A real-time rendering method for medical image boundary value enhancement
By transforming explicit geometric representations into implicit geometric representations and utilizing multilayer perceptron functions and center feature selection methods, the problem of uneven point density at the intersection of complex models in medical graphics rendering is solved, achieving more efficient, smoother, and more realistic rendering effects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF SHANGHAI FOR SCI & TECH
- Filing Date
- 2022-11-10
- Publication Date
- 2026-06-09
AI Technical Summary
Existing medical graphics rendering methods struggle to effectively handle uneven density at intersections of complex models, resulting in rendering complexity and inefficiency, making it difficult to achieve smooth and realistic rendering effects.
The explicit geometric representation is transformed into an implicit geometric representation. A virtual network is constructed using a multilayer perceptron function and a center feature selection method. Combined with Thiessen polygons and a virtual mesh, region partitioning and rendering optimization are performed.
It improves the simplicity and efficiency of rendering, enhances the smoothness and realism of rendering, and increases the processing speed.
Smart Images
Figure CN115830213B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of computer rendering, and in particular to a real-time rendering method for enhancing boundary values in medical graphics. Background Technology
[0002] Medical graphics often require rendering multiple tissues or organs simultaneously. To describe complex medical graphics simply and effectively, point clouds or subdivision surfaces are usually used. Since point clouds and subdivision surfaces are explicit geometric representations, they have the advantages of explicit geometric representations: they can represent the shape of complex objects relatively simply. However, it is difficult to determine the positional relationship between a point and the object. Furthermore, at the intersection of complex models, the density of points is often much greater than at other locations on the model itself. Summary of the Invention
[0003] To address the shortcomings of existing technologies, the present invention aims to provide a real-time rendering method for medical graphics boundary value enhancement, improving rendering simplicity and efficiency, increasing computational speed, and enhancing rendering smoothness and realism. To achieve the above-mentioned objectives and other advantages of the present invention, a real-time rendering method for medical graphics boundary value enhancement is provided, comprising:
[0004] S1. Perform preprocessing, and in the preprocessing stage, use the transformation from explicit geometric representation to implicit geometric representation to construct a complete geometric model of the medical graphics, given the input point cloud data:
[0005] Where x i Let I be the i-th point in the point cloud space, and R be the set of integers. 3 It is a three-dimensional space
[0006] Normal dataset Where n i Let I be the i-th normal data, where I is the set of integers and R is the set of integers. 3 It is a three-dimensional space;
[0007] S2. An approximate multilayer perceptron function, including a loss function, can be obtained through point cloud and normal data.
[0008] S3. Calculate the gradient in the loss function using numerical methods;
[0009] S4. Obtain a distance-signed function as a new model representation using the chain rule;
[0010] S5. Create a virtual network marker to define the graphical boundary;
[0011] S6. Draw a triangular mesh to create a lighting model and load the relevant texture maps;
[0012] S7. The graphics are drawn using rasterization, and the final rendering result is obtained.
[0013] Preferably, the loss function in step S2 can be expressed as:
[0014]
[0015] Where λ is a parameter greater than 0, and ||·|| is the Euclidean 2-norm.
[0016]
[0017] In the second term of formula (1), we find the mathematical expectation of the 2-norm of the gradient of the activation function f(x; θ) following a certain probability distribution.
[0018] Preferably, when using numerical methods to calculate the gradient in the loss function, the error in the numerical calculation should be included in the loss. Each layer of the activation function f of the multilayer perceptron model has a y... λ+1 =σ(Wy λ +b) where σ is a nonlinear differentiable activation function, and W and b are the learning parameters of each layer of the multilayer perceptron. The gradients can satisfy the following relationship through the chain rule:
[0019]
[0020] From formula (3), it can be seen that... Construct a multilayer perceptron model using f(x; θ).
[0021] Preferably, by training a multilayer perceptron model to convergence, any number of distance-signed functions can be obtained on any point cloud set X with arbitrary normal directions N.
[0022] Preferably, in step S5, the center feature selection method is used to divide the regions of points on different models, and the volume of the virtual mesh is adjusted according to the density of points. The virtual mesh uses a spatial regular hexahedron, and the following empirical formula is used to represent the relationship between the edge length of the spatial regular hexahedron and the number of points:
[0023] l=2+[c / 7.1]c>20 (4)
[0024] Where l is the edge length of the spatial hexahedron, and c is the number of points.
