Medical instrument personalized matching design method and system based on three-dimensional reconstruction

Abstract

The application provides a medical instrument personalized matching design method and system based on three-dimensional reconstruction, relates to the technical field of medical instrument design, and comprises the following steps: acquiring a medical image sequence, extracting feature points and calculating a feature vector, and realizing spatial alignment; segmenting and three-dimensionally reconstructing the aligned image; determining a design parameter range based on the extracted morphological feature parameters, combining a design rule library and an expert knowledge base; and obtaining optimal design parameters through multi-objective optimization. The application can realize precise personalized matching of medical instruments, and improve clinical adaptability and patient comfort.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical device design, and particularly relates to a medical device personalized matching design method and system based on three-dimensional reconstruction. BACKGROUND

[0002] With the development of medical technology and the increasing demand for individualized medical care, personalized design of medical devices has attracted more and more attention. Traditional medical devices are usually designed with standardization, which is difficult to meet the individual differences of patients. The personalized matching design method of medical devices based on three-dimensional reconstruction can significantly improve the matching degree and treatment effect of medical devices by obtaining medical image data of patients, performing three-dimensional reconstruction, and performing parameterized design of medical devices according to individual characteristics of patients.

[0003] At present, medical image processing technology, three-dimensional reconstruction technology and computer-aided design technology have made great progress. Medical images such as CT and MRI can provide detailed information of internal structure of human body, which provides data basis for personalized design of medical devices. Three-dimensional reconstruction technology can convert two-dimensional medical image sequence into three-dimensional model, and intuitively display the spatial form of human body structure. Computer-aided design technology can perform parameterized design and optimization of medical devices based on three-dimensional model. SUMMARY

[0004] The embodiment of the present application provides a personalized matching design method and system of medical devices based on three-dimensional reconstruction, which can solve the problems in the prior art.

[0005] In a first aspect, the embodiment of the present application provides a personalized matching design method of medical devices based on three-dimensional reconstruction, comprising:

[0006] Obtaining a medical image sequence of a matching part of a target object;

[0007] Extracting structure feature points and boundary feature points of the medical image sequence, and calculating feature vectors of each feature point; calculating an image transformation matrix based on the similarity and spatial consistency of the feature vectors, projecting each medical image to a reference image coordinate system and performing region fusion to obtain a spatially aligned medical image sequence;

[0008] Extracting edge features of the spatially aligned medical image sequence, determining initial segmentation points and performing region growing segmentation, fitting an initial reconstruction surface using non-uniform rational B-spline, extracting texture features and performing feature mapping to obtain a three-dimensional reconstruction model of the matching part of the target object, and extracting morphological feature parameters of the matching part;

[0009] Based on the morphological feature parameters, a multi-path reasoning algorithm is used to extract design rules from a medical instrument design rule library, and combined with an expert experience knowledge base, design parameter ranges of the personalized medical instrument are obtained.

[0010] Based on the mechanical properties and structural variables of the design parameter ranges, a multi-objective parameter optimization function is constructed, and a particle swarm optimization algorithm is used to solve it, to obtain the optimal design parameters of the personalized medical instrument.

[0011] The structural feature points and boundary feature points of the medical image sequence are extracted, and the feature vectors of each feature point are calculated; based on the similarity and spatial consistency of the feature vectors, an image transformation matrix is calculated, each medical image is projected into a reference image coordinate system and regionally fused, to obtain a spatially aligned medical image sequence, including:

[0012] The medical image sequence is decomposed by a Gaussian difference operator, the gradient direction distribution histogram of each medical image is calculated, the inflection points of the edge profile of the medical image are detected as boundary feature points; the local gray scale of each medical image pixel is calculated, a feature response map is generated and non-maximum suppression is performed, and the local maximum points are taken as structural feature points;

[0013] The gradient direction distribution in the neighborhood of the boundary feature points and the structural feature points is counted to generate a feature vector; between adjacent medical images, the similarity of the feature vectors is calculated to determine the initial matching point pairs, and the motion continuity and spatial distribution consistency of the initial matching point pairs are calculated to eliminate abnormal matching point pairs, and based on the similarity and spatial position relationship of the remaining matching point pairs, an image transformation matrix between adjacent images is calculated;

[0014] Based on the image transformation matrix, a global transformation matrix of each medical image relative to the reference image is calculated, the medical image sequence is projected into the reference image coordinate system according to the global transformation matrix, the overlapping regions are weighted and fused, and the interpolation filling is performed on the vacant regions, to generate a spatially aligned medical image sequence.

[0015] The edge features of the spatially aligned medical image sequence are extracted, the initial segmentation points are determined and region growing segmentation is performed, the initial reconstruction surface is fitted using non-uniform rational B-spline, the texture features are extracted and feature mapping is performed, to obtain a three-dimensional reconstruction model of the target object matching part, and the morphological feature parameters of the matching part are extracted, including:

[0016] An edge feature of the spatially aligned medical image sequence is extracted by using a multi-scale convolution to generate a tissue boundary contour map, and a tissue density distribution map is obtained by calculating a gray gradient value of each pixel point; an initial segmentation point is determined based on the tissue boundary contour map, and region growing is performed with gray gradient similarity as an expansion criterion, and adjacent pixel points with a gray gradient difference less than an average gradient of a tissue surface are marked as a same structure region;

[0017] Boundary points of the structure region are taken as control points, a three-dimensional curved surface is fitted based on a non-uniform rational B-spline method, and an initial reconstruction curved surface of a to-be-matched part of the target object is obtained;

[0018] Local texture features in each structure region are extracted, a matching score of the texture features and a vertex of the initial reconstruction curved surface is calculated, the texture features are mapped to a corresponding curved surface position according to the matching score, and a three-dimensional reconstruction model of the to-be-matched part of the target object is obtained by using weighted superposition of local region texture features to realize smooth transition for a position where the texture features jump, and morphological feature parameters of the to-be-matched part are extracted from the three-dimensional reconstruction model.

[0019] Boundary points of the structure region are taken as control points, a three-dimensional curved surface is fitted based on a non-uniform rational B-spline method, and an initial reconstruction curved surface of a to-be-matched part of the target object is obtained, including:

[0020] The coordinates of the boundary points of the structure region are taken as control point coordinates, local curvatures of the control points are calculated, the control point coordinates are mapped to a parameter space, a B-spline basis function of a specified order is constructed, a weight coefficient is calculated based on the local curvatures, and a rational B-spline basis function is constructed by combining the weight coefficient and the B-spline basis function;

[0021] The control point coordinates are substituted into the rational B-spline basis function to obtain a non-uniform rational B-spline curved surface, a fitting error between the non-uniform rational B-spline curved surface and the original boundary points is calculated, when a relative error of a structure region is greater than a preset error threshold, the control points of the structure region are gradient-optimized and adjusted to obtain an initial reconstruction curved surface of a to-be-matched part of the target object.

[0022] Based on the morphological feature parameters, a multi-path reasoning algorithm is used to extract design rules from a medical instrument design rule library, and combined with an expert experience knowledge base, design parameter ranges of the individualized medical instrument are obtained, including:

[0023] The design rules in the medical device design rule library are taken as branch nodes, and a rule path tree is constructed through the rule correlation degree between the nodes; the morphological feature parameters are reconstructed into a feature vector, the effectiveness score of the feature vector under the rule constraint corresponding to each branch node is calculated from the root node of the rule path tree, and the next layer of nodes to be accessed is determined; when the rule path search depth exceeds a preset layer number or the effectiveness score is lower than a rule screening threshold, the search of the branch is stopped, and multiple candidate rule paths are obtained;

[0024] Based on the rule correlation degree between each branch node, the integrity score of each candidate rule path is calculated, the candidate rule path with the highest integrity score is selected as the current design rule, the historical design parameter range of the current design rule is extracted from the expert experience knowledge base, the use frequency of each parameter value is calculated, the historical design parameter range is modified through the use frequency, and the design parameter range of the personalized medical device is obtained.

