Vascular stenosis three-dimensional reconstruction method based on random field depth modeling

Abstract

The invention discloses a vascular stenosis three-dimensional reconstruction method based on random field depth modeling. The method comprises the following steps: S1, obtaining processed two-dimensional medical image data; s2, generating a corresponding blood vessel structure feature map; s3, constructing a random field model comprising a local energy term and a neighborhood smooth constraint energy term based on the blood vessel structure feature map, and modeling a target function in a blood vessel three-dimensional reconstruction process; s4, introducing a gold washing optimization algorithm to optimize the random field model, and performing iterative solution on the target function of the random field model through global search and local fine search by the gold washing optimization algorithm to obtain a final optimal solution; s5, reconstructing the three-dimensional surface grid data by using the final optimal solution, and automatically marking the reconstructed three-dimensional surface grid data; and S6, generating a blood vessel three-dimensional reconstruction model. According to the method, the global and locality performance of the model can be balanced more effectively, and the total calculation time of three-dimensional reconstruction is shortened by about 20%.

Description

Technical Field

[0001] The present invention relates to the field of vascular technology, and in particular to a three-dimensional reconstruction method of vascular stenosis based on random field depth modeling. Background Art

[0002] In recent years, with the continuous advancement of medical imaging technology, the diagnosis and treatment of vascular-related diseases have gradually increased the demand for accurate analysis of imaging data. Clinically, the detection and evaluation of vascular stenosis has an important impact on the treatment and prognosis of cardiovascular and cerebrovascular diseases, peripheral vascular diseases, etc. Therefore, how to perform high-precision three-dimensional reconstruction of vascular stenosis has become a key task in the field of medical image processing.

[0003] In the field of 3D reconstruction of blood vessels, existing technologies usually perform image segmentation, registration and simple 3D reconstruction based on 2D tomographic images. Some methods use traditional image processing algorithms based on thresholds or region growing to reconstruct the rough shape of the blood vessels. Some studies have also introduced basic machine learning algorithms and used manually designed features to realize the recognition and segmentation of vascular regions. Existing methods can provide a certain degree of reconstruction capability when the vascular structure is relatively simple or the degree of lesions is relatively mild. However, in vascular bifurcation areas, stenosis sites and lesion areas with complex morphology, the reconstruction accuracy is often insufficient, which can easily lead to reconstruction breaks or misjudgment of stenosis.

[0004] In practical applications, the existing technologies mainly have the following defects: First, the traditional reconstruction methods based on thresholds or rules are sensitive to vascular image noise and grayscale changes, and are not robust enough when facing image data of different quality and contrast; Second, the structures of vascular bifurcations and stenosis areas are complex, and it is difficult to obtain high-precision three-dimensional reconstruction results by relying solely on manual or simple algorithms, which is prone to missed detection or false detection; Third, there is a lack of common modeling methods for the overall connectivity and local details of the blood vessels, so the geometric details of the stenosis segments and lesions are often ignored, and it is difficult to form an accurate model that can be used for clinical decision-making; Fourth, the degree of automation is not high, and the processing process requires a lot of manual intervention or experience adjustment, resulting in low overall efficiency, which is difficult to meet the needs of modern medicine for rapid and accurate diagnosis. The existing deficiencies have formed a bottleneck in the diagnosis and treatment of vascular stenosis, and it is urgent to improve and enhance the three-dimensional modeling, global optimization and automatic labeling. Summary of the invention

[0005] One purpose of the present invention is to propose a three-dimensional reconstruction method for vascular stenosis based on random field deep modeling. The present invention can more effectively balance the global and local performance of the model, and the total calculation time of the three-dimensional reconstruction is shortened by about 20%.

[0006] A method for three-dimensional reconstruction of vascular stenosis based on random field depth modeling according to an embodiment of the present invention comprises the following steps:

[0007] S1. Acquire two-dimensional medical image data, perform image denoising, image grayscale normalization and image contrast enhancement on the two-dimensional medical image data, and obtain processed two-dimensional medical image data;

[0008] S2. Using a deep learning model to perform local feature extraction and global feature extraction on the processed two-dimensional medical image data to generate a corresponding vascular structure feature map;

[0009] S3. Based on the vascular structure feature map, a random field model including a local energy term and a neighborhood smoothness constraint energy term is constructed to model the objective function in the vascular 3D reconstruction process;

[0010] S4. The gold panning optimization algorithm is introduced to optimize the random field model. The gold panning optimization algorithm iteratively solves the objective function of the random field model through global search and local fine search to obtain the final optimization solution;

[0011] S5. reconstructing the three-dimensional surface mesh data using the final optimization solution, and automatically marking the reconstructed three-dimensional surface mesh data;

[0012] S6. Generate a 3D reconstruction model of blood vessels, which includes a 3D representation of the complete vascular structure, vascular bifurcation regions, and vascular stenosis regions.

[0013] Optionally, S1 includes the following specific steps:

[0014] S11. Acquisition of two-dimensional medical imaging data I raw (x, y), and perform format unification processing on the two-dimensional medical image data, converting all the two-dimensional medical image data into standardized grayscale images;

[0015] S12. Using a non-local mean filtering method to perform image denoising on the standardized grayscale image, retaining edge information, and obtaining denoised two-dimensional medical image data;

[0016] S13. Performing grayscale normalization processing on the denoised two-dimensional medical image data, mapping the pixel values ​​to the range of [0, 1], and obtaining normalized two-dimensional medical image data;

[0017] S14. Perform image contrast enhancement processing on the normalized two-dimensional medical image data using a histogram equalization method to obtain contrast-enhanced two-dimensional medical image data, whose grayscale distribution satisfies the following conditions:

[0018]

[0019] Where H'(g) is the cumulative distribution of gray value g in the enhanced image, H(g') is the number of pixels with gray value g' in the original image, M and N are the number of rows and columns of the image respectively;

[0020] S15. The two-dimensional medical image data I after image denoising, grayscale normalization and contrast enhancement processing processed (x,y) are saved in a unified format.

