degree of increase. To enable precise asymmetric correction of the basic thrust in three-dimensional space, the system constructs a three-dimensional chemotaxis weight parameter vector that perfectly corresponds to the coordinate space dimensions. The calculation of this weight parameter vector is not a simple linear superposition, but rather introduces a smooth nonlinear rectified function to avoid abnormal thrust reversal effects in the negative gradient direction. The system inputs the local three-dimensional vascular density spatial gradient vector into the smooth nonlinear rectified function to obtain a positive gradient gain distribution. This gain distribution is then multiplied by the system's preset chemotactic response sensitivity coefficient and summed with a basic three-dimensional unit column vector where all elements are constants. Through this mathematical model correction, when a dense vascular network exists outside the grid nodes, the chemotactic weight parameter pointing to this high-vascular-density region will exhibit a significant nonlinear amplification trend; while on the tissue side lacking vascular supply, the weight parameter remains at the basic unit critical value, thus perfectly representing, at the digital computation level, the invasive desire of tumor cells to blood-rich areas.

[0083] After calculating the chemiluminescence weight parameters specific to each grid node, the tensor acceleration core then performs the synthesis operation of the force vector. The system extracts the isotropic initial virtual extrapolation vector generated in the basic scheme, which has only a uniform physical size along the center outward. The graphics processor calls low-level instructions to perform a strict Hadamard product operation on this isotropic initial virtual extrapolation vector and the aforementioned generated three-dimensional chemiluminescence weight parameter vector. The Hadamard product operation requires that the two mathematical objects involved in the calculation have completely identical dimensional structures; the calculation process is manifested as scalar multiplication of corresponding elements within the two vectors one by one. Through the solution of the Hadamard product operation, the components of the initial force vector on the three orthogonal coordinate axes are independently and non-uniformly scaled and corrected by the chemiluminescence weight parameters. Finally, the graphics processor outputs an anisotropic virtual extrapolation force vector with spatial directional differences. By superimposing microscopic blood supply environment data extracted from real images, this anisotropic virtual extrapolation vector exhibits a significantly larger value towards the direction of dense blood vessel density compared to other directions. This breaks the perfect spherical expansion morphology and displays highly realistic irregular tentacle-like biomechanical expansion characteristics. After obtaining the corrected core power source, the subsequent resistance and stress calculation processes of the system are seamlessly integrated into the basic tensor model. The computing device, combining the local Gaussian curvature at this node and the spatial gradient of virtual tissue mass density reflecting the resistance of healthy tissue, transforms the anisotropic virtual extrapolation vector with directional bias characteristics into the final anisotropic virtual solid stress tensor matrix through tensor cross-product and trace operation formulas.

[0084] The anisotropic virtual solid stress tensor matrix is ​​also mapped to high-dimensional graph structure data node features, which are then fed into a graph convolutional neural network configured in the graphics processor's memory for deep extraction of physical topological features. In the subsequent heterogeneous fusion classification step, the system continues to use a cross-attention mechanism module, using the graph network output rich in vascular chemotaxis physical priors as the query basis, and globally weighting the visual feature matrix extracted by the three-dimensional convolutional neural network. This forces the final classification decision network to focus its evaluation on the malignant infiltration boundary regions of tumors with extremely high vascular density traction and irregular grayscale textures. This extended embodiment successfully constructs a cross-disciplinary triple-nested analysis model by cleverly integrating the Hessian matrix image filtering algorithm, biological vascular chemotaxis prior laws, and solid mechanics tensor solution architecture. This scheme completely overcomes the inherent defect of conventional computer vision models that rely purely on grayscale morphology, which are prone to missed diagnosis blind spots when facing early malignant nodules with extremely blurred boundaries but rich vascular microcirculation. By relying entirely on the internal computing platform to perform inverse reconstruction calculations of anisotropic virtual stress environments on static basic medical images, the solution unexpectedly and significantly improves the classification accuracy and confidence of the evaluation coefficient of high-risk malignant tumors with crab-like invasive features. Furthermore, it completely isolates the dependence on any expensive external elastic imaging hardware and excitation equipment throughout the entire complex physical and biological feature extraction chain, forming a logically rigorous and highly preventive high-value technical implementation solution.

[0085] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0086] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in the claims, they should all fall within the protection scope of the present invention.

Claims

1. A tumor image intelligent classification method based on three-dimensional medical images, characterized in that, include: The process involves acquiring an original 3D medical image, extracting a 3D surface mesh model containing the tumor region, inputting the original 3D medical image into a 3D convolutional neural network to extract a visual feature matrix, and inputting the fused features into a classifier to output the classification result. The voxel gray values ​​of the original three-dimensional medical image are mapped to virtual tissue mass density; Using the internal center point of the three-dimensional surface mesh model as a reference, a virtual thrust vector pointing towards the external mesh node is constructed; The local Gaussian curvature at the grid node and the spatial gradient of the virtual tissue mass density in the external adjacent region are extracted. Based on the local Gaussian curvature and the spatial gradient of the virtual tissue mass density, the virtual thrust vector is analyzed by tensor analysis to generate a three-dimensional virtual solid stress tensor field covering the tumor edge. The three-dimensional virtual solid stress tensor field is transformed into graph structure data containing nodes and edges, and the graph structure data is input into a graph convolutional neural network to extract physical topological features. The physical topology features are used as query vectors and input into the cross-attention mechanism module to perform weight allocation and feature fusion on the visual feature matrix.