[0025] Preferably, after the virtual mesh is created, the geometric centroid of the i-th virtual mesh is selected as the central feature point p. i The maximum and minimum values are calculated based on the number of points of different models contained in this virtual grid, and finally the central feature point p is determined. i The entire virtual mesh is labeled with model feature values that contain the maximum number of points.
[0026] Preferably, Thiessen polygons are created using central feature points and their markers. The central feature points are the "seeds" of these Thiessen polygons. Several statistics within each Thiessen polygon are calculated iteratively until convergence. The weighted sum of the standard deviations of the statistics in each subdivision is calculated and compared with the data of the previous subdivision until convergence. In the continuous iteration, the minimized S is obtained.
[0027] Compared with existing technologies, the beneficial effects of this invention are: it is integrated into the hardware subdivision pipeline, uses a new paradigm to preprocess and obtain implicit geometric models, and uses a central feature selection method to divide the regions of points on different models, thereby improving the simplicity and efficiency of rendering. It not only improves the computing speed, but also speeds up the smoothness and realism of rendering. Attached Figure Description
[0028] Figure 1 This is a three-dimensional structural diagram of the real-time rendering method for medical graphics boundary value enhancement according to the present invention. Detailed Implementation
[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] Reference Figure 1 A real-time rendering method for boundary value enhancement of medical graphics includes: S1, performing preprocessing, and in the preprocessing stage, using implicit geometric representations transformed from explicit geometric representations to construct a complete geometric model of the medical graphics, given the input point cloud data:
[0031] Where x i Let I be the i-th point in the point cloud space, and R be the set of integers. 3 It is a three-dimensional space
[0032] Normal dataset Where n i Let I be the i-th normal data, where I is the set of integers and R is the set of integers. 3 It is a three-dimensional space;
[0033] S2. An approximate multilayer perceptron function, including a loss function, can be obtained through point cloud and normal data.
[0034] S3. Calculate the gradient in the loss function using numerical methods;
[0035] S4. Obtain a distance-signed function as a new model representation using the chain rule;
[0036] S5. Create a virtual network marker to define the graphical boundary;
[0037] S6. Draw a triangular mesh to create a lighting model and load the relevant texture maps;
[0038] S7. The graphics are drawn using rasterization, and the final rendering result is obtained.
[0039] Furthermore, the loss function in step S2 can be expressed as:
[0040]
[0041] Where λ is a parameter greater than 0, and ||·|| is the Euclidean 2-norm.
[0042]
[0043] In the second term of formula (1), the expected value of the L2 norm of the gradient of the activation function f(x; θ) following a certain probability distribution is calculated. When using numerical methods to calculate the gradient in the loss function, the error in the numerical calculation needs to be included in the loss. Each layer of the activation function f in the multilayer perceptron model has y λ+1 =σ(Wy λ +b) where σ is a nonlinear differentiable activation function, and W and b are the learning parameters of each layer of the multilayer perceptron. The gradients can satisfy the following relationship through the chain rule:
[0044]
[0045] From formula (3), it can be seen that... A multilayer perceptron model is constructed using f(x; θ). By training the multilayer perceptron model to convergence, any number of distance-signed functions can be obtained on any point cloud set X with arbitrary normal directions N. Using this representation method, the boundary features of the model can be distinguished more easily than with point clouds.
[0046] Furthermore, in step S5, the central feature selection method is used to divide the regions of points on different models. Virtual meshes are established at the intersection points of the models and divided according to formula (4). First, the density of points at this point is calculated, and the volume of the virtual mesh is estimated based on the density. After obtaining the edge length of the virtual mesh, the corresponding virtual mesh is created. At the intersection of complex models, the density of points is often much greater than that of other positions on the model itself. The volume of the virtual mesh is adjusted according to the density of points. The virtual mesh uses a spatial regular hexahedron. Based on multiple experiments, the following empirical formula is used to represent the relationship between the edge length of the spatial regular hexahedron and the number of points:
[0047] l=2+[c / 7.1]c>20 (4)
[0048] Where l is the edge length of the spatial regular hexahedron, and c is the number of points. After the virtual mesh is created, the geometric centroid of the i-th virtual mesh is selected as the central feature point p. i The maximum and minimum values are calculated based on the number of points of different models contained in this virtual grid, and finally the central feature point p is determined. i The entire virtual mesh is labeled with model feature values that contain the maximum number of points.
