Artificial blood vessel helical structure parameter optimization method and system and artificial blood vessel

Abstract

The application discloses a kind of artificial blood vessel helical structure parameter optimization method, system and artificial blood vessel, and wherein, method includes: the data set of helical structure parameter and adverse hemodynamic area ratio of proportion with corresponding relationship is constructed;Attention-enhanced residual deep neural network model is established with helical structure parameter as input, with adverse hemodynamic area ratio as output;The depth neural network model is trained and tested;On this basis, using the dynamic optimization strategy of bayesian optimization and genetic algorithm cooperation, the multi-objective optimization of neural network model implemented after training is carried out, so as to obtain the optimal helical structure parameter combination.The application significantly improves the optimization efficiency and design quality of artificial blood vessel helical structure parameter, and provides a reliable data-driven solution for artificial blood vessel performance optimization.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of biomedical engineering and artificial intelligence, and specifically relates to a method, system and artificial blood vessel for optimizing the parameters of the spiral structure of artificial blood vessels. Background Technology

[0002] Artificial blood vessels play a crucial role in hemodialysis, providing adequate blood flow by establishing arteriovenous graft (AVG) access. However, traditional artificial blood vessels have long faced the problem of anastomotic stenosis between the artificial vessel and the natural vein, resulting in access failure rates as high as 30%-50%, severely impacting patients' quality of life. Studies have shown that abnormal hemodynamics (such as high oscillatory shear stress and low time-averaged wall shear stress) are the core factors inducing these complications. To improve hemodynamic performance, spiral-structured artificial blood vessels have emerged. Clinical studies have shown that these vessels, by guiding blood to form a spiral flow, can more effectively maintain blood flow stability, and their one-year patency rate shows a significant advantage over traditional artificial blood vessels. However, despite the overall improvement in the performance of spiral-structured artificial blood vessels, the problem of anastomotic stenosis has not been fundamentally solved, with some clinical reports indicating a failure rate still exceeding 20%. This highlights that the design of existing helical structures may not be globally optimal, and the mapping relationship between their key geometric parameters and complex hemodynamic responses has not yet been precisely grasped. The optimization of helical structure parameters faces multiple bottlenecks: First, there are significant conflicts between multiple objectives. For example, reducing the pitch can enhance blood flow stability, but an excessively small pitch can induce turbulence due to excessive tortuosity of the flow path, leading to mechanical damage to vascular endothelial cells; increasing the ridge height can optimize the distribution of wall shear stress, but an excessively large ridge can form oscillating blood flow at the helical ridge, leading to abnormal blood flow; Second, traditional optimization methods rely on trial and error, making it difficult to quickly converge to the global optimum in a high-dimensional parameter space, and they do not dynamically adjust weights based on clinical data.

[0003] Therefore, there is an urgent need for an intelligent artificial design method that integrates hemodynamic simulation, machine learning, and multi-objective collaborative optimization to overcome the technical bottleneck of spiral structure parameter optimization and effectively suppress AVG venous anastomosis stenosis. Summary of the Invention

[0004] This invention addresses the problems of low optimization efficiency and high risk of postoperative anastomotic stenosis caused by relying on experience-based trial and error, difficulty in coordinating conflicting multi-objective parameters, lack of efficient prediction models, etc. in the design of traditional artificial blood vessel spiral structures. It proposes a method, system, and artificial blood vessel for optimizing artificial blood vessel spiral structure parameters. By integrating data-driven prediction and multi-objective dynamic optimization, it breaks through the limitations of traditional experience-based design, significantly shortens the artificial blood vessel design cycle, and reduces the risk of postoperative stenosis of artificial blood vessels.

[0005] To achieve the objectives of this invention, the technical solution provided by this invention is as follows:

[0006] First aspect

[0007] This application provides a method for optimizing the parameters of an artificial blood vessel spiral structure, including the following steps:

[0008] Step S1: Construct a dataset with corresponding spiral structure parameters and the area ratio of unfavorable hemodynamic regions, and divide the dataset into a training set and a test set;

[0009] Step S2: Establish a deep neural network model with spiral structure parameters as input and the area ratio of unfavorable hemodynamic regions as output;

[0010] Step S3: Based on the training set, the deep neural network model is trained using K-fold cross-validation. The number of hidden layer nodes and the Dropout rate are optimized through grid search. The prediction error index is calculated based on the test set. The deep neural network model with the best generalization ability is selected as the final model for predicting hemodynamic parameters.