[0025] Based on the mechanical properties and structural variables of the design parameter range, a multi-objective parameter optimization function is constructed, and a particle swarm optimization algorithm is used for solving, to obtain the optimal design parameters of the personalized medical device, including:

[0026] The stress distribution diagram of the medical device under physiological load is calculated, the maximum stress, maximum strain and total strain energy are extracted as mechanical optimization variables; the compressive strength, friction coefficient and deformation variable of the surface of the medical device are calculated as structural optimization variables, and the correlation coefficient between each optimization variable is calculated to construct a multi-objective parameter optimization function;

[0027] The design parameter range is mapped to a particle swarm search space, a normalization coding rule is set for each design parameter, a particle position vector is coded as a design parameter combination, and the multi-objective optimization function value of each design parameter combination is calculated as the particle fitness, the fitness of the initial population particles is evaluated, and the particle position with the highest fitness is selected as the global optimal solution;

[0028] In the iterative optimization process, the adaptive inertia weight is set based on the Euclidean distance between the particle and the global optimal solution, the particle velocity is updated according to the inertia weight, and the velocity size is limited to ensure that the particle moves in the effective search space; the fitness of the updated particle position is calculated, and the global optimal solution is updated, and when the improvement amplitude of the global optimal solution in continuous preset iterations is less than a convergence threshold, the iteration is stopped, and the global optimal solution is decoded as the optimal design parameters of the personalized medical device.

[0029] The second aspect of the embodiment of the application provides a medical device personalized matching design system based on three-dimensional reconstruction, including:

[0030] The first unit is used for acquiring a medical image sequence of a target object to be matched;

[0031] a second unit configured to extract structure feature points and boundary feature points of the medical image sequence, and calculate a feature vector of each feature point; calculate an image transformation matrix based on similarity and spatial consistency of the feature vectors, project each medical image to a reference image coordinate system, and perform region fusion to obtain a spatially aligned medical image sequence;

[0032] a third unit configured to extract edge features of the spatially aligned medical image sequence, determine initial segmentation points and perform region growing segmentation, fit an initial reconstruction surface using a non-uniform rational B-spline, extract texture features and perform feature mapping to obtain a three-dimensional reconstruction model of a matching part of the target object, and extract morphological feature parameters of the matching part;

[0033] a fourth unit configured to extract design rules from a medical instrument design rule library using a multi-path reasoning algorithm based on the morphological feature parameters, and combine an expert experience knowledge base to obtain a design parameter range of the individualized medical instrument;

[0034] a fifth unit configured to construct a multi-objective parameter optimization function based on mechanical properties and structural variables of the design parameter range, and solve the function using a particle swarm optimization algorithm to obtain optimal design parameters of the individualized medical instrument.

[0035] a third aspect of the embodiment of the application,

[0036] An electronic device is provided, comprising:

[0037] a processor;

[0038] a memory for storing processor-executable instructions;

[0039] The processor is configured to invoke the instructions stored in the memory to perform the method described above.

[0040] a fourth aspect of the embodiment of the application,

[0041] A computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0042] The beneficial effects of the present application are as follows:

[0043] The medical instrument individualized matching design method based on three-dimensional reconstruction provided by the application realizes accurate spatial alignment by extracting medical image feature points and calculating feature vectors, solves the problem of low image registration accuracy in traditional methods, and improves the accuracy of three-dimensional reconstruction.

[0044] The method combines non-uniform rational B-spline surface fitting and texture feature mapping technology, realizes high-fidelity three-dimensional reconstruction of complex anatomical structure, and can accurately extract individual-specific morphological feature parameters of patients, thereby providing a reliable anatomical basis for personalized design of medical devices.

[0045] By fusing the multi-path reasoning algorithm and the expert experience knowledge base, and combining the multi-objective particle swarm optimization technology, the method can realize optimal matching design of the medical device while meeting the mechanical performance requirements, greatly improve the matching degree of the medical device and the patient's anatomical structure and the clinical applicability, and reduce the risk of complications. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of a medical device personalized matching design method based on three-dimensional reconstruction according to an embodiment of the present application is shown in

[0047] Figure 2 A personalized medical device design parameter optimization flowchart is shown in DETAILED DESCRIPTION

[0048] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0049] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0050] Figure 1 A flowchart of a medical device personalized matching design method based on three-dimensional reconstruction according to an embodiment of the present application is shown in Figure 1 As shown in the figure, the method comprises:

[0051] Obtaining a medical image sequence of a target object to be matched part;

[0052] Extracting structure feature points and boundary feature points of the medical image sequence, and calculating feature vectors of each feature point; based on the similarity and spatial consistency of the feature vectors, calculating an image transformation matrix, projecting each medical image to a reference image coordinate system and performing region fusion to obtain a spatially aligned medical image sequence;

[0053] extracting edge features of the spatially aligned medical image sequence, determining initial segmentation points and performing region growing segmentation, fitting an initial reconstruction surface using non-uniform rational B-splines, extracting texture features and performing feature mapping to obtain a three-dimensional reconstruction model of a matching part of the target object, and extracting morphological feature parameters of the matching part;

[0054] Based on the morphological feature parameters, a multi-path reasoning algorithm is used to extract design rules from a medical instrument design rule library, and combined with an expert experience knowledge base, design parameter ranges of the individualized medical instrument are obtained.

[0055] Based on the mechanical properties and structural variables of the design parameter ranges, a multi-objective parameter optimization function is constructed, and a particle swarm optimization algorithm is used to solve it to obtain the optimal design parameters of the individualized medical instrument.

[0056] In an alternative embodiment, structural feature points and boundary feature points of the medical image sequence are extracted, and feature vectors of each feature point are calculated; based on the similarity and spatial consistency of the feature vectors, an image transformation matrix is calculated, each medical image is projected into a reference image coordinate system and region fusion is performed to obtain a spatially aligned medical image sequence, including:

[0057] The medical image sequence is decomposed using a Gaussian difference operator, the gradient direction distribution histogram of each medical image is calculated, the inflection points of the edge profile of the medical image are detected as boundary feature points, the local gray scale of each medical image pixel is calculated, a feature response map is generated and non-maximum suppression is performed, and the local maximum points are taken as structural feature points;

[0058] The gradient direction distribution in the neighborhood of the boundary feature points and the structural feature points is counted to generate a feature vector; between adjacent medical images, the similarity of the feature vectors is calculated to determine initial matching point pairs, and the motion continuity and spatial distribution consistency of the initial matching point pairs are calculated to eliminate abnormal matching point pairs, and based on the similarity and spatial position relationship of the remaining matching point pairs, an image transformation matrix between adjacent images is calculated;

[0059] Based on the image transformation matrix, a global transformation matrix of each medical image relative to a reference image is calculated, the medical image sequence is projected into a reference image coordinate system according to the global transformation matrix, the overlapping regions are weighted and fused, and the interpolation filling is performed on the blank regions to generate a spatially aligned medical image sequence.

[0060] In the embodiment of the present application, the spatial alignment of the medical image sequence requires the extraction of feature points of each image to establish a matching relationship. By extracting structural feature points and boundary feature points, important structural and contour information of the medical image is captured. For the extraction of the boundary feature points, multi-scale decomposition of the medical image is performed using a Gaussian difference operator. Specifically, Gaussian filtering is performed on each medical image at five scales, and the difference between adjacent scales constitutes a Gaussian difference image. On each difference image, a gradient direction distribution histogram is generated by calculating the gradient amplitude and direction of the pixel points. The gradient direction is quantized into eight direction intervals, and each pixel increases the corresponding value of the direction interval according to its gradient direction and amplitude.