[0021] Optionally, S2 includes the following specific steps:

[0022] S21. Using 2D medical imaging data I processed (x, y) is input into the deep learning model to perform multi-scale block processing on the two-dimensional medical image data, and the two-dimensional medical image data is divided into a set of image blocks according to the multi-scale window. Where P k (x, y) represents the image block at the kth scale, and n represents the total number of multi-scale divisions;

[0023] S22. On the basis of multi-scale division, a multi-path deep learning network based on an adaptive feature selection mechanism is constructed, wherein the multi-path deep learning network includes the following modules:

[0024] The local feature extraction module extracts the image blocks P at each scale. k (x,y) is input into the deep convolutional network to extract the local features of blood vessels F local,k (x,y), local features of blood vessels represent small-scale blood vessel edges, textures, and bifurcation areas;

[0025] The global feature extraction module fuses local features of all scales through a multi-path parallel network to generate a global feature F global (x,y), global features represent the overall morphology, bifurcation characteristics and potential stenosis areas of the blood vessels;

[0026] Adaptive feature weight allocation module, using attention mechanism to generate weight w k (x, y), for each scale feature blood vessel local feature F local,k (x,y) and the global feature F global (x,y) for weighted fusion:

[0027]

[0028] Among them, F adaptive (x,y) represents the adaptive fusion feature, w k (x,y) and w global (x, y) are the weights of local features and global features respectively;

[0029] S23. Adaptive fusion feature F adaptive (x,y) introduces a feature residual optimization mechanism, constructs a feature correction module, and uses the residual between the original feature and the reconstructed feature to correct the feature representation:

[0030] F final (x,y)=F adaptive (x,y)+γ(F global (x,y)-F adaptive (x,y));

[0031] Among them, F final (x, y) represents the optimized feature, and γ is the correction coefficient, which is used to control the amplitude of residual correction;

[0032] S24. Using the optimized feature F final (x,y) Generate vascular structure feature map T vascular (x, y), the vascular structure feature map includes the overall vascular structure features, vascular bifurcation area features and vascular stenosis area features:

[0033] T vascular (x,y)=δ1F final (x,y)+δ2F global (x,y)

[0034] Among them, δ1 and δ2 are feature fusion coefficients, which are used to enhance the weight ratio of optimized features and global features respectively.

[0035] Optionally, S3 includes the following specific steps:

[0036] S31. Based on the vascular structure feature map T vascular (x,y) defines the set of pixel states of the random field model:

[0037] X={x (x,y) ∣(x,y)∈Ω};

[0038] Among them, x (x,y) represents the pixel state with coordinates (x, y) in the vascular structure feature map, and Ω represents the coordinate domain of the image plane;

[0039] S32. Introducing adaptive penalty coefficients and offsets to construct the local energy term U(x (x,y) ), used to measure the pixel state x (x,y) Figure 1. Vascular structure characteristics vascular The degree of match between (x,y):

[0040] U(x (x,y) )=κ(|x (x,y) -T vascular(x,y)|+θ) η ;

[0041] Among them, κ is the overall weight factor, which is used to balance the impact of the difference between local pixels and feature maps, θ is the offset, and η is the nonlinear index;

[0042] S33. Constructing the neighborhood smoothing constraint energy term V(x (x,y) ,x (u,v) ), which is used to constrain the state consistency between the pixels in the blood vessel neighborhood and introduce the similarity of blood vessel features to dynamically adjust the smoothness:

[0043] V(x (x,y) ,x (u,v) )=α (x,y),(u,v) exp(-β∥T vascular (x,y)-T vascular (u,v)∥)·(x (x,y) -x (u,v) ) 2 ;

[0044] Among them, α (x,y),(u,v) is the basic smoothing coefficient, β is the feature similarity adjustment factor, and the difference between adjacent pixels in the vascular feature map is reflected by exponential decay. vascular (x,y)-T vascular (u,v)∥ represents the feature difference between pixel (x,y) and pixel (u,v) in the vascular structure feature map. (x,y) -x (u,v) ) 2 is the state difference term, which is used to measure the inconsistency of the state of the neighborhood pixels;

[0045] S34. The local energy term is combined with the neighborhood smoothness constraint energy term to construct the total energy function E(X) of the random field model to characterize the overall optimization goal in the 3D reconstruction of blood vessels:

[0046] E(X)=∑ (x,y)∈Ω ω (x,y) U(x (x,y) )+∑ (x,y)∈Ω ∑ (u,v)∈N(x,y) γ (x,y),(u,v) V(x (x,y) ,x (u,v) );

[0047] Among them, ω (x,y) is the local energy weight coefficient, which is used to assign differential weights to pixels at different locations to better focus on the stenosis site. N(x, y) represents the neighborhood set of pixel (x, y), and γ (x,y),(u,v)is the neighborhood smoothing weight coefficient, which is used to balance the proportion of local energy terms and smoothing constraint energy terms in 3D vascular modeling. The minimization result of E(X) is used to generate the pixel state distribution in the 3D vascular reconstruction process.