[0011] Step S4: Using a dynamic optimization strategy that combines Bayesian optimization and genetic algorithm, the output parameters of the final model are optimized for multiple objectives to obtain the optimized spiral structure parameters of the artificial blood vessel.

[0012] Further, step S1 specifically includes:

[0013] Step S1.1: Construct an anastomosis model of a spiral artificial blood vessel and a natural vein with different spiral structure parameters using 3D modeling software. The spiral structure parameters include pitch P, ridge height H, and anastomosis angle A.

[0014] Step S1.2: Perform computational fluid dynamics (CFD) calculations on the anastomosis models of each spiral artificial blood vessel and natural vein. Extract the area percentage of unfavorable hemodynamic regions at the anastomosis site for each spiral artificial blood vessel and natural vein anastomosis model. Among them, the area percentages of regions with wall shear stress WSS>7, average wall shear stress TAWSS<0.1, and oscillating shear index OSI>0.25 are respectively SA WSS SA TAWSS SA OSI ;

[0015] Step S1.3: After normalizing the spiral structure parameters and the corresponding area proportions of unfavorable hemodynamic regions, filtering out noise, and extracting features, the dataset is obtained;

[0016] Step S1.4: Divide the dataset into training sets according to a preset ratio. With test set .

[0017] Furthermore, the deep neural network model in step S2 includes: an input layer, a feature interaction layer, a self-attention mechanism, n residual block layers, and an output layer.

[0018] The input layer is used to perform Min-Max standardization on the spiral structure parameters and output a standardized feature vector X.

[0019] The feature interaction layer is used to expand the standardized feature vector X using a second-order polynomial to generate a feature set Φ(X) containing interaction terms.

[0020] The self-attention mechanism is used to obtain weighted features by calculating the attention weights of the attention weight matrix based on the extended feature set Φ(X);

[0021] The n residual block layers are used to output a high-level semantic representation of the hemodynamic prediction task based on weighted features;

[0022] The output layer is used to obtain the final prediction result based on the high-level semantic representation.

[0023] Further, step S3 specifically includes:

[0024] Step S3.1: Obtain the training set Divide the data into K=5 mutually exclusive subsets and define the k-th fold verification set. and training set ,satisfy: k=1,2,...,5;

[0025] Step S3.2: Update the parameters of the deep neural network model using the AdamW optimizer, as shown in the following formula:

[0026] ;

[0027] in, The parameters are updated at time t. Here are the parameters before the update at time t-1, and η is the learning rate. and For estimation of first-order and second-order matrices, λ is the weight decay coefficient;

[0028] Step S3.3: Calculate the difference between the model's predicted value and the true value using the Huber loss function, and output the loss value L;

[0029] Step S3.4: Based on the loss value L, optimize the hyperparameter combination through grid search to determine the optimal hyperparameters;

[0030] Step S3.5: Based on the optimal hyperparameters and test set The deep neural network model was tested, and its performance was evaluated based on mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). The deep neural network model with the best generalization ability was selected as the final model for predicting hemodynamic parameters.

[0031] Further, step S4 specifically includes:

[0032] Step S4.1: Construct the multi-objective optimization function:

[0033] ;

[0034] in, α , β , γ These are the weighting coefficients, and the objective is to minimize them. The weighting coefficient is used to reduce the risk of venous anastomotic stenosis, and the weighting coefficient can be dynamically adjusted according to clinical feedback to adapt to the patient's blood flow environment needs.

[0035] Step S4.2: Perform dynamic optimization using a combination of Bayesian optimization and genetic algorithm, including:

[0036] Bayesian optimization phase:

[0037] Define the parameter space of the helical structure as follows: ,in, , , ;in, This is the set minimum pitch value. The maximum set pitch. The minimum ridge height is set. The maximum value of the ridge height is set. This is the minimum set matching angle. The maximum value of the matching angle is set; a Gaussian process surrogate model is constructed, using the Matérn kernel function, and the acquisition function is the desired improved EI, iterating N times;

[0038] Genetic Algorithm Stage:

[0039] Individuals with a predetermined fitness level are selected from the Bayesian optimization results as the initial population; offspring are generated through crossover and mutation operations, and the individuals with the highest fitness are retained according to the fitness function. M Individual;

[0040] Dynamic collaborative strategy: per iteration T After the genetic algorithm, the current optimal parameter set is added to the Bayesian optimized observation dataset to update the surrogate model; the termination condition is the rate of change of the fitness function. Or the total number of iterations reaches the maximum number of iterations. This is the threshold parameter.