[0061] For the extraction of the structural feature points, the local gray level characteristics of each pixel point in each medical image are calculated. For each pixel point, the gray level distribution in its 9x9 pixel neighborhood window is investigated, and the gray level mean, standard deviation, and gray level difference from the center pixel in the neighborhood are calculated. According to these statistical characteristics, a feature response value is calculated for each pixel point to construct a feature response map. The feature response value calculation takes into account the gray level contrast, local gray level change rate, and gray level uniformity. Specifically, the square of the gray level difference is multiplied by the standard deviation and then divided by the neighborhood area. Non-maximum suppression is performed on the feature response map. A local window of 11x11 is set, and only the pixel point with the maximum response value in the window is retained, and the response value must exceed a preset response threshold, which is set to 2.5 times the overall average of the response value. The local maximum value point that meets the condition is marked as a structural feature point.

[0062] A feature vector is calculated for each feature point for subsequent matching. The gradient direction distribution in the 16x16 pixel neighborhood around each boundary feature point and structural feature point is counted. The neighborhood is divided into 4x4 sub-blocks, and the gradient direction in each sub-block is quantized into eight direction intervals. For each interval, the gradient amplitude falling into the interval is accumulated to form an 8-dimensional vector. The 8-dimensional vectors of the 16 sub-blocks are concatenated to obtain a 128-dimensional feature vector. In order to improve the stability of the feature vector, the vector is normalized, i.e., each element of the vector is divided by the Euclidean norm of the vector to make it a unit vector.

[0063] The matching relationship of feature points between adjacent medical images is established. For each pair of feature points in the two images, the Euclidean distance between their feature vectors is calculated. For each feature point in image 1, the feature point in image 2 with the smallest distance is found. If the ratio of the smallest distance to the second smallest distance is less than 0.8, it is considered that the pair of feature points constitutes an initial matching pair. The initial matching pair may contain false matches and needs to be further screened. The motion vector of all initial matching pairs is calculated, that is, the size and direction of the position difference of the corresponding points in the two images. The motion continuity and spatial distribution consistency of each matching pair are calculated through cluster analysis of the motion vector. The motion continuity is measured by comparing the motion vector of the point pair with the motion vectors of the surrounding point pairs. The point pair whose difference exceeds twice the local average is considered abnormal. The spatial distribution consistency is evaluated by constructing a Delaunay triangulation diagram of the matching point pairs and checking the deformation degree of each triangle. The matching point pairs corresponding to the triangles whose area ratio change exceeds 50% or angle change exceeds 30 degrees are removed.

[0064] According to the reliable matching point pairs after screening, the transformation matrix between adjacent images is calculated. The image transformation parameters are estimated by weighted least squares method. The weight of each matching point pair consists of two parts: similarity weight and spatial position weight. The similarity weight directly uses the similarity value of the matching point pair. For example, if the similarity of a matching point pair is 0.85, the similarity weight is also 0.85. The spatial position weight considers the distribution of matching points in the image and encourages more evenly distributed points to obtain higher weight. In specific implementation, the image is divided into 5x5 grids, and the number of matching points in each grid is calculated. If there are n matching points in a certain grid, the spatial position weight of each matching point in the grid is 1 / n. In this way, the contributions of different regions can be balanced, and the maximum weight is the product of the similarity weight and the spatial position weight. For example, if the similarity of a matching point is 0.85 and there are 5 matching points in the grid, the maximum weight is 0.85x(1 / 5)=0.17.

[0065] In this embodiment, an affine transformation model is selected to describe the transformation relationship between images. Affine transformation includes four basic transformations: translation, rotation, scaling and shear. For a two-dimensional image, affine transformation has six parameters: a, b, c, d, t x and t y , where a, b, c, d describe rotation, scaling and shear, and t x and t y describe translation. For example, in medical CT sequence image registration, the transformation parameters between adjacent slices are a=0.98, b=0.02, c=-0.01, d=0.99, t x =5, t y =3, indicating that there is slight rotation and scaling, and several pixel translation. For each matching point pair (xs , y s ) and (x t , y t ), two linear equations can be established according to the affine transformation relationship, for example, for the matched point pair (100, 150) and (115, 160), two equations can be obtained: a x 100 + b x 150 + t x = 115 and c x 100 + d x 150 + t y = 160, for the 150 pairs of matched points after screening, 300 equations can be obtained to form an over-determined equation set.

[0066] When solving the over-determined equation set by using the weighted least squares method, the weight of each equation is considered, and in the implementation process, the coefficient matrix A and the constant vector b are established, each two rows of A correspond to the coefficients of a pair of matched points, and each two elements of b correspond to the target coordinates of the matched points. For example, for the aforementioned matched point pair, the corresponding two rows in the A matrix are [100, 150, 0, 0, 1, 0] and [0, 0, 100, 150, 0, 1], and the corresponding two elements in the b vector are 115 and 160. At the same time, a weight matrix W is constructed according to the weight of each pair of matched points, W is a diagonal matrix, and the diagonal elements are the square roots of the weights of the matched points. The weight matrix is applied to the coefficient matrix and the constant vector to obtain the modified coefficient matrix W A and the constant vector W b , and the transformation parameter vector x is obtained by solving the normal equation, and in the specific implementation, the singular value decomposition (SVD) method can be used to solve it to improve the numerical stability.

[0067] To reduce the influence of noise on the estimation of the transformation parameters, a regularization term is introduced to control the deviation of the transformation matrix from the unit matrix, and an additional constraint row is added to the coefficient matrix, which corresponds to each parameter of the unit matrix, that is, a = 1, b = 0, c = 0, d = 1, t x = 0, t y = 0, for example, for the constraint of a = 1, a row [1, 0, 0, 0, 0, 0] is added to the coefficient matrix A, and an element 1 is added to the corresponding position in the constant vector b, and the weight of the additional constraint row is determined by the regularization coefficient λ, the larger λ is, the closer the transformation matrix is to the unit matrix.

[0068] The setting of the regularization coefficient λ adopts a dynamic adjustment strategy and is adaptively set according to the number of matching points. Generally, λ takes the reciprocal of the number of matching points, for example, for 150 pairs of matching points, λ = 1 / 150 = 0.0067. When the matching points are unevenly distributed or the number of matching points is small, the value of λ is appropriately increased. When the matching points are evenly distributed and the number of matching points is sufficient, the value of λ can be reduced. In the experiment, the optimal value range of λ is determined to be 0.005 to 0.01 through cross-validation.

[0069] After the transformation parameter solving is completed, a transformation matrix T can be constructed, T is a 3x3 matrix, the first two rows contain the six parameters obtained by solving, and the last row is fixed as [0, 0, 1], for example, if a = 0.98, b = 0.02, c = -0.01, d = 0.99, t x = 5, t y = 3, then the transformation matrix T is [[0.98, 0.02, 5], [-0.01, 0.99, 3], [0, 0, 1]]. Using the transformation matrix, any point in the source image can be mapped to the target image coordinate system.

[0070] After the transformation matrix between adjacent images is calculated, the global transformation matrix of each image relative to the selected reference image is calculated by matrix multiplication. Selecting an image at a middle position in the sequence can minimize the cumulative error. According to the global transformation matrix, each image in the medical image sequence is projected into the reference image coordinate system, the overlapping area is weighted and fused, the weight is proportional to the distance of the pixel to the image edge, the farther the distance to the edge, the greater the weight, and the empty area is filled by bilinear interpolation to ensure the continuity of the image. Finally, a spatially aligned medical image sequence is generated, providing a basis for subsequent clinical analysis and processing.