[0048] Optionally, S4 includes the following specific steps:

[0049] S41. Based on the total energy function E(X) of the random field model, using the vascular structure feature map T vascular (x,y) Initialize the candidate solution set:

[0050]

[0051] in, is the i-th initial solution, f init is the initialization function based on the vascular feature map, ξ i represents the random perturbation introduced to increase the diversity of candidate solutions, and n is the number of candidate solutions;

[0052] S42. In the candidate solution set C0, several low-energy solutions are selected by calculating the value of the total energy function E(X) to enter the next stage. The optimization strategy of global search is defined as:

[0053]

[0054] in, Represents the optimal solution in the global search phase and defines the screening rules:

[0055] C global ={X i ∈C0|E(X i )≤τ};

[0056] Among them, τ is the global search energy threshold, and the set of candidate solutions C is selected. global Solutions with lower energy values ​​are retained;

[0057] S43. The candidate solution set C obtained by global search global , using the optimization strategy of gradient descent combined with dynamic perturbation to perform local fine optimization in the vascular bifurcation area and stenosis area:

[0058]

[0059] in, is the candidate solution at the tth iteration, η1 is the dynamic step size factor, which is dynamically adjusted according to the energy gradient of the local search. represents the gradient of the energy function, δ t is a local disturbance term, simulating the random exploration mechanism in the gold mining optimization algorithm;

[0060] S44. During the local fine search process, the energy change trend of the candidate solution is evaluated in real time, and the step size and disturbance term are dynamically updated:

[0061] η t+1 =h(η t );

[0062] δ t+1 =g(δ t );

[0063] Among them, h(η t ) represents the step size adjustment function, which is used to increase the step size when the gradient change is less than the threshold, and to reduce the step size when the gradient change is greater than the threshold. t ) represents the disturbance adjustment function;

[0064] S45. Combine the results of global search and local fine search to select the solution corresponding to the minimum value of the total energy function E(X) as the final optimization solution:

[0065]

[0066] Among them, X optimal is the global optimal solution of the random field model.

[0067] Optionally, S5 includes the following specific steps:

[0068] S51. From the final optimization solution optimal Extract the optimal state of each pixel (x, y)

[0069] S52. The optimal state Convert to three-dimensional space coordinate point S (x,y) , construct the mapping function:

[0070]

[0071] Among them, Γ(·; ·) is the state-coordinate mapping function, which comprehensively utilizes the optimal state of pixels and the vascular structure feature map T vascular (x, y) information to generate the three-dimensional coordinates S (x,y) To preserve the overall structure of blood vessels, the spatial characteristics of the vascular bifurcation area and the vascular stenosis area;

[0072] S53. According to all pixel coordinate points S (x,y) Constructing 3D blood vessel surface mesh data M 3D :

[0073]

[0074] Where V = {S (x,y)} is the node set of the three-dimensional grid, E is the edge set of the three-dimensional grid, Λ(·,·) is the connectivity judgment function, which determines whether the nodes are spatially connected based on the neighborhood relationship of the two-dimensional image plane and the similarity of the optimal state, so as to describe the overall topological structure of the blood vessel, the bifurcation of the blood vessel and the location of the stenosis;

[0075] S54. For the three-dimensional surface mesh data M 3D Perform partition marking and sort nodes according to the optimal state The value characteristics of are divided into normal blood vessel segment, blood vessel bifurcation segment and blood vessel stenosis segment:

[0076]

[0077] Among them, C(·) is the classification function, Δ(·) represents the operation of feature extraction for the optimal state, and Γ norm ,Γ bif ,Γ sten It is the threshold or judgment range corresponding to the normal segment, bifurcation segment and stenosis segment, and is used to highlight the spatial distribution of the overall vascular structure, vascular bifurcation area and vascular stenosis area in the three-dimensional surface mesh data.

[0078] The beneficial effects of the present invention are:

[0079] (1) The present invention introduces a random field model, and through the joint modeling of local energy terms and neighborhood smoothing constraint energy terms, utilizes the local features and global topological information of vascular images, thereby overcoming the problem that traditional methods only rely on local features, resulting in broken or discontinuous reconstruction results. By applying the gold panning optimization algorithm in the global search and local fine search of the random field model, the consistency and accuracy of the vascular bifurcation area and the stenosis area are ensured, and the complex structure of the blood vessel can be reconstructed more accurately, showing higher robustness in low-contrast and noisy image data.

[0080] (2) In the process of constructing the three-dimensional surface mesh of the blood vessel, the present invention introduces the state-coordinate mapping function and the node classification function to realize the automatic partitioning and marking of the overall structure of the blood vessel, the bifurcation area and the stenosis area. Combining the optimized optimal solution with the global constraints of the random field model, the specific location, range and diameter change of the stenosis area can be efficiently identified and accurately marked.