[0041] Furthermore, the process also includes the following steps: verifying the hemodynamic performance of the three-dimensional artificial blood vessel model generated using the optimized spiral structure parameters through CFD simulation; if the simulation verification is incorrect, the parameter is removed and resampled.

[0042] Second aspect

[0043] Corresponding to the above method, this application also provides an artificial blood vessel spiral structure parameter optimization system, including the following units: a dataset acquisition unit, a deep neural network model building unit, a model training and testing unit, and a multi-objective optimization unit;

[0044] The dataset acquisition unit is used to construct a dataset with corresponding spiral structure parameters and the area ratio of unfavorable hemodynamic regions, and divide the dataset into a training set and a test set.

[0045] The deep neural network model building unit is used to build a deep neural network model with spiral structure parameters as input and the area ratio of unfavorable hemodynamic regions as output.

[0046] The model training and testing unit is used to train the deep neural network model using K-fold cross-validation based on the training set, optimize the number of hidden layer nodes and the Dropout rate through grid search, calculate the prediction error index based on the test set, and select the deep neural network model with the best generalization ability as the final model for hemodynamic parameter prediction.

[0047] The multi-objective optimization unit is used to perform multi-objective optimization on the output parameters of the final model using a dynamic optimization strategy that combines Bayesian optimization and genetic algorithm, so as to obtain the optimized spiral structure parameters of the artificial blood vessel.

[0048] Third aspect

[0049] This application also provides an artificial blood vessel whose helical structure parameters are obtained according to the artificial blood vessel helical structure parameter optimization method.

[0050] Furthermore, the artificial blood vessel has a diameter of 6 mm and spiral structure parameters of 29.2 mm pitch, 1.75 mm ridge height, and 38.6° anastomosis angle.

[0051] Furthermore, the fabrication process of the artificial blood vessel includes the following:

[0052] First, a sugar helical structure is generated by melt spinning maltitol.

[0053] Then, using electrospinning technology, an optimized spiral artificial blood vessel is prepared by spinning an electrospun fiber skeleton using biopolymer materials as raw materials; wherein, the biopolymer materials are one or a mixture of several of them in any proportion, including polycaprolactone PCL, poly-L-lactide-caprolactone PLCL, polylactic acid PLA, polyglycolic acid PGA, polylactic acid-glycolic acid copolymer PLGA, polyurethane PU and its modified copolymers.

[0054] Compared with existing technologies, traditional artificial blood vessel spiral structure parameter optimization methods suffer from significant multi-objective conflicts, reliance on trial and error making it difficult to quickly converge to the global optimum, and a lack of dynamic weight adjustment based on clinical data. The advantages of this invention are as follows:

[0055] This invention addresses the challenges of traditional methods by innovatively proposing an optimization approach based on a hybrid algorithm and dynamic weight adjustment. This method leverages attention-enhanced residual deep neural networks to rapidly and accurately predict the hemodynamic performance of artificial blood vessels, avoiding the blind spots of traditional methods. Simultaneously, by combining a collaborative search mechanism of Bayesian optimization and genetic algorithms, efficient global optimization of structural parameters can be achieved in a high-dimensional parameter space, effectively resolving multi-objective conflict problems. Furthermore, this invention integrates manufacturing constraints, ensuring that the design is not only theoretically feasible but also practically implementable in manufacturing. Through these innovative measures, this invention can effectively suppress AVG venous anastomosis stenosis, ultimately significantly improving the long-term patency rate of artificial blood vessels and providing hemodialysis patients with a higher quality and more reliable treatment solution.

[0056] The invention utilizes the proposed scheme combined with melt spinning, electrospinning, and 3D printing technologies to actually fabricate an optimized spiral artificial blood vessel. This vessel was then applied to a large animal model of porcine carotid artery and vein artificial blood vessel grafts. Through long-term monitoring of hemodynamic parameters and vascular patency, the effectiveness of the proposed scheme was verified. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the method for optimizing the parameters of the spiral structure of artificial blood vessels provided in an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of the deep neural network model structure provided in an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the spiral structure of the artificial blood vessel and the anastomosis in an embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of the macroscopic structure of the spiral artificial blood vessel in an embodiment of the present invention;

[0061] Figure 5This is a schematic diagram showing the HE histological staining effect of venous anastomosis in animals after surgery using a straight cylindrical graft and a spiral graft of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0063] It should be noted that this invention combines computational fluid dynamics (CFD) simulation, deep neural network modeling, and hybrid optimization algorithms to optimize the helical structure parameters of artificial blood vessels—pitch, ridge height, and anastomosis angle—based on the hemodynamic characteristics of arteriovenous anastomoses, in order to reduce the risk of thrombosis and improve the long-term patency of blood vessels.