[0071] In an optional implementation, edge features of the spatially aligned medical image sequence are extracted, initial segmentation points are determined, and region growing segmentation is performed. Non-uniform rational B-spline fitting is used to fit the initial reconstruction surface, texture features are extracted and feature mapping is performed to obtain a three-dimensional reconstruction model of a to-be-matched part of the target object, and morphological feature parameters of the to-be-matched part are extracted, including:

[0072] Edge features of the spatially aligned medical image sequence are extracted using multi-scale convolution to generate a tissue boundary contour map, and a tissue density distribution map is obtained by calculating the gray gradient value of each pixel point; initial segmentation points are determined based on the tissue boundary contour map, and region growing is performed with gray gradient similarity as the expansion criterion, and adjacent pixel points with a gray gradient difference less than the average gradient of the tissue surface are marked as the same structure region;

[0073] The boundary points of the structure region are taken as control points, a three-dimensional curved surface is fitted based on a non-uniform rational B-spline method, and an initial reconstruction curved surface of the to-be-matched part of the target object is obtained;

[0074] Local texture features in each structure region are extracted, a matching score of the texture features and vertices of the initial reconstruction curved surface is calculated, the texture features are mapped to corresponding curved surface positions according to the matching score, for a position where the texture features jump, a weighted superposition of local region texture features is used to realize smooth transition, a three-dimensional reconstruction model of the to-be-matched part of the target object is obtained, and morphological feature parameters of the to-be-matched part are extracted from the three-dimensional reconstruction model.

[0075] In the embodiment, for the spatially aligned medical image sequence, a multi-scale convolution algorithm is used to extract edge features, specifically, a 3x3, 5x5 and 7x7 convolution kernel of three different scales is used to perform convolution operation on the image respectively to capture edge information at different scales, for each pixel point (x, y), the convolution response value at three scales is calculated, and the maximum response value is taken as the edge intensity value E(x, y) of the point. When E(x, y) is greater than a preset response threshold T edge , T edge =15, the point is marked as an edge point, and thus a tissue boundary contour map is generated. At the same time, the gray gradient value G(x, y) of each pixel point is calculated, and a tissue density distribution map is constructed. In actual application, for a certain CT image sequence, after multi-scale convolution processing, the edge at the junction of soft tissue and bone tissue is successfully identified, and the edge intensity value is between 18-25, which is obviously higher than the threshold 15.

[0076] Based on the above tissue boundary contour map, initial segmentation points are determined, and points with a gray gradient value greater than 1.5 times the average gradient value of the region are automatically selected as initial seed points in a region where the boundary contour is obvious. Taking a certain hip joint CT image as an example, the average gradient value of the region is 12, and when a point with a gradient value of 20 is detected, it is marked as an initial segmentation point. Taking these initial segmentation points as starting points, region growing segmentation is performed. During the region growing process, for each adjacent point Q of the current region pixel point P, the difference Diff(P, Q) between the gray gradient value of the point Q and the point P is calculated, and if Diff(P, Q) is less than the average gradient value G avg of the tissue surface (set as 10), the point Q is marked as the same structure region and added to the processing queue. In this way, the region is gradually expanded until the point that cannot meet the gray gradient similarity criterion, experimental data shows that for bone tissue segmentation, this method can successfully stop growing when the boundary gradient changes significantly, i.e. the gradient difference is about 15-20, effectively avoiding the error fusion between structures.

[0077] After the region growing segmentation is completed, the boundary points of the structure region are taken as control points, and a three-dimensional curved surface is fitted by using a non-uniform rational B-spline (NURBS) method. In the specific implementation, a boundary point set {P i} is selected as the control points, each control point is assigned a weight value w i , and the initial value is 1.0. Node vectors U and V are defined to control the shape of the curved surface in two parameter directions. The node vectors are selected according to the uniform distribution principle, such as U = [0, 0, 0, 0, 1 / 4, 1 / 2, 3 / 4, 1, 1, 1, 1], wherein the repetition degree of the end points is 4, and the curved surface is ensured to pass through the boundary control points. For a complex anatomical structure such as a hip joint, a cubic NURBS curved surface is selected, and the number of control points is about 400, which can accurately fit the semispherical surface of the hip joint, and the fitting error is controlled within 0.5 mm. The initial reconstructed curved surface obtained in this way can basically reflect the geometric shape of the matching part of the target object.

[0078] Local texture features in each structure region are extracted, and four feature values including contrast, correlation, energy and homogeneity in different directions (0°, 45°, 90° and 135°) are calculated by using a gray level co-occurrence matrix (GLCM) method. For example, in a 3*3 neighborhood of a certain bone region, the calculated texture feature values are as follows: contrast = 2.8, correlation = 0.72, energy = 0.15 and homogeneity = 0.81. The extracted texture features are matched with the vertices of the initial reconstructed curved surface, and the Euclidean distance is calculated as the matching score. The lower the score is, the higher the matching degree is. For each vertex v i of the curved surface, the lowest-scored feature T j in the texture feature space is found, and T j is mapped to the corresponding position of v i .

[0079] During the texture feature mapping process, special attention should be paid to the position where the features jump. When the difference between the texture features of adjacent vertices exceeds a preset difference threshold, the weighted superposition of the texture features in the local region is used to realize smooth transition. The preset difference threshold can be set to 3. Specifically, for the vertex v k of the jump region, the weighted average value of the texture features of all the vertices within a radius r around the vertex is calculated. The radius is usually 2-3 vertex distances, and the weight is inversely proportional to the distance. This processing method can effectively eliminate the texture discontinuity at the joint cartilage and bone tissue junction, and make the texture transition of the reconstructed model surface natural.

[0080] After the texture mapping is completed, a three-dimensional reconstructed model of the matching part of the target object is obtained. The morphological feature parameters of the matching part, including the curved surface volume, surface area, maximum diameter and average curvature, are extracted from the model, which provides reliable morphological parameters for subsequent matching.

[0081] In an alternative embodiment, the boundary points of the structure region are taken as control points, and a three-dimensional curved surface is fitted based on a non-uniform rational B-spline method to obtain an initial reconstructed curved surface of the matching part of the target object, including:

[0082] The coordinates of the boundary points of the structure region are taken as control point coordinates, the local curvature of the control points is calculated, the control point coordinates are mapped to a parameter space, a B-spline basis function of a specified order is constructed, a weight coefficient is calculated based on the local curvature, and the weight coefficient is combined with the B-spline basis function to construct a rational B-spline basis function;

[0083] The control point coordinates are substituted into the rational B-spline basis function to obtain a non-uniform rational B-spline curved surface, the fitting error between the non-uniform rational B-spline curved surface and the original boundary points is calculated, and when the relative error of the structure region is greater than a preset error threshold, the control points of the structure region are gradient-optimized and adjusted to obtain an initial reconstructed curved surface of the matching part of the target object.

[0084] In the specific embodiment, the boundary point coordinate data of the structure region is read, and it is assumed that the obtained boundary point set is P, containing n points, and the three-dimensional coordinates of each point are represented as P i (x i , y i , z i ), where i is from 1 to n, and for each boundary point, the local curvature value is calculated, which can be realized by fitting a local curved surface composed of the boundary point and its adjacent points. It is assumed that the coordinates of point P i are (10.5, 15.2, 8.7), the points in the neighborhood with a radius of 2 mm around it are selected, 20 neighborhood points are obtained, and the coordinates of these points are (10.3, 15.0, 8.8), (10.7, 15.3, 8.6), etc., all of which are in the sphere with P i as the center and a radius of 2 mm.