[0081] (3) The present invention dynamically adjusts the search step size and perturbation mechanism through the gold panning optimization algorithm to ensure the global optimal solution while avoiding falling into the local optimum. The global search stage is used to quickly lock the main features of the vascular area, and the local fine search is used to dynamically optimize the stenosis and bifurcation areas. Combined with the dynamic step size adjustment and the introduction of the perturbation function, the efficiency and convergence speed of the optimization process are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0083] Figure 1 This is a flow chart of a method for three-dimensional reconstruction of vascular stenosis based on random field depth modeling proposed by the present invention. DETAILED DESCRIPTION

[0084] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0085] refer to Figure 1 A three-dimensional reconstruction method of vascular stenosis based on random field deep modeling includes the following steps:

[0086] S1. Acquire two-dimensional medical image data, perform image denoising, image grayscale normalization and image contrast enhancement on the two-dimensional medical image data, and obtain processed two-dimensional medical image data;

[0087] S2. Using a deep learning model to perform local feature extraction and global feature extraction on the processed two-dimensional medical image data to generate a corresponding vascular structure feature map;

[0088] S3. Based on the vascular structure feature map, a random field model including a local energy term and a neighborhood smoothness constraint energy term is constructed to model the objective function in the vascular 3D reconstruction process;

[0089] S4. The gold panning optimization algorithm is introduced to optimize the random field model. The gold panning optimization algorithm iteratively solves the objective function of the random field model through global search and local fine search to obtain the final optimization solution;

[0090] S5. reconstructing the three-dimensional surface mesh data using the final optimization solution, and automatically marking the reconstructed three-dimensional surface mesh data;

[0091] S6. Generate a 3D reconstruction model of blood vessels, which includes a 3D representation of the complete vascular structure, vascular bifurcation regions, and vascular stenosis regions.

[0092] In this implementation, S1 includes the following specific steps:

[0093] S11. Acquisition of two-dimensional medical imaging data I raw (x, y), and perform format unification processing on the two-dimensional medical image data, converting all the two-dimensional medical image data into standardized grayscale images;

[0094] S12. Using a non-local mean filtering method to perform image denoising on the standardized grayscale image, retaining edge information, and obtaining denoised two-dimensional medical image data;

[0095] S13. Performing grayscale normalization processing on the denoised two-dimensional medical image data, mapping the pixel values ​​to the range of [0, 1], and obtaining normalized two-dimensional medical image data;

[0096] S14. Perform image contrast enhancement processing on the normalized two-dimensional medical image data using a histogram equalization method to obtain contrast-enhanced two-dimensional medical image data, whose grayscale distribution satisfies the following conditions:

[0097]

[0098] Where H'(g) is the cumulative distribution of gray value g in the enhanced image, H(g') is the number of pixels with gray value g' in the original image, M and N are the number of rows and columns of the image respectively;

[0099] S15. The two-dimensional medical image data I after image denoising, grayscale normalization and contrast enhancement processing processed (x,y) are saved in a unified format.

[0100] In this implementation, S2 includes the following specific steps:

[0101] S21. Using 2D medical imaging data I processed (x, y) is input into the deep learning model to perform multi-scale block processing on the two-dimensional medical image data, and the two-dimensional medical image data is divided into a set of image blocks according to the multi-scale window. Where P k (x, y) represents the image block at the kth scale, and n represents the total number of multi-scale divisions;

[0102] S22. Based on multi-scale division, a multi-path deep learning network based on adaptive feature selection mechanism is constructed. The multi-path deep learning network includes the following modules:

[0103] The local feature extraction module extracts the image blocks P at each scale. k (x,y) is input into the deep convolutional network to extract the local features of blood vessels F local,k (x,y), local features of blood vessels represent small-scale blood vessel edges, textures, and bifurcation areas;

[0104] The global feature extraction module fuses local features of all scales through a multi-path parallel network to generate a global feature F global (x,y), global features represent the overall morphology, bifurcation characteristics and potential stenosis areas of the blood vessels;

[0105] Adaptive feature weight allocation module, using attention mechanism to generate weight w k (x, y), for each scale feature blood vessel local feature F local,k (x,y) and the global feature F global (x,y) for weighted fusion:

[0106]

[0107] Among them, F adaptive (x,y) represents the adaptive fusion feature, w k (x,y) and w global (x, y) are the weights of local features and global features respectively;

[0108] S23. Adaptive fusion feature F adaptive (x,y) introduces a feature residual optimization mechanism, constructs a feature correction module, and uses the residual between the original feature and the reconstructed feature to correct the feature representation:

[0109] F final (x,y)=F adaptive (x,y)+γ(F global (x,y)-F adaptive (x,y));

[0110] Among them, F final (x, y) represents the optimized feature, and γ is the correction coefficient, which is used to control the amplitude of residual correction;

[0111] S24. Using the optimized feature F final (x,y) Generate vascular structure feature map T vascular (x, y), the vascular structure feature map includes the overall vascular structure features, vascular bifurcation area features and vascular stenosis area features:

[0112] T vascular (x,y)=δ1F final (x,y)+δ2F global (x,y)

[0113] Among them, δ1 and δ2 are feature fusion coefficients, which are used to enhance the weight ratio of optimized features and global features respectively.