[0064] Specifically, such as Figure 1 As shown in the figure, the method for optimizing the parameters of an artificial blood vessel spiral structure provided in this embodiment includes the following steps:

[0065] Step S1: Construct a dataset with corresponding spiral structure parameters and the area proportion of adverse hemodynamic regions, and divide the dataset into a training set and a test set; specifically including:

[0066] Step S1.1: Construct an anastomosis model of a spiral artificial blood vessel and a natural vein with different spiral structure parameters using 3D modeling software. The spiral structure parameters include pitch P, ridge height H, and anastomosis angle A.

[0067] It should be noted that the anastomosis model of the spiral artificial blood vessel and the natural vein was constructed using the 3D modeling software Solidworks, with the following parameters set: pitch P∈[3mm,100mm], ridge height H∈[0.1mm,3mm], and anastomosis angle A∈[15°,75°]. Figure 3 As shown.

[0068] Step S1.2: Perform computational fluid dynamics (CFD) calculations on the anastomosis models of each spiral artificial blood vessel and natural vein. Extract the area percentage of unfavorable hemodynamic regions at the anastomosis site for each spiral artificial blood vessel and natural vein anastomosis model. Among them, the area percentages of regions with wall shear stress WSS>7, average wall shear stress TAWSS<0.1, and oscillating shear index OSI>0.25 are respectively SA WSS SA TAWSS SA OSI ;

[0069] It should be noted that computational fluid dynamics (CFD) simulations were performed on each spiral artificial blood vessel and natural vein anastomosis model. The inlet velocity of the artificial blood vessel was set to a typical pulsatile blood flow velocity curve for hemodialysis, and transient solutions were obtained using COMSOL Multiphysics simulation software. Subsequently, the area percentage of unfavorable hemodynamic regions at the anastomosis site of the artificial blood vessel and natural vein for each model was extracted from the simulation results, including the SA (subtraction anastomosis). WSS (Percentage of area with WSS > 7Pa), SA TAWSS (Percentage of area where TAWSS < 0.1Pa), SA OSI (Percentage of areas with OSI>0.25).

[0070] Step S1.3: After normalizing the spiral structure parameters and the corresponding area proportions of unfavorable hemodynamic regions, filtering out noise, and extracting features, the dataset is obtained;

[0071] It should be noted that the input parameters (P,H,A) are normalized using Min-Max and mapped to the [0,1] interval; a median filter is used to remove outlier data points.

[0072] Step S1.4: Divide the dataset into training sets according to a preset ratio. With test set .

[0073] It should be noted that the training set is divided in a 7:3 ratio. (210 sets of data) and test set (90 sets of data) The dataset completed in step S1 provides standardized input data for the deep neural network model established in step S2.

[0074] Step S2: Establish a deep neural network model with spiral structure parameters as input and the area ratio of unfavorable hemodynamic regions as output;

[0075] The deep neural network model in step S2 includes: an input layer, a feature interaction layer, a self-attention mechanism, n residual block layers, and an output layer.

[0076] The input layer is used to perform Min-Max standardization on the spiral structure parameters and output a standardized feature vector X.

[0077] It should be noted that the helical structure parameters (pitch P, ridge height H, and coincidence angle A) obtained in step S1 are standardized to output a standardized feature vector X:

[0078] ;

[0079] Where norm represents minimum-maximum standardization, and each parameter is standardized by Min-Max: ,in, and These are the minimum and maximum values ​​of the pitch P in the training dataset, respectively. The same method is applied to... and The output of this layer is a standardized feature vector X, avoiding training bias caused by differences in parameter numerical ranges. The standardized feature vector X serves as the input to the feature interaction layer, providing a numerically stable foundation for subsequent feature expansion, enabling parameters of different magnitudes to interact within the same numerical range.

[0080] The feature interaction layer is used to expand the standardized feature vector X using a second-order polynomial to generate a feature set Φ(X) containing interaction terms.