[0085] After obtaining the neighborhood point set of point P i , a least squares method is used to fit a local curved surface, and a quadratic curved surface is used as the fitting model. The quadratic curved surface can be expressed as a quadratic function of Z value with respect to X and Y, specifically: Z = aX 2 + bY 2 + cXY + dX + eY + f, where a, b, c, d, e, and f are coefficients to be solved. Each point (X j , Y j , Z j ) in the neighborhood is substituted into the above equation to obtain an equation about the unknown coefficients, forming an over-determined equation set.

[0086] Using matrix decomposition techniques such as singular value decomposition (SVD) to solve the coefficients, for the neighborhood points in the above example, the coefficient values are: a = 0.05, b = 0.03, c = -0.01, d = 0.2, e = 0.15, f = 7.8, near point P i , the local surface approximation is Z = 0.05X 2 + 0.03Y 2 - 0.01XY + 0.2X + 0.15Y + 7.8, after obtaining the parameters of the fitted surface, the curvature value at point P i can be calculated, the curvature calculation is based on the first and second derivatives of the surface at the point. For a quadratic surface model, the first derivatives at point P i are: ∂Z / ∂X = 2aX + cY + d and ∂Z / ∂Y = 2bY + cX + e. The second derivatives are: ∂ 2 Z / ∂X 2 = 2a, ∂ 2 Z / ∂Y 2 = 2b, ∂ 2 Z / ∂X∂Y = c, based on the above derivative values, the shape operator of the surface is constructed, and the average curvature is calculated as the curvature value K i of point P i , the average curvature is the arithmetic mean of the principal curvatures, which are obtained by solving the eigenvalues of the surface shape operator. For the surface parameters fitted in the previous example, the calculated curvature value K i is 0.08, indicating the degree of curvature of the surface at the point, a larger curvature value indicates that the surface is more curved at the point, while a smaller curvature value indicates that the surface is relatively flat at the point. In practical applications, the curvature value at the sharp corners or edges of the point cloud surface is usually significantly higher than that of the flat area.

[0087] Map all boundary point coordinates to a two-dimensional parameter space to obtain parameter points (u i , v i ), the mapping method uses the chord length parameterization technique, in specific implementation, the boundary points are sorted to ensure that adjacent points are close in physical space, for adjacent boundary points P i and P i+1 , calculate the chord length distance d i,i+1 between them, and determine the parameter value according to the cumulative chord length ratio. For example, the parameter value u i can be distributed in the interval from 0 to 1, and v i is determined according to the distribution of the point in the other dimension. In practical applications, when the total number of boundary points is 500, the parameter space can be uniformly divided into a 20x25 grid.

[0088] The B-spline basis function of a specified order is constructed, and in this embodiment, a 3-order B-spline is selected as the basis function, that is, the surface has second-order continuity in each direction. To construct the non-uniform B-spline basis function, the node vectors U and V need to be defined. For the parameter u direction, the node vector U can be set to {0, 0, 0, 0, 1 / 4, 1 / 2, 3 / 4, 1, 1, 1, 1}, and for the parameter v direction, the node vector V can be set to {0, 0, 0, 0, 1 / 5, 2 / 5, 3 / 5, 4 / 5, 1, 1, 1, 1}. This node distribution method can ensure good fitting performance of the surface at the boundary.

[0089] Based on the previously calculated local curvature K i , the weight coefficient w i of each control point is calculated. The weight coefficient is proportional to the local curvature, and the specific calculation method is w i = 1 + α·K i , where α is a weight adjustment coefficient, which can be set to 0.5 in actual application. When the local curvature K i of a certain point is 0.2, the corresponding weight coefficient w i is 1.1. The introduction of the weight coefficient can make the surface have better fitting accuracy in high curvature areas.

[0090] The weight coefficient is combined with the B-spline basis function to construct the rational B-spline basis function. The rational B-spline basis function is obtained by multiplying the B-spline basis function by the weight coefficient and performing normalization processing. At the parameter point (u, v), the point coordinates on the NURBS surface are determined by the product of the control point coordinates and the rational B-spline basis function.

[0091] The constructed rational B-spline basis function and control point coordinates are used to calculate the non-uniform rational B-spline surface. For any point (u, v) in the parameter space, its corresponding three-dimensional space point coordinates are calculated. By uniformly sampling in the parameter space, for example, setting the step size to 0.01, point cloud data on the surface can be generated, forming a complete non-uniform rational B-spline surface representation.

[0092] The fitting error between the generated NURBS surface and the original boundary points is calculated. For each original boundary point P i , the nearest point Q i on the NURBS surface is found, the Euclidean distance e i = ||P i - Q i || is calculated, and the relative error is calculated, which is defined as the sum of all errors e iThe ratio of the root mean square of the fitting error to the characteristic value of the structure region size, in practical applications, assuming that there are 500 boundary points, the average error between the generated NURBS surface and the boundary points is 0.08 mm, and the characteristic size of the structure region is 50 mm, then the relative error is 0.16%.

[0093] When the relative error of the structure region is greater than the preset error threshold, the control points are adjusted by gradient optimization, the error threshold is set to 0.1%, if the calculated relative error is greater than the threshold, the control points are optimized, the gradient descent method is used in the optimization process, the fitting error is reduced by adjusting the control point coordinates, for each control point, the gradient of the error function with respect to the control point coordinates is calculated, and the control point is moved in the opposite direction of the gradient, the moving step can be set to 0.01 mm, the NURBS surface and the fitting error are recalculated after each adjustment, and the relative error is lower than the threshold or the maximum iteration number is reached 50.

[0094] After the above optimization, the initial reconstruction surface of the target object to be matched part is obtained, which has good smoothness and accuracy, and can accurately express the geometric characteristics of the target object to be matched part.

[0095] The three-dimensional surface fitted by the above non-uniform rational B-spline method provides a high-quality initial reconstruction model for subsequent matching operations, and can be effectively applied to the fields of medical implant design, personalized prosthesis manufacturing, etc.

[0096] In an optional implementation, based on the morphological feature parameters, a multi-path reasoning algorithm is used to extract design rules from a medical instrument design rule library, and combined with an expert experience knowledge base, design parameter ranges of the personalized medical instrument are obtained, including:

[0097] The design rules in the medical instrument design rule library are taken as branch nodes, and a rule path tree is constructed through the rule correlation degree between the nodes; the morphological feature parameters are reconstructed into a feature vector, the effectiveness score of the feature vector under the constraint of the rules corresponding to each branch node is calculated from the root node of the rule path tree, and the next layer of nodes to be accessed is determined; when the search depth of the rule path exceeds a preset layer number or the effectiveness score is lower than a rule screening threshold, the search of the branch is stopped, and a plurality of candidate rule paths are obtained;

[0098] Based on the rule correlation degree between the branch nodes, the completeness score of each candidate rule path is calculated, the candidate rule path with the highest completeness score is selected as the current design rule, the historical design parameter ranges of the current design rule are extracted from the expert experience knowledge base, the use frequency of each parameter value is calculated, the historical design parameter ranges are modified through the use frequency, and the design parameter ranges of the personalized medical instrument are obtained.

[0099] The application provides a personalized medical instrument design parameter range determination method based on morphological characteristic parameters and a multi-path reasoning algorithm.

[0100] In actual application, morphological characteristic parameters of a patient are acquired, and taking personalized knee joint prosthesis design as an example, morphological characteristic parameters of the patient include distal femur width 45.6 mm, femoral intercondylar width 22.3 mm, tibial plateau width 70.2 mm, knee joint flexion angle 135 degrees and other key parameters, which are obtained by medical imaging equipment and used as basic data for subsequent design rule derivation.

[0101] A multi-path reasoning algorithm is used to extract suitable design rules from a medical instrument design rule library according to the acquired morphological characteristic parameters. The medical instrument design rule library stores a large number of design rules, such as "if the distal femur width is within the range of 40-50 mm, the prosthesis intercondylar width should be within the range of 20-25 mm", "if the knee joint flexion angle of the patient is greater than 130 degrees, the prosthesis posterior condyle height should be controlled within the range of 10-12 mm" and other rules. These rules are established based on a large amount of clinical practice and research data and have high reliability.