[0114] In this implementation, S3 includes the following specific steps:

[0115] S31. Based on the vascular structure feature map T vascular (x,y) defines the set of pixel states of the random field model:

[0116] X={x (x,y) ∣(x,y)∈Ω};

[0117] Among them, x (x,y) represents the pixel state with coordinates (x, y) in the vascular structure feature map, and Ω represents the coordinate domain of the image plane;

[0118] S32. Introducing adaptive penalty coefficients and offsets to construct the local energy term U(x (x,y) ), used to measure the pixel state x (x,y) Figure 1. Vascular structure characteristics vascular The degree of match between (x,y):

[0119] U(x (x,y) )=κ(|x (x,y) -T vascular (x,y)|+θ) η ;

[0120] Among them, κ is the overall weight factor, which is used to balance the impact of the difference between local pixels and feature maps, θ is the offset, and η is the nonlinear index;

[0121] S33. Constructing the neighborhood smoothing constraint energy term V(x (x,y) ,x (u,v) ), which is used to constrain the state consistency between the pixels in the blood vessel neighborhood and introduce the similarity of blood vessel features to dynamically adjust the smoothness:

[0122] V(x (x,y) ,x (u,v) )=α (x,y),(u,v) exp(-β∥T vascular (x,y)-T vascular (u,v)∥)·(x (x,y) -x (u,v) ) 2 ;

[0123] Among them, α (x,y),(u,v) is the basic smoothing coefficient, β is the feature similarity adjustment factor, and the difference between adjacent pixels in the vascular feature map is reflected by exponential decay. vascular (x,y)-T vascular (u,v)∥ represents the feature difference between pixel (x,y) and pixel (u,v) in the vascular structure feature map. (x,y) -x (u,v) ) 2 is the state difference term, which is used to measure the inconsistency of the state of the neighborhood pixels;

[0124] S34. The local energy term is combined with the neighborhood smoothness constraint energy term to construct the total energy function E(X) of the random field model to characterize the overall optimization goal in the 3D reconstruction of blood vessels:

[0125] E(X)=∑ (x,y)∈Ω ω (x,y) U(x (x,y) )+∑ (x,y)∈Ω ∑ (u,v)∈N(x,y) γ (x,y),(u,v) V(x (x,y) ,x (u,v) );

[0126] Among them, ω (x,y) is the local energy weight coefficient, which is used to assign differential weights to pixels at different locations to better focus on the stenosis site. N(x, y) represents the neighborhood set of pixel (x, y), and γ (x,y),(u,v) is the neighborhood smoothing weight coefficient, which is used to balance the proportion of local energy terms and smoothing constraint energy terms in 3D vascular modeling. The minimization result of E(X) is used to generate the pixel state distribution in the 3D vascular reconstruction process.

[0127] In this implementation, S4 includes the following specific steps:

[0128] S41. Based on the total energy function E(X) of the random field model, using the vascular structure feature map T vascular (x,y) Initialize the candidate solution set:

[0129]

[0130] in, is the i-th initial solution, f init is the initialization function based on the vascular feature map, ξ i represents the random perturbation introduced to increase the diversity of candidate solutions, and n is the number of candidate solutions;

[0131] S42. In the candidate solution set C0, several low-energy solutions are selected by calculating the value of the total energy function E(X) to enter the next stage. The optimization strategy of global search is defined as:

[0132]

[0133] in, Represents the optimal solution in the global search phase and defines the screening rules:

[0134] C global ={X i ∈C0|E(X i )≤τ};

[0135] Among them, τ is the global search energy threshold, and the set of candidate solutions C is selected. global Solutions with lower energy values ​​are retained;

[0136] S43. The candidate solution set C obtained by global search global , using the optimization strategy of gradient descent combined with dynamic perturbation to perform local fine optimization in the vascular bifurcation area and stenosis area:

[0137]

[0138] in, is the candidate solution at the tth iteration, η1 is the dynamic step size factor, which is dynamically adjusted according to the energy gradient of the local search. represents the gradient of the energy function, δ t is a local disturbance term, simulating the random exploration mechanism in the gold mining optimization algorithm;

[0139] S44. During the local fine search process, the energy change trend of the candidate solution is evaluated in real time, and the step size and disturbance term are dynamically updated:

[0140] η t+1 =h(η t );

[0141] δ t+1 =g(δ t );

[0142] Among them, h(η t ) represents the step size adjustment function, which is used to increase the step size when the gradient change is less than the threshold, and to reduce the step size when the gradient change is greater than the threshold. t ) represents the disturbance adjustment function;

[0143] S45. Combine the results of global search and local fine search to select the solution corresponding to the minimum value of the total energy function E(X) as the final optimization solution:

[0144]

[0145] Among them, X optimal is the global optimal solution of the random field model.

[0146] In this implementation, S5 includes the following specific steps:

[0147] S51. From the final optimization solution optimal Extract the optimal state of each pixel (x, y)

[0148] S52. The optimal state Convert to three-dimensional space coordinate point S (x,y) , construct the mapping function:

[0149]

[0150] Among them, Γ(·; ·) is the state-coordinate mapping function, which comprehensively utilizes the optimal state of pixels and the vascular structure feature map T vascular (x, y) information to generate the three-dimensional coordinates S (x,y) To preserve the overall structure of blood vessels, the spatial characteristics of the vascular bifurcation area and the vascular stenosis area;

[0151] S53. According to all pixel coordinate points S (x,y) Constructing 3D blood vessel surface mesh data M 3D :

[0152]

[0153] Where V = {S (x,y)} is the node set of the three-dimensional grid, E is the edge set of the three-dimensional grid, Λ(·,·) is the connectivity judgment function, which determines whether the nodes are spatially connected based on the neighborhood relationship of the two-dimensional image plane and the similarity of the optimal state, so as to describe the overall topological structure of the blood vessel, the bifurcation of the blood vessel and the location of the stenosis;

[0154] S54. For the three-dimensional surface mesh data M 3D Perform partition marking and sort nodes according to the optimal state The value characteristics of are divided into normal blood vessel segment, blood vessel bifurcation segment and blood vessel stenosis segment:

[0155]

[0156] Among them, C(·) is the classification function, Δ(·) represents the operation of feature extraction for the optimal state, and Γ norm ,Γ bif ,Γ sten It is the threshold or judgment range corresponding to the normal segment, bifurcation segment and stenosis segment, and is used to highlight the spatial distribution of the overall vascular structure, vascular bifurcation area and vascular stenosis area in the three-dimensional surface mesh data.