[0081] It should be noted that,

[0082]

[0083] This extended feature set is used to enhance the model's nonlinear expressive power in order to better capture the complex relationship between helical structure parameters and hemodynamic response. As input to the self-attention mechanism, it provides rich feature combinations, enabling attention weight calculation to be based on more comprehensive feature information.

[0084] The self-attention mechanism is used to adjust the feature set based on the extended feature set. We obtain the weighted features by calculating the attention weights in the attention weight matrix;

[0085] It should be noted that the expanded feature set Input from the attention layer, calculate the attention weight matrix:

[0086] ;

[0087] Here, Q, K, and V are obtained from Φ(X) through linear transformation, representing the dimensions of the key vector. By calculating the attention weights among the query matrix Q, key matrix K, and value matrix V, adaptive weighting of the importance of different features is achieved, highlighting the features more critical to hemodynamic prediction. The output of the attention mechanism is the weighted feature. ,and As the direct input to a deep residual feature extraction network with n residual blocks, this feature weighting process lays an important foundation for subsequent deep feature extraction, ensuring that the network can focus on feature information that contributes more to the prediction task.

[0088] The n residual block layers are used to output a high-level semantic representation of the hemodynamic prediction task based on weighted features;

[0089] It should be noted that the weighted features As the input to n residual block layers, denoted as Z(0), Z(0) is used as the input to the 0th residual block. This input feature has been refined by the self-attention mechanism and contains feature information that is more important for hemodynamic prediction tasks.

[0090] The design uses multiple residual blocks, and the calculation process for each residual block is as follows:

[0091] ;

[0092] ;

[0093] ;

[0094] Z(l) and Z(l+1) represent the input and output of the l-th residual block, respectively. This represents the intermediate feature transformation results during the residual block computation process. and Here are the weight matrix and bias vector for the first linear transformation layer. and Here, represents the weight matrix and bias vector of the second linear transformation layer, BN represents batch normalization, and ReLU is the activation function.

[0095] Residual connections ensure effective gradient propagation, progressively extracting higher-order feature representations. The output Z(L) of the last residual block serves as the input to the output layer. This deep feature extraction process further abstracts the weighted features into high-level semantic representations, providing rich feature information for the final prediction task.

[0096] The output layer is used to obtain the final prediction result based on the high-level semantic representation.

[0097] It should be noted that the output layer takes the output Z(L) of the last residual block as input to obtain the final prediction result:

[0098] ;

[0099] in The predicted output vector represents the predicted proportion of the three adverse hemodynamic regions, and its specific structure is as follows: ,in, This represents the predicted percentage of the area of ​​the region with unfavorable wall shear stress (WSS) in blood vessels. This represents the predicted percentage of the area of ​​the mean wall shear stress (TAWSS) region when blood vessels are unfavorable. This represents the predicted percentage of the area affected by the unfavorable oscillation shear index (OSI). The predicted output values ​​corresponding to different helical structure parameters (P, H, A) constitute an end-to-end mapping from helical structure parameters to hemodynamic responses. The complete deep neural network model architecture constructed in step S2 serves as the training foundation for step S3, providing a structural framework for model parameter optimization. The entire S2 step, through a progressive process of feature preprocessing, expansion, weighting, extraction, and prediction, constructs a neural network model capable of effectively modeling the complex mapping relationship between helical parameters and hemodynamic responses. This architecture captures the nonlinear relationships between parameters through feature interaction layers, identifies key features using a self-attention mechanism, and deepens the network's learning capabilities with the help of residual structures, effectively modeling the complex mapping relationship between helical parameters and hemodynamic responses, such as... Figure 2 As shown.

[0100] Steps S1 and S2 provide the basic framework for the dataset and deep neural network model for step S3, respectively.

[0101] Step S3: Based on the training set, the deep neural network model is trained using K-fold cross-validation. The number of hidden layer nodes and the Dropout rate are optimized through grid search. The prediction error index is calculated based on the test set. The deep neural network model with the best generalization ability is selected as the final model for hemodynamic parameter prediction. Specifically, this includes:

[0102] Step S3.1: Obtain the training set Divide the data into K=5 mutually exclusive subsets and define the k-th fold verification set. and training set ,satisfy: k=1,2,...,5;

[0103] Step S3.2: Update the parameters of the deep neural network model using the AdamW optimizer, as shown in the following formula:

[0104] ;

[0105] in, The parameters are updated at time t. Here are the parameters before the update at time t-1, and η is the learning rate. and The first-order and second-order matrices are estimated, and λ is the weight decay coefficient. The configured optimizer will provide a parameter update mechanism for the loss function in step S3.3, realizing the optimization of model parameters based on gradient descent.