[0049] Furthermore, Thiessen polygons are created using central feature points and their labels. These central feature points serve as the "seeds" for these Thiessen polygons. Each Thiessen polygon defines a proximal region where any point is closer to its "seed" than any other point in the set. Several statistics within each Thiessen polygon are iteratively calculated until convergence. The key statistics are the number of models, the region point density, and the number of disjoint models. The weighted sum of the standard deviations of the statistics for each subdivision, S, is calculated and compared with the data from the previous subdivision until convergence. Through continuous iteration, the minimized S is obtained. S can be calculated using the following formula:
[0050]
[0051] Where n is the number of Thiessen polygons, M is the number of points used to create the virtual mesh, N is the number of virtual meshes, and k i t i w i These represent the number of models, the region point density, and the number of non-intersecting models for the i-th Thiessen polygon. It is the mean of the number of models, the density of regional points, and the number of non-intersecting models in this subdivision; minimizing S means that the selected parameters are distributed as evenly as possible in each polygon in the subdivision, and the labels of the selected central feature points more realistically reflect the boundary conditions of the models.
[0052] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention, and applications, modifications and variations thereof will be apparent to those skilled in the art.
[0053] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.
Claims
1. A real-time rendering method for medical graphics boundary value enhancement, characterized in that, Includes the following steps: S1. Perform preprocessing, and in the preprocessing stage, use the method of converting explicit geometric representation to implicit geometric representation to construct the geometric model of the complete medical graphics, given the input point cloud data: ; Where x i Let I be the i-th point in the point cloud space, and R be the set of integers. 3 It is a three-dimensional space; Normal dataset Where n i Let I be the i-th normal data, where I is the set of integers and R is the set of integers. 3 It is a three-dimensional space; S2. Obtain an approximate multilayer perceptron function using point cloud and normal data, including a loss function; the loss function in step S2 can be expressed as: (1)； in It is a parameter greater than 0. It is the Euclidean 2-norm. ； The second term in formula (1) is for finding the excitation function. The 2-norm of the gradient follows the mathematical expectation of a certain probability distribution; S3. Calculate the gradient in the loss function using numerical methods; S4. Obtain a distance-signed function as a new model representation using the chain rule; S5. Create a virtual network to mark the boundary of the graphic; in step S5, the center feature selection method is used to divide the regions of points on different models, and the volume of the virtual mesh is adjusted according to the density of points. The virtual mesh uses a spatial hexahedron, and the following empirical formula is used to express the relationship between the edge length of the spatial hexahedron and the number of points: c>20 (4)； Where l is the edge length of the spatial regular hexahedron, and c is the number of points; S6. Create a lighting model by drawing a triangular mesh and load the relevant texture maps; S7. The graphics are drawn using rasterization, and the final rendering result is obtained.
2. The real-time rendering method for medical graphics boundary value enhancement as described in claim 1, characterized in that, When calculating the gradient in the loss function using numerical methods, the error in the numerical calculation must be included in the loss. The activation function of the multilayer perceptron model... Each layer has ,in is a nonlinear differentiable activation function, and W and b are the learning parameters of each layer of the multilayer perceptron. The gradients satisfy the following relationship through the chain rule: (3)； From formula (3), it can be seen that... and Construct a multilayer perceptron model.
3. The real-time rendering method for medical graphics boundary value enhancement as described in claim 2, characterized in that, By training a multilayer perceptron model to converge, any number of distance-signed functions can be obtained on any point cloud set X with arbitrary normal directions N.
4. The real-time rendering method for medical graphics boundary value enhancement as described in claim 1, characterized in that, After the virtual mesh is created, the geometric centroid of the i-th virtual mesh is selected as the central feature point. The maximum and minimum values are calculated based on the number of points of different models contained in this virtual grid, and finally the central feature point is determined. The entire virtual mesh is labeled with model feature values that contain the maximum number of points.
5. A real-time rendering method for medical graphics boundary value enhancement as described in claim 4, characterized in that, Thiessen polygons are created using central feature points and their labels. The central feature points are the "seeds" of these Thiessen polygons. Several statistics within each Thiessen polygon are calculated iteratively until convergence. The weighted sum of the standard deviations of the statistics for each subdivision is calculated and compared with the data from the previous subdivision until convergence. In continuous iteration, the minimized S is obtained. The several statistics include the number of models, the density of region points, and the number of non-intersecting models.