[0157] Embodiment 1:

[0158] Example In October 2024, a 67-year-old male patient received a physical examination at the Department of Cardiology of a Class III hospital in City A. The patient had a history of hypertension and diabetes and had frequent chest tightness and chest pain in the past two months. The hospital performed a coronary CT angiography on the patient and obtained a total of 700 image slices. The preliminary diagnosis suspected that the patient had severe stenosis of the left anterior descending artery. Three-dimensional reconstruction of the blood vessels is crucial for surgical planning and risk assessment.

[0159] After receiving the image data, the radiologist started the 3D reconstruction system for vascular stenosis based on the method of the present invention. The system completed the preprocessing of the 2D image data within 5 minutes, which specifically included the following key steps: First, the system detected that the image noise level was high, and used non-local mean filtering to reduce the noise of the image, effectively eliminating more than 90% of the artifacts while maintaining the clarity of the blood vessel edges. The image grayscale was then normalized, and the grayscale value was mapped to the range of [0,1], which significantly improved the image contrast. Finally, the system used histogram equalization technology to further enhance the contrast between the blood vessels and background tissues, making the vascular structure more prominent.

[0160] After completing preprocessing, the system uses a deep learning model to extract features from the image data. For the left anterior descending branch image area, the model extracts the local texture features and global topological features of the blood vessels, and generates a high-resolution vascular structure feature map. The doctor observed that the system has a strong feature extraction capability for the bifurcation areas and stenosis segments of the blood vessels, especially near the calcified lesions. The feature map clearly shows the range of the lesions.

[0161] Subsequently, the system constructed an energy model based on random fields to model the three-dimensional structure of blood vessels. Through the gold panning optimization algorithm, the system first performed a global search and screened out a set of low-energy candidate solutions. Then, in response to the complex morphology of the stenotic segment, the system entered the local fine search stage, dynamically adjusted the step size in each iteration, and combined with the random perturbation mechanism to effectively avoid the local optimal problem. After 17 optimization iterations, the system finally determined the optimal solution for three-dimensional reconstruction.

[0162] After the three-dimensional reconstruction was completed, the system generated the three-dimensional surface mesh data of the blood vessel within 3 seconds and automatically marked the normal segment, bifurcation area and stenosis area of ​​the blood vessel. The doctor found that the diameter of the stenosis segment marked by the system was only 0.9 mm and the stenosis length was 6.3 mm. In order to verify the reliability of the results, the doctor used the traditional three-dimensional reconstruction method based on regional growth for a comparative test and found that the diameter of the stenosis segment generated by the traditional method was 1.2 mm, which had obvious errors.

[0163] In the system's 3D visualization interface, doctors can intuitively observe the overall structure of the left anterior descending branch, bifurcation flow direction, and the geometry of the stenosis. To further evaluate the system's performance, doctors and the imaging team conducted comparative experiments on 50 similar cases. The following are key comparative data:

[0164] index Method of the present invention Traditional methods Average 3D reconstruction time (minutes) 11.2 22.8 Stenosis diameter measurement error (mm) 0.12 0.32 3D model completeness score (out of 10) 9.6 7.3 Stenosis identification accuracy (%) 97.4 84.1

[0165] The above experimental results show that the method of the present invention is significantly superior to traditional methods in terms of efficiency, accuracy and robustness, especially in the reconstruction of bifurcation areas and severe stenosis segments, showing higher stability and reliability. Finally, through the high-precision three-dimensional vascular model generated by the method of the present invention, the doctor determined the specific morphology and range of the patient's left anterior descending branch stenosis, successfully planned the surgical path, and reduced intraoperative risks. The examples fully demonstrated the practical application value and technical advantages of the method of the present invention in the three-dimensional reconstruction of vascular stenosis.

[0166] The present invention introduces a random field model, and through the joint modeling of local energy terms and neighborhood smoothing constraint energy terms, utilizes the local features and global topological information of vascular images, thereby overcoming the problem that traditional methods only rely on local features, resulting in broken or discontinuous reconstruction results. The application of the gold panning optimization algorithm in the global search and local fine search of the random field model ensures the continuity and accuracy of vascular bifurcation areas and stenosis areas, and can more accurately reconstruct complex vascular structures, showing higher robustness in low-contrast and noisy image data.

[0167] In the process of constructing the three-dimensional surface mesh of the blood vessel, the present invention introduces the state-coordinate mapping function and the node classification function to realize the automatic partitioning and marking of the overall structure of the blood vessel, the bifurcation area and the stenosis area. Combining the optimized optimal solution with the global constraints of the random field model, the specific position, range and diameter change of the stenosis area can be efficiently identified and accurately marked.

[0168] The present invention dynamically adjusts the search step size and perturbation mechanism through the gold panning optimization algorithm to avoid falling into the local optimum while ensuring the global optimal solution. The global search stage is used to quickly lock the main features of the vascular area, and the local fine search dynamically optimizes the stenosis and bifurcation areas. Combined with the dynamic step size adjustment and the introduction of the perturbation function, the efficiency and convergence speed of the optimization process are improved.