[0106] Step S3.3: Calculate the difference between the model's predicted value and the true value using the Huber loss function, and output the loss value L;

[0107] It should be noted that the Huber loss function is used, and its mathematical expression is as follows:

[0108] ;

[0109] ;

[0110] in, The threshold parameter is set to 0.5. This loss function is used to calculate the model predictions output in step S2.5. The difference between the actual value y and the true value y is calculated, and the loss value L is output, representing the prediction error of the current batch of data, thereby guiding the model hyperparameter optimization process in step S3.4.

[0111] Step S3.4: Based on the loss value L, optimize the hyperparameter combination through grid search to determine the optimal hyperparameters;

[0112] It should be noted that in this step, the hyperparameter combination is optimized through grid search, and the search space is defined as follows:

[0113] Number of hidden layer nodes ;

[0114] Dropout rate ;

[0115] Learning rate ;

[0116] Batch size ;

[0117] A grid search strategy is employed to traverse all hyperparameter combinations. For each hyperparameter combination, K-fold cross-validation training is performed, recording the validation set performance at each fold, and calculating the average performance metric as the final score for that hyperparameter combination. The average performance metric for each hyperparameter combination on the validation set, as well as the optimal hyperparameter settings that perform best on the validation set, are output. The optimal hyperparameters determined in step S3.4 will be used for model training and performance evaluation in step S3.5.

[0118] Step S3.5: Based on the optimal hyperparameters and test set The deep neural network model was tested, and its performance was evaluated based on mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). The deep neural network model with the best generalization ability was selected as the final model for predicting hemodynamic parameters.

[0119] It should be noted that in this step, model selection is based on cross-validation performance, and the model performance evaluation metrics include:

[0120] ;

[0121] ;

[0122] ;

[0123] Where q is the test set The number of samples in Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample. This is the average of the true values ​​of all samples in the test set. The input to this step is the deep neural network model trained with optimal hyperparameters, and the independent test set retained in step S1. Finally, the model's predictive performance is comprehensively evaluated based on multiple indicators, including mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²) on the test set. The model with the best generalization ability is selected as the final model for hemodynamic parameter prediction. The optimal prediction model output in this step will serve as the core component of the multi-objective optimization in step S4, used to evaluate the hemodynamic performance of different helical structure parameters.

[0124] Step S4: Employing a dynamic optimization strategy combining Bayesian optimization and genetic algorithms, multi-objective optimization is performed on the output parameters of the final model to obtain the optimized spiral structure parameters of the artificial blood vessel, specifically including:

[0125] Step S4.1: Construct the multi-objective optimization function:

[0126] ;

[0127] in, α , β , γ These are the weighting coefficients, and the objective is to minimize them. The weighting coefficient is used to reduce the risk of venous anastomotic stenosis, and the weighting coefficient can be dynamically adjusted according to clinical feedback to adapt to the patient's blood flow environment needs.

[0128] The optimization objective is to minimize This value reduces the risk of anastomotic stenosis. The output of this step is the value corresponding to different spiral structure parameters. and will This information is then used as input to step S4.2 to guide the parameter optimization process.

[0129] It should be noted that in this step, the area of ​​the adverse hemodynamic region predicted by the best-performing deep neural network model obtained in step S3 is used. Based on this, a comprehensive objective function is established.

[0130] Step S4.2: Perform dynamic optimization using a combination of Bayesian optimization and genetic algorithm, including:

[0131] Bayesian optimization phase:

[0132] Define the parameter space of the helical structure as follows: ,in, , , ;in, This is the set minimum pitch value. The maximum set pitch. The minimum ridge height is set. The maximum value of the ridge height is set. This is the minimum set matching angle. The maximum value of the matching angle is set; a Gaussian process surrogate model is constructed, using the Matérn kernel function, and the acquisition function is the desired improved EI, iterating N times;

[0133] It should be noted that in this stage, the output f(P,H,A) from step S4.1 is used as input to explore the parameter space, and the parameter space of the spiral structure is defined as follows: Where P∈[3mm,100mm], H∈[0.1mm,3mm], and A∈[15°,75°]; a Gaussian process surrogate model is constructed to fit the objective function, the kernel function is Matérn5 / 2, the acquisition function is Expectation Improvement (EI), the iteration is performed 50 times, and 10 sets of candidate parameters are output as the initialization basis for the genetic algorithm stage.