2. The intelligent tumor image classification method based on three-dimensional medical images according to claim 1, characterized in that, The method also includes a step of extracting virtual hydrodynamic features in parallel, specifically including: The inverse of the voxel grayscale gradient in the original three-dimensional medical image is mapped to the virtual permeability of the porous medium; The interior of the three-dimensional surface mesh model is defined as a virtual fluid source, and the virtual fluid diffusion velocity field in the solution space is calculated based on the virtual permeability of the porous medium. The gradient abrupt change value of the virtual fluid diffusion velocity field at the boundary of the three-dimensional surface mesh model is extracted to generate a virtual fluid shear force map. The virtual fluid shear force map and the three-dimensional virtual solid stress tensor field are then concatenated and input into the graph convolutional neural network.

3. The intelligent tumor image classification method based on three-dimensional medical images according to claim 2, characterized in that, The steps for calculating the virtual fluid diffusion velocity field in the computational solution domain include: A three-dimensional partial differential equation is established based on Darcy's law of fluid diffusion; The discrete voxel coordinates of the original three-dimensional medical image are used as the boundary conditions for solving the three-dimensional spatial partial differential equations by using the finite difference method to obtain the velocity vector set of each voxel node.

4. The intelligent tumor image classification method based on three-dimensional medical images according to claim 2, characterized in that, The steps for generating a virtual fluid shear force map include: Calculate the dot product matrix between the normal vector of the grid node and the adjacent velocity vector set; Based on the dot product matrix, the tangential velocity components parallel to the tangential plane of the three-dimensional surface mesh model are extracted, and the virtual fluid shear force map is constructed based on the distribution of the tangential velocity components.

5. The intelligent tumor image classification method based on three-dimensional medical images according to claim 1, characterized in that, The method also includes a step of extracting virtual thermodynamic features in parallel, specifically including: The voxel gray values ​​of the original three-dimensional medical image are mapped to a virtual thermal conductivity tensor; The center of the three-dimensional surface mesh model is set as an isothermal energy source, and the spatial distribution field of the virtual temperature gradient in the solution domain is calculated based on the three-dimensional distribution law of the virtual thermal conductivity tensor. The three-dimensional matrix of the virtual temperature gradient spatial distribution field is input as an auxiliary attention map into the cross-attention mechanism module.

6. The intelligent tumor image classification method based on three-dimensional medical images according to claim 5, characterized in that, The steps for calculating the spatial distribution of the virtual temperature gradient within the solution domain include: Extract the local maxima of the virtual temperature gradient spatial distribution field under the multi-scale Gaussian pyramid; The three-dimensional coordinates corresponding to the local maxima are added to the graph structure data as thermodynamic anomaly nodes.

7. The intelligent tumor image classification method based on three-dimensional medical images according to claim 1, characterized in that, The steps to construct a virtual thrust vector pointing to an external mesh node include: The original three-dimensional medical image is subjected to tubular structure filtering to extract the three-dimensional blood vessel density distribution matrix in the area surrounding the tumor; Calculate the chemotaxis weight parameters for each grid node pointing to regions with high blood vessel density; The initial isotropic extrapolation vector is multiplied by the chemitropic weighting parameter to generate an anisotropic virtual extrapolation vector with spatial directional differences.

8. The intelligent tumor image classification method based on three-dimensional medical images according to claim 7, characterized in that, The steps of extracting the three-dimensional blood vessel density distribution matrix of the region surrounding the tumor by tubular structure filtering of the original three-dimensional medical image include: Calculate the Hessian matrix for each voxel in the original three-dimensional medical image; Extract the three eigenvalues ​​of the Hessian matrix; A set of voxels is selected that satisfies the condition that two large eigenvalues ​​are less than zero and a third eigenvalue is close to zero. The distribution of the voxel set is then used to construct the three-dimensional blood vessel density distribution matrix.

9. The intelligent tumor image classification method based on three-dimensional medical images according to claim 2, characterized in that, The step of mapping the inverse of the voxel grayscale gradient in the original three-dimensional medical image to the virtual permeability of the porous medium includes: Extract the gray-level co-occurrence matrix of a single voxel in three-dimensional space; Calculate the local texture anisotropy vector of the gray-level co-occurrence matrix; The local texture anisotropy vector is used as a direction correction multiplier, and combined with the reciprocal of the voxel grayscale gradient, to generate an anisotropic virtual permeability tensor.

10. The intelligent tumor image classification method based on three-dimensional medical images according to claim 9, characterized in that, When calculating the virtual fluid diffusion velocity field in the solution domain based on the virtual permeability of the porous medium, the anisotropic virtual permeability tensor is substituted as the flow field conduction coefficient of the partial differential equation for iterative solution.

Images