[0169] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A three-dimensional reconstruction method for vascular stenosis based on random field deep modeling, characterized in that: The steps include: S1. Acquire two-dimensional medical image data, perform image denoising, image grayscale normalization and image contrast enhancement on the two-dimensional medical image data, and obtain processed two-dimensional medical image data; S2. Using a deep learning model to perform local feature extraction and global feature extraction on the processed two-dimensional medical image data to generate a corresponding vascular structure feature map; S3. Based on the vascular structure feature map, a random field model including a local energy term and a neighborhood smoothness constraint energy term is constructed to model the objective function in the vascular 3D reconstruction process; S4. The gold panning optimization algorithm is introduced to optimize the random field model. The gold panning optimization algorithm iteratively solves the objective function of the random field model through global search and local fine search to obtain the final optimization solution; S5. reconstructing the three-dimensional surface mesh data using the final optimization solution, and automatically marking the reconstructed three-dimensional surface mesh data; S6. Generate a 3D reconstruction model of blood vessels, which includes a 3D representation of the complete vascular structure, vascular bifurcation regions, and vascular stenosis regions.

2. The method for three-dimensional reconstruction of vascular stenosis based on random field deep modeling according to claim 1, characterized in that: The S1 comprises the following specific steps: S11. Acquisition of two-dimensional medical imaging data I raw (x, y), and perform format unification processing on the two-dimensional medical image data, converting all the two-dimensional medical image data into standardized grayscale images; S12. Using a non-local mean filtering method to perform image denoising on the standardized grayscale image, retaining edge information, and obtaining denoised two-dimensional medical image data; S13. Performing grayscale normalization processing on the denoised two-dimensional medical image data, mapping the pixel values ​​to the range of [0, 1], and obtaining normalized two-dimensional medical image data; S14. Perform image contrast enhancement processing on the normalized two-dimensional medical image data using a histogram equalization method to obtain contrast-enhanced two-dimensional medical image data, whose grayscale distribution satisfies the following conditions: Where H'(g) is the cumulative distribution of gray value g in the enhanced image, H(g') is the number of pixels with gray value g' in the original image, M and N are the number of rows and columns of the image respectively; S15. The two-dimensional medical image data I after image denoising, grayscale normalization and contrast enhancement processing processed (x,y) are saved in a unified format.

3. The method for three-dimensional reconstruction of vascular stenosis based on random field depth modeling according to claim 1, characterized in that: The S2 comprises the following specific steps: S21. Using 2D medical imaging data I processed (x, y) is input into the deep learning model to perform multi-scale block processing on the two-dimensional medical image data, and the two-dimensional medical image data is divided into a set of image blocks according to the multi-scale window. Where P k (x, y) represents the image block at the kth scale, and n represents the total number of multi-scale divisions; S22. On the basis of multi-scale division, a multi-path deep learning network based on an adaptive feature selection mechanism is constructed, wherein the multi-path deep learning network includes the following modules: The local feature extraction module extracts the image blocks P at each scale. k (x,y) is input into the deep convolutional network to extract the local features of blood vessels F local,k (x,y), local features of blood vessels represent small-scale blood vessel edges, textures, and bifurcation areas; The global feature extraction module fuses local features of all scales through a multi-path parallel network to generate a global feature F global (x,y), global features represent the overall morphology, bifurcation characteristics and potential stenosis areas of the blood vessels; Adaptive feature weight allocation module, using attention mechanism to generate weight w k (x, y), for each scale feature blood vessel local feature F local,k (x,y) and the global feature F global (x,y) for weighted fusion: Among them, F adaptive (x,y) represents the adaptive fusion feature, w k (x,y) and w global (x, y) are the weights of local features and global features respectively; S23. Adaptive fusion feature F adaptive (x,y) introduces a feature residual optimization mechanism, constructs a feature correction module, and uses the residual between the original feature and the reconstructed feature to correct the feature representation: F final (x,y)=F adaptive (x,y)+γ(F global (x,y)-F adaptive (x,y)); Among them, F final (x, y) represents the optimized feature, and γ is the correction coefficient, which is used to control the amplitude of residual correction; S24. Using the optimized feature F final (x,y) Generate vascular structure feature map T vascular (x, y), the vascular structure feature map includes the overall vascular structure features, vascular bifurcation area features and vascular stenosis area features: T vascular (x,y)=δ1F final (x,y)+δ2F global (x,y) Among them, δ1 and δ2 are feature fusion coefficients, which are used to enhance the weight ratio of optimized features and global features respectively.