[0102] In the rule extraction process, a rule path tree needs to be constructed. Taking knee joint prosthesis design as an example, the root node of the rule path tree can be "knee joint prosthesis basic design rule", the first layer branch nodes include "femur part design rule", "tibia part design rule" and "patella part design rule", and under the "femur part design rule", it can be further divided into "femur size rule", "femur curvature rule", "femur cutting amount rule" and other finer granularity rule nodes. The rule correlation degree between nodes is pre-set according to the correlation strength in clinical practice, such as the correlation degree between "femur size rule" and "tibia size rule" is 0.85, indicating that the two types of rules are highly related.

[0103] After the morphological characteristic parameters of the patient are reconstructed into a feature vector, the root node of the rule path tree is traversed, and taking the distal femur width 45.6 mm of the patient as an example, when the "femur size rule" node is accessed, the effectiveness score of the feature under the rule constraint is calculated. The specific calculation method is to quantify the matching degree of the feature value and the rule allowed range into a score. For example, the distal femur width 45.6 mm of the patient completely falls within the range of 40-50 mm defined by the rule, so the effectiveness score is 1.0. If a feature value only partially meets the rule constraint, the score will be correspondingly reduced.

[0104] During the traversal process, the next layer of nodes to be accessed is determined according to the validity score of the current node, and when the rule path search depth reaches the preset 5 layers or the validity score of a certain node is lower than the set threshold of 0.6, the search of the branch will stop, in this way, multiple candidate rule paths are obtained, such as the validity score of the path "knee prosthesis basic design rule→femur part design rule→femur size rule→femur front and back size constraint→femur component thickness constraint" is 0.92.

[0105] For all candidate rule paths, the integrity score is calculated based on the rule correlation degree between the nodes in each branch, and the integrity score reflects the comprehensiveness of the rule path in covering the design requirements. The calculation method is to multiply the weighted average value of the correlation degree between all nodes in the path with the proportion of the characteristic parameters covered by the path. For example, path 1 covers 80% of the key characteristic parameters, and the average correlation degree between the nodes is 0.85, so the integrity score is 0.68; while path 2 covers 75% of the characteristic parameters, but the average correlation degree between the nodes is 0.92, so the integrity score is 0.69, therefore, path 2 is selected as the current design rule.

[0106] After determining the current design rule, the historical design parameter range corresponding to the rule is extracted from the expert experience knowledge base. The expert experience knowledge base stores a large number of design parameters of successful cases, such as "for patients with femur distal width of 45-46mm, the prosthesis medial-lateral condylar distance is usually set to 21-23mm in historical cases", these historical design parameters are extracted, and the usage frequency of each parameter value is counted, for example, for patients with similar morphological characteristics, the usage frequency of the prosthesis medial-lateral condylar distance value of 21.5mm is 15%, the usage frequency of the value of 22.0mm is 35%, the usage frequency of the value of 22.5mm is 40%, and the usage frequency of the value of 23.0mm is 10%.

[0107] Based on the usage frequency, the historical design parameter range is corrected, and the parameter values with high usage frequency will obtain higher weights, so that the corrected parameter range is more inclined to the commonly used values in historical successful cases. For this patient, the final design parameter range of the personalized knee prosthesis medial-lateral condylar distance is 21.8-22.7mm, which takes into account the morphological characteristics of the patient, the design rule constraints and the historical successful experience.

[0108] Through the above method, the system can provide personalized medical device design parameter ranges for patients with different morphological characteristics, and improve the adaptability and treatment effect of medical devices.

[0109] Figure 2An illustration of a personalized medical device design parameter optimization process. In an optional embodiment, based on the mechanical properties and structural variables of the design parameter range, a multi-objective parameter optimization function is constructed, and a particle swarm optimization algorithm is used to solve it to obtain the optimal design parameters of the personalized medical device, including:

[0110] The stress distribution of the medical device under physiological load is calculated, and the maximum stress, maximum strain and total strain energy are extracted as the mechanical optimization variables; the compressive strength, friction coefficient and deformation variable of the surface of the medical device are calculated as the structural optimization variables, and the correlation coefficients between the optimization variables are calculated to construct a multi-objective parameter optimization function;

[0111] The design parameter range is mapped to the particle swarm search space, and the normalized coding rule is set for each design parameter. The particle position vector is encoded as a design parameter combination, and the multi-objective optimization function value of each design parameter combination is calculated as the particle fitness. The fitness of the initial population particles is evaluated, and the particle position with the highest fitness is selected as the global optimal solution;

[0112] In the iterative optimization process, the adaptive inertia weight is set based on the Euclidean distance between the particle and the global optimal solution, the particle velocity is updated according to the inertia weight, and the velocity size is limited to ensure that the particle moves within the effective search space; the fitness of the updated particle position is calculated, and the global optimal solution is updated. When the improvement amplitude of the global optimal solution of continuous preset number of iterations is less than the convergence threshold, the iteration is stopped, and the global optimal solution is decoded as the optimal design parameters of the personalized medical device.

[0113] In the present embodiment, a finite element model of the medical device needs to be established, a three-dimensional structure of the medical device is constructed based on patient image data, and a design parameter range to be optimized is determined. The construction of the three-dimensional structure of the medical device needs to combine patient-specific data with standardized medical device design templates. In the present embodiment, based on the anatomical structure characteristics of the proximal tibia of the patient, a parameterized design method is used to construct a tibial platform prosthesis model. Specifically, based on the contour shape of the patient's tibial platform, a uniform fitting gap of 2 mm is set to construct the basic shape of the prosthesis. Considering the mechanical properties and biocompatibility requirements of the prosthesis, a composite structure including a solid support structure and a porous structure is designed. The solid support structure is located at the load-bearing part to ensure the overall stiffness of the prosthesis. The porous structure is distributed on the surface in contact with the bone tissue to promote bone integration. Based on the previous clinical research data, the design parameter range to be optimized is determined: the wall thickness range is set to 1.2 mm to 2.5 mm, the porous structure pore size range is set to 0.5 mm to 2.0 mm, and the material is titanium alloy with an elastic modulus range of 60 GPa to 120 GPa.

[0114] The generation of finite element mesh is crucial to the accuracy of the calculation results. For complex medical device structures, tetrahedral elements are used for meshing. In this embodiment, the tibial platform prosthesis model is meshed, with the maximum mesh size set to 0.5 mm and the minimum mesh size set to 0.1 mm. Local refinement is performed in the stress concentration areas of the structure, such as the connection between the porous structure and the solid structure, and the area with varying wall thickness. The generated finite element model contains about 500,000 elements. Element quality inspection shows that 99.5% of the elements meet the quality requirement of a Jacobian ratio greater than 0.7.

[0115] The assignment of material properties is the basis of finite element analysis. In this embodiment, for titanium alloy materials, the density is set to 4.5 g / cm³, the elastic modulus is set to vary in the range of 60 GPa to 120 GPa, the Poisson's ratio is set to 0.33, and the yield strength is set to 880 MPa. For bone tissue, a density-elastic modulus mapping relationship is established based on CT values to convert the patient-specific bone density distribution into the corresponding mechanical property distribution. The elastic modulus of cortical bone is about 15 GPa, and the elastic modulus of cancellous bone is about 0.5 GPa.