[0134] Genetic Algorithm Stage:

[0135] Individuals with a predetermined fitness level are selected from the Bayesian optimization results as the initial population; offspring are generated through crossover and mutation operations, and the individuals with the highest fitness are retained according to the fitness function. M Individual;

[0136] Specifically, it includes:

[0137] Genetic algorithm initialization: Select the top 20% of individuals with the best fitness (i.e., 2 sets of parameters) from the Bayesian optimization results as output. This forms the initial population, and the population size is set to 20.

[0138] Genetic operation execution: on the population (initially the input of the previous stage, and subsequently the output of the previous generation);

[0139] Perform crossover and mutation operations: crossover probability set to PC =0.8, through linear combination of parent individuals pi and pj Generate offspring, that is ,in, The random scaling factor is within the range (0,1); the probability of variation is set to pm=0.1, and a normal distribution perturbation is applied to the individual parameters, with the perturbation amplitude being... =0.1×parameter range, enhancing population diversity. Outputs a new population after genetic manipulation (containing parent and offspring individuals).

[0140] Input: Output is the genetically modified population, which performs individual selection and population maintenance. Based on the fitness function... Assess individual traits, retain the top 50% of fittest individuals, maintain a population size of 20, ensuring superior genes are passed on to the next generation. Output a new generation population of 20. This selected population will be used to determine convergence conditions in the next stage or to continue the next round of genetic evolution.

[0141] Dynamic collaborative strategy: per iteration T After the genetic algorithm, the current optimal parameter set is added to the Bayesian optimized observation dataset to update the surrogate model; the termination condition is the rate of change of the fitness function. Or the total number of iterations reaches the maximum number of iterations. This is the threshold parameter.

[0142] It should be noted that the input consists of the currently optimal parameters of the individual and the population state. After every 5 generations of the genetic algorithm, the current optimal parameters are added to the observation dataset of the Bayesian optimization, updating the posterior distribution of the Gaussian process surrogate model, thus realizing information exchange and collaborative search between the two optimization methods; when the rate of change of the fitness function Δ f< When the value is 0.01, the algorithm is considered to have converged, the optimization process is terminated, and the optimal combination of spiral structure parameters is output.

[0143] Furthermore, the process also includes the following steps: verifying the hemodynamic performance of the three-dimensional artificial blood vessel model generated using the optimized spiral structure parameters through CFD simulation; if the simulation verification is incorrect, the parameter is removed and resampled.

[0144] In this embodiment, the final optimized result is a pitch of 29.2 mm, a ridge height of 1.75 mm, and an anastomosis angle of 38.6°. This parameter combination is designed for a 6 mm diameter artificial blood vessel. In practical applications, the parameter range can be adjusted to adapt to different blood vessel diameter requirements.

[0145] In addition, this application also includes the step of generating an optimized design scheme: generating a three-dimensional artificial blood vessel model based on the optimized spiral structure parameters, and verifying its hemodynamic performance through CFD simulation.

[0146] Maltitol was used for melt spinning to generate a sugar helical structure, and then electrospinning was employed with polycaprolactone (PCL) as the raw material to spin an electrospun fiber skeleton, thus actually preparing an optimized helical artificial blood vessel. Figure 4As shown. Furthermore, in this embodiment, the prepared helical artificial blood vessel is made by adding a melt-spun support ring using PCL as the raw material, thereby meeting the clinical requirements for the anti-kink performance of artificial blood vessels.

[0147] The optimized spiral artificial blood vessel (spiral graft) and the control traditional straight graft were experimentally validated in a large animal porcine carotid artery and vein artificial blood vessel graft fistula model. The anastomosis angle was controlled by a 3D-printed angle fixator. HE histological staining of the venous anastomosis site one month post-operatively in the experimental porcine artificial blood vessel fistula was performed. Figure 5 As shown, the venous anastomosis corresponding to the spiral graft showed no obvious signs of hyperplasia, with a neointima-to-lumen ratio of only 2.67%. In contrast, the venous anastomosis corresponding to the straight graft, especially at the suture site between the graft and the natural vein, showed significant neointima hyperplasia, with a neointima-to-lumen ratio as high as 27.03%. The results of three control and three implementation animal experiments indicate that this optimized spiral artificial blood vessel demonstrated 100% patency in the porcine carotid artery and vein AVG model for one month (see Table 1).