4. The method for three-dimensional reconstruction of vascular stenosis based on random field deep modeling according to claim 1, characterized in that: The S3 includes the following specific steps: S31. Based on the vascular structure feature map T vascular (x,y) defines the set of pixel states of the random field model: X={x (x,y) ∣(x,y)∈Ω}; Among them, x (x,y) represents the pixel state with coordinates (x, y) in the vascular structure feature map, and Ω represents the coordinate domain of the image plane; S32. Introducing adaptive penalty coefficients and offsets to construct the local energy term U(x (x,y) ), used to measure the pixel state x (x,y) Figure 2. Vascular structure characteristics. vascular The degree of match between (x,y): U(x (x,y) )=κ(|x (x,y) -T vascular (x,y)|+θ) η ; Among them, κ is the overall weight factor, which is used to balance the impact of the difference between local pixels and feature maps, θ is the offset, and η is the nonlinear index; S33. Constructing the neighborhood smoothing constraint energy term V(x (x,y) ,x (u,v) ), which is used to constrain the state consistency between the pixels in the blood vessel neighborhood and introduce the similarity of blood vessel features to dynamically adjust the smoothness: V(x (x,y) ,x (u,v) )=α (x,y),(u,v) exp(-β∥T vascular (x,y)-T vascular (u,v)∥)·(x (x,y) -x (u,v) ) 2 ; Among them, α (x,y),(u,v) is the basic smoothing coefficient, β is the feature similarity adjustment factor, and the difference between adjacent pixels in the vascular feature map is reflected by exponential decay. vascular (x,y)-T vascular (u,v)∥ represents the feature difference between pixel (x,y) and pixel (u,v) in the vascular structure feature map. (x,y) -x (u,v) ) 2 is the state difference term, which is used to measure the inconsistency of the state of the neighborhood pixels; S34. The local energy term is combined with the neighborhood smoothness constraint energy term to construct the total energy function E(X) of the random field model to characterize the overall optimization goal in the 3D reconstruction of blood vessels: E(X)=∑ (x,y)∈Ω ω (x,y) U(x (x,y) )+∑ (x,y)∈Ω ∑ (u,v)∈N(x,y) γ (x,y),(u,v) V(x (x,y) ,x (u,v) ); Among them, ω (x,y) is the local energy weight coefficient, which is used to assign differential weights to pixels at different locations to better focus on the stenosis site. N(x, y) represents the neighborhood set of pixel (x, y), and γ (x,y),(u,v) is the neighborhood smoothing weight coefficient, which is used to balance the proportion of local energy terms and smoothing constraint energy terms in 3D vascular modeling. The minimization result of E(X) is used to generate the pixel state distribution in the 3D vascular reconstruction process.

5. The method for three-dimensional reconstruction of vascular stenosis based on random field depth modeling according to claim 1, characterized in that: The S4 comprises the following specific steps: S41. Based on the total energy function E(X) of the random field model, using the vascular structure feature map T vascular (x,y) Initialize the candidate solution set: in, is the i-th initial solution, f init is the initialization function based on the vascular feature map, ξ i represents the random perturbation introduced to increase the diversity of candidate solutions, and n is the number of candidate solutions; S42. In the candidate solution set C0, several low-energy solutions are selected by calculating the value of the total energy function E(X) to enter the next stage. The optimization strategy of global search is defined as: in, Represents the optimal solution in the global search phase and defines the screening rules: C global ={X i ∈C0∣E(X i )≤τ}; Among them, τ is the global search energy threshold, and the set of candidate solutions C is selected. global Solutions with lower energy values ​​are retained; S43. The candidate solution set C obtained by global search global , using the optimization strategy of gradient descent combined with dynamic perturbation to perform local fine optimization in the vascular bifurcation area and stenosis area: in, is the candidate solution at the tth iteration, η1 is the dynamic step size factor, which is dynamically adjusted according to the energy gradient of the local search. represents the gradient of the energy function, δ t is a local disturbance term, simulating the random exploration mechanism in the gold mining optimization algorithm; S44. During the local fine search process, the energy change trend of the candidate solution is evaluated in real time, and the step size and disturbance term are dynamically updated: or t+1 =h(η t ); d t+1 =g(δ t ); Among them, h(η t ) represents the step size adjustment function, which is used to increase the step size when the gradient change is less than the threshold, and to reduce the step size when the gradient change is greater than the threshold. t ) represents the disturbance adjustment function; S45. Combine the results of global search and local fine search to select the solution corresponding to the minimum value of the total energy function E(X) as the final optimization solution: Among them, X optimal is the global optimal solution of the random field model.

6. The method for three-dimensional reconstruction of vascular stenosis based on random field deep modeling according to claim 1, characterized in that: The S5 comprises the following specific steps: S51. From the final optimization solution optimal Extract the optimal state of each pixel (x, y) S52. The optimal state Convert to three-dimensional space coordinate point S (x,y) , construct the mapping function: Among them, Γ(·; ·) is the state-coordinate mapping function, which comprehensively utilizes the optimal state of pixels and the vascular structure feature map T vascular (x, y) information to generate the three-dimensional coordinates S (x,y) To preserve the overall structure of blood vessels, the spatial characteristics of the vascular bifurcation area and the vascular stenosis area; S53. According to all pixel coordinate points S (x,y) Constructing 3D blood vessel surface mesh data M 3D : Where V = {S (x,y) } is the node set of the three-dimensional grid, E is the edge set of the three-dimensional grid, Λ(·,·) is the connectivity judgment function, which determines whether the nodes are spatially connected based on the neighborhood relationship of the two-dimensional image plane and the similarity of the optimal state, so as to describe the overall topological structure of the blood vessel, the bifurcation of the blood vessel and the location of the stenosis; S54. For the three-dimensional surface mesh data M 3D Perform partition marking and sort nodes according to the optimal state The value characteristics of are divided into normal blood vessel segment, blood vessel bifurcation segment and blood vessel stenosis segment: Among them, C(·) is the classification function, Δ(·) represents the operation of feature extraction for the optimal state, and Γ norm ,Γ bif ,Γ sten It is the threshold or judgment range corresponding to the normal segment, bifurcation segment and stenosis segment, and is used to highlight the spatial distribution of the overall vascular structure, vascular bifurcation area and vascular stenosis area in the three-dimensional surface mesh data.

Images