[0116] The setting of boundary conditions and load conditions is based on physiological mechanics analysis. According to biomechanical research data, in the standing state, the tibial plateau bears about 3 times the body weight. For a patient weighing 70 kg, a vertical load of 2100 N is applied to the tibial plateau, which is transmitted through the knee joint contact surface with a contact area of about 1200 mm². In the finite element model, the distal tibia is fixed to simulate the support provided by the bone shaft. A distributed load is applied to the tibial plateau, with the load distribution determined based on the results of knee joint contact analysis. The medial side bears about 60% of the load, and the lateral side bears about 40% of the load.

[0117] Under the above conditions, finite element analysis is performed to calculate the stress distribution of the medical device under physiological load. The analysis results show that for the design scheme with a wall thickness of 2.0 mm, a pore diameter of 1.0 mm, and an elastic modulus of 100 GPa, the maximum von Mises stress is 320 MPa, located in the transition area between the porous structure and the solid structure. The maximum strain is 0.3%, located at the small connection of the porous structure. The total strain energy is 0.85 J, mainly distributed in the porous structure area. These mechanical indicators constitute one of the target variables for optimization.

[0118] According to the finite element analysis results, the compressive strength, friction coefficient and deformation of the surface of the medical device are extracted as the structure optimization variables. In this embodiment, the compressive strength of the tibial platform prosthesis is obtained by stress analysis of the key cross section, which is 430 MPa; the surface friction coefficient is determined by the material properties and surface treatment process, which is 0.35; and the maximum deformation is 0.18 mm, which is located in the area with the largest load. These structure indexes together with the aforementioned mechanical indexes constitute the variable set of the multi-objective optimization function.

[0119] The correlation between variables needs to be considered in the construction of the multi-objective parameter optimization function, and the multi-objective parameter optimization function F is constructed based on correlation analysis. Through correlation analysis, it is found that the correlation coefficient between wall thickness and maximum stress is -0.85, indicating that increasing the wall thickness will significantly reduce the maximum stress; the correlation coefficient between the pore size of the porous structure and the compressive strength is -0.72, indicating that increasing the pore size will reduce the compressive strength; the correlation coefficient between the elastic modulus and the total strain energy is -0.68, indicating that increasing the elastic modulus will reduce the energy absorption. According to these correlation data, a comprehensive evaluation function containing mechanical properties and structural characteristics is constructed. The function is the weighted sum of the ratio of the maximum stress to the material yield strength, the ratio of the maximum deformation to the allowable deformation, and the ratio of the compressive strength to the target strength. The weight coefficients are determined by expert scoring method, which are 0.4, 0.3 and 0.3 respectively. The optimization goal is to minimize the value of F, that is, to find the best balance point of mechanical properties and structural properties.

[0120] In the implementation process of the particle swarm optimization algorithm, the design parameter range is mapped to the search space of the particle swarm. For each design parameter, normalized encoding is performed to map its value to the range of 0-1. For example, for the wall thickness parameter, the encoding formula is: normalized value = (actual wall thickness value - 1.2) / (2.5 - 1.2). When initializing the particle swarm, 50 particles are randomly generated, each representing a combination of design parameters. For example, the position vector of a particle [0.4, 0.6, 0.3] corresponds to the actual design parameters after decoding: wall thickness 1.72 mm, pore size 1.4 mm, and elastic modulus 78 GPa.

[0121] The fitness of each particle in the initial population is evaluated by calculating the mechanical properties and structural variables of the corresponding design parameter combination through finite element analysis, and then substituting them into the multi-objective optimization function to calculate the fitness value. For example, the particle with the highest fitness in the initial population has a fitness value of 0.68, and the corresponding design parameters are wall thickness 1.8 mm, pore size 1.2 mm, and elastic modulus 85 GPa. This set of parameters is set as the initial global optimal solution.

[0122] In the iterative optimization process, the adaptive inertia weight strategy is adopted to improve the convergence efficiency of the algorithm, the Euclidean distance of each particle and the global optimal solution is calculated, the inertia weight is larger when the distance is farther, which promotes the global search of the particle, and the inertia weight is smaller when the distance is closer, which is beneficial to the local fine search.For example, when the normalized distance of the particle and the global optimal solution is 0.5, the inertia weight is set to 0.8; when the distance is 0.1, the inertia weight is reduced to 0.4.

[0123] According to the calculated inertia weight, the particle velocity is updated, the velocity update considers the influence of the current particle velocity, the individual historical optimal position and the global optimal position, the individual cognitive coefficient is set to 2.0, and the social learning coefficient is set to 2.0, in order to ensure that the particle moves in the effective search space, the velocity is limited to be less than 20% of the search range in each dimension, for example, the upper limit of the velocity of the wall thickness parameter is (2.5-1.2) * 20% = 0.26mm.

[0124] Based on the updated velocity, the new position of the particle is calculated, and the boundary check is performed to ensure that all design parameters are in the effective range, the fitness of the new position is evaluated, the individual historical optimal position and the global optimal position are updated, and the convergence condition is set to stop iteration when the global optimal solution improvement amplitude is less than 0.001 for 15 consecutive iterations.

[0125] After 87 iterations, the algorithm converges, and the global optimal solution fitness value is 0.42, and the optimal design parameters after decoding are: wall thickness 2.1mm, hole diameter 0.8mm, elastic modulus 92GPa, through finite element verification analysis, the maximum stress of the medical instrument under the group of parameters is 145MPa, which is lower than 72.5% of the material yield strength 200MPa; the maximum strain is 0.0018, and the total strain energy is 0.37J; the compressive strength reaches 188MPa, the friction coefficient is 0.08, and the maximum deformation is 0.25mm, and each index meets the clinical use requirements.

[0126] The embodiment of the application is based on a three-dimensional reconstruction-based personalized matching design system for medical instruments, which comprises:

[0127] The first unit is used for acquiring a medical image sequence of a target object to be matched;

[0128] The second unit is used for extracting structural feature points and boundary feature points of the medical image sequence, and calculating feature vectors of the feature points; based on the similarity and spatial consistency of the feature vectors, an image transformation matrix is calculated, each medical image is projected into a reference image coordinate system and regionally fused, and a spatially aligned medical image sequence is obtained;

[0129] The third unit is configured to extract edge features of the spatially aligned medical image sequence, determine an initial segmentation point, perform region growing segmentation, fit an initial reconstruction surface using a non-uniform rational B-spline, extract texture features and perform feature mapping, obtain a three-dimensional reconstruction model of a to-be-matched part of the target object, and extract morphological feature parameters of the to-be-matched part.

[0130] The fourth unit is configured to extract design rules from a medical instrument design rule library using a multi-path reasoning algorithm based on the morphological feature parameters, and obtain a design parameter range of the individualized medical instrument in combination with an expert experience knowledge base.

[0131] The fifth unit is configured to construct a multi-objective parameter optimization function based on mechanical properties and structural variables of the design parameter range, and obtain optimal design parameters of the individualized medical instrument by solving the function using a particle swarm optimization algorithm.

[0132] The third aspect of the embodiment of the present application provides an electronic device, comprising:

[0133] a processor;

[0134] a memory for storing processor-executable instructions;

[0135] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0136] The fourth aspect of the embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon, and the computer program instructions are executed by a processor to implement the method described above.

[0137] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which are used to perform various aspects of the present application.

[0138] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for personalized matching design of a medical instrument based on three-dimensional reconstruction, characterized in that, The method comprises the following steps: obtaining a medical image sequence of a target object part to be matched; extracting structural feature points and boundary feature points of the medical image sequence, and calculating feature vectors of the feature points; calculating an image transformation matrix based on the similarity and spatial consistency of the feature vectors, projecting each medical image to a reference image coordinate system and performing region fusion to obtain a spatially aligned medical image sequence; extracting edge features of the spatially aligned medical image sequence, determining initial segmentation points and performing region growing segmentation, fitting an initial reconstruction surface using non-uniform rational B-splines, extracting texture features and performing feature mapping to obtain a three-dimensional reconstruction model of the target object part to be matched, and extracting morphological feature parameters of the part to be matched; based on the morphological feature parameters, extracting design rules from a medical instrument design rule library using a multi-path reasoning algorithm, and combining an expert experience knowledge base to obtain a design parameter range of the individualized medical instrument; based on the mechanical properties and structural variables of the design parameter range, constructing a multi-objective parameter optimization function, and solving it using a particle swarm optimization algorithm to obtain the optimal design parameters of the individualized medical instrument, including: calculating the stress distribution of the medical instrument under physiological load, and extracting the maximum stress, maximum strain and total strain energy as mechanical optimization variables; calculating the compressive strength, friction coefficient and deformation variable of the surface of the medical instrument as structural optimization variables, and calculating the correlation coefficients between the optimization variables to construct a multi-objective parameter optimization function; mapping the design parameter range to a particle swarm search space, setting a normalization coding rule for each design parameter, encoding the particle position vector as a design parameter combination, and calculating the multi-objective optimization function value of each design parameter combination as the particle fitness, evaluating the fitness of the initial population particles, and selecting the particle position with the highest fitness as the global optimal solution; in the iterative optimization process, setting an adaptive inertia weight based on the Euclidean distance between the particle and the global optimal solution, updating the particle velocity according to the inertia weight, and limiting the velocity to ensure that the particle moves within the effective search space; calculating the fitness of the updated particle position and updating the global optimal solution, and stopping iteration when the improvement amplitude of the global optimal solution for a continuous preset number of iterations is less than a convergence threshold, and decoding the global optimal solution into the optimal design parameters of the individualized medical instrument.

2. The method of claim 1, wherein, extracting structural feature points and boundary feature points of the medical image sequence, and calculating feature vectors of the feature points; calculating an image transformation matrix based on the similarity and spatial consistency of the feature vectors, projecting each medical image to a reference image coordinate system and performing region fusion to obtain a spatially aligned medical image sequence, including: performing multi-scale decomposition on the medical image sequence using a Gaussian difference operator, calculating the gradient direction distribution histogram of each medical image, detecting the inflection points of the edge profile of the medical image as boundary feature points; calculating the local gray level of each pixel point of the medical image, generating a feature response map and performing non-maximum suppression, and taking the local maximum points as structural feature points; statistically analyzing the gradient direction distribution in the neighborhood of the boundary feature points and the structural feature points to generate feature vectors; The similarity of the feature vectors between adjacent medical images is calculated to determine initial matching point pairs, and motion continuity and spatial distribution consistency of the initial matching point pairs are calculated, and abnormal matching point pairs are removed, and an image transformation matrix between adjacent images is calculated according to the similarity and spatial position relationship of the remaining matching point pairs; Based on the image transformation matrix, a global transformation matrix of each medical image relative to a reference image is calculated, the medical image sequence is projected into a reference image coordinate system according to the global transformation matrix, a weighted fusion is performed on the overlapping area, an interpolation filling is performed on the vacancy area, and a spatially aligned medical image sequence is generated.

3. The method of claim 1, wherein, Edge features of the spatially aligned medical image sequence are extracted, initial segmentation points are determined, and region growing segmentation is performed, an initial reconstruction surface is fitted using a non-uniform rational B-spline, texture features are extracted and feature mapping is performed, a three-dimensional reconstruction model of the target object matching part is obtained, and morphological feature parameters of the matching part are extracted, including: Edge features of the spatially aligned medical image sequence are extracted using multi-scale convolution, a tissue boundary contour map is generated, and a tissue density distribution map is obtained by calculating the gray gradient value of each pixel point; Based on the tissue boundary contour map, initial segmentation points are determined, and region growing is performed with gray gradient similarity as the expansion criterion, and adjacent pixel points with a gray gradient difference less than the average gradient of the tissue surface are marked as the same structure region; The boundary points of the structure region are taken as control points, a three-dimensional surface is fitted based on a non-uniform rational B-spline method, and an initial reconstruction surface of the target object matching part is obtained; Local texture features in each structure region are extracted, a matching score of the texture features and the initial reconstruction surface vertex is calculated, the texture features are mapped to the corresponding surface position according to the matching score, and for the position where the texture features jump, a smooth transition is realized by weighted superposition of local region texture features, a three-dimensional reconstruction model of the target object matching part is obtained, and morphological feature parameters of the matching part are extracted from the three-dimensional reconstruction model.

4. The method of claim 3, wherein, The boundary points of the structure region are taken as control points, a three-dimensional surface is fitted based on a non-uniform rational B-spline method, and an initial reconstruction surface of the target object matching part is obtained, including: The local curvature of the control points is calculated by taking the boundary point coordinates of the structure region as control point coordinates; The control point coordinates are mapped to a parameter space, a B-spline basis function of a specified order is constructed, a weight coefficient is calculated based on the local curvature, and a rational B-spline basis function is constructed by combining the weight coefficient and the B-spline basis function; The control point coordinates are substituted into the rational B-spline basis function to obtain a non-uniform rational B-spline surface, and the fitting error between the non-uniform rational B-spline surface and the original boundary points is calculated, and when the relative error of the structure region is greater than a preset error threshold, the control points of the structure region are gradient optimized and adjusted to obtain an initial reconstruction surface of the target object matching part.

5. The method of claim 1, wherein, Based on the morphological feature parameters, a multi-path reasoning algorithm is used to extract design rules from a medical instrument design rule library, and combined with an expert experience knowledge base, design parameter ranges of individualized medical instruments are obtained, including: The design rules in the medical device design rule library are taken as branch nodes, and a rule path tree is constructed through the rule correlation degrees between the nodes; The morphological feature parameters are reconstructed into a feature vector, the effectiveness score of the feature vector under the rule constraints corresponding to each branch node is calculated from the root node of the rule path tree, and the next layer of to-be-accessed nodes is determined; When the rule path search depth exceeds a preset layer number or the effectiveness score is lower than a rule screening threshold, the search of the branch is stopped, and multiple candidate rule paths are obtained; Based on the rule correlation degrees between the branch nodes, the completeness scores of the candidate rule paths are calculated, the candidate rule path with the highest completeness score is selected as the current design rule, the historical design parameter range of the current design rule is extracted from the expert experience knowledge base, the use frequency of each parameter value is calculated, the historical design parameter range is modified through the use frequency, and the design parameter range of the individualized medical device is obtained.

6. A medical instrument individualized matching design system based on three-dimensional reconstruction, for implementing the method according to any one of claims 1-5, characterized in that, Comprise: A first unit configured to acquire a medical image sequence of a target object to-be-matched part; A second unit configured to extract structural feature points and boundary feature points of the medical image sequence and calculate feature vectors of the feature points; Based on the similarity and spatial consistency of the feature vectors, an image transformation matrix is calculated, each medical image is projected to a reference image coordinate system and regionally fused, and a spatially aligned medical image sequence is obtained; A third unit configured to extract edge features of the spatially aligned medical image sequence, determine initial segmentation points and perform region growing segmentation, fit an initial reconstruction surface using a non-uniform rational B-spline, extract texture features and perform feature mapping, obtain a three-dimensional reconstruction model of the target object to-be-matched part, and extract morphological feature parameters of the to-be-matched part; A fourth unit configured to extract design rules from a medical device design rule library based on the morphological feature parameters, and obtain a design parameter range of an individualized medical device in combination with an expert experience knowledge base; A fifth unit configured to construct a multi-objective parameter optimization function based on the mechanical properties and structural variables of the design parameter range, and solve the function using a particle swarm optimization algorithm to obtain optimal design parameters of the individualized medical device.

7. An electronic device, comprising: Comprise: A processor; A memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 5.

Images