[0148] Table 1

[0149] ;

[0150] As can be seen, this invention, by establishing a multi-dimensional mapping model between helical structure parameters and hemodynamic performance, combined with intelligent optimization algorithms and an automated parameter screening platform, not only overcomes the challenge of co-optimizing multiple objective parameters such as anastomosis angle, pitch, and ridge height in traditional artificial blood vessel design, but also achieves efficient convergence of the optimal solution in a high-dimensional parameter space—completing the iterative verification process that would take months using traditional methods within 72 hours, significantly shortening the R&D cycle. Addressing the real-time hemodynamic weight requirements from clinical feedback, this method, through a dynamic weight allocation mechanism, can automatically adapt hemodynamic priority weights, outputting the optimal combination of helical blood vessel structure parameters and anastomosis angle in real time. Furthermore, by combining various blood vessel preparation techniques such as melt spinning and electrospinning, it can accurately and efficiently complete the preparation of optimized helical artificial blood vessels and perform in vitro functional verification. This forms a closed-loop optimization system from virtual optimization to physical verification, providing a precise and batch-producible artificial blood vessel structure optimization scheme for hemodynamically adaptable artificial blood vessel design for complex vascular lesions.

[0151] Corresponding to the above method, this embodiment also provides an artificial blood vessel spiral structure parameter optimization system, including the following units: a dataset acquisition unit, a deep neural network model building unit, a model training and testing unit, and a multi-objective optimization unit;

[0152] The dataset acquisition unit is used to construct a dataset with corresponding spiral structure parameters and the area ratio of unfavorable hemodynamic regions, and divide the dataset into a training set and a test set.

[0153] The deep neural network model building unit is used to build a deep neural network model with spiral structure parameters as input and the area ratio of unfavorable hemodynamic regions as output.

[0154] The model training and testing unit is used to train the deep neural network model using K-fold cross-validation based on the training set, optimize the number of hidden layer nodes and the Dropout rate through grid search, calculate the prediction error index based on the test set, and select the deep neural network model with the best generalization ability as the final model for hemodynamic parameter prediction.

[0155] The multi-objective optimization unit is used to perform multi-objective optimization on the output parameters of the final model using a dynamic optimization strategy that combines Bayesian optimization and genetic algorithm, so as to obtain the optimized spiral structure parameters of the artificial blood vessel.

[0156] This embodiment also provides an artificial blood vessel, the spiral structure parameters of which are obtained according to the artificial blood vessel spiral structure parameter optimization method described above.

[0157] Furthermore, the artificial blood vessel has a diameter of 6 mm and spiral structure parameters of 29.2 mm pitch, 1.75 mm ridge height, and 38.6° anastomosis angle.

[0158] Furthermore, the fabrication process of the artificial blood vessel includes the following:

[0159] First, a sugar helical structure is generated by melt spinning maltitol.

[0160] Then, using electrospinning technology, an optimized spiral artificial blood vessel is prepared by spinning an electrospun fiber skeleton using biopolymer materials as raw materials; wherein, the biopolymer materials are one or a mixture of several of them in any proportion, including polycaprolactone PCL, poly-L-lactide-caprolactone PLCL, polylactic acid PLA, polyglycolic acid PGA, polylactic acid-glycolic acid copolymer PLGA, polyurethane PU and its modified copolymers.

[0161] The embodiments of this invention are merely illustrative of the technical solutions for the multi-objective intelligent optimization method and system of the spiral artificial blood vessel, and are not intended to limit the scope of protection of this invention. After reading the technical content disclosed in this invention, those skilled in the art can reasonably modify or equivalently replace the definition of spiral structure parameters, hybrid optimization algorithm flow, or weight adjustment strategy based on actual application scenarios. Such modifications or adjustments, if they do not depart from the core technical features defined in the claims of this invention, should be considered to fall within the patent protection scope of this invention.

Claims

1. A method for optimizing the parameters of an artificial blood vessel spiral structure, characterized in that, Includes the following steps: Step S1: Construct a dataset with corresponding spiral structure parameters and the area ratio of unfavorable hemodynamic regions, and divide the dataset into a training set and a test set; Step S2: Establish a deep neural network model with helical structure parameters as input and the area ratio of unfavorable hemodynamic regions as output; the helical structure parameters include pitch P, ridge height H, and anastomosis angle A; the area ratio of unfavorable hemodynamic regions includes the area ratios of regions with wall shear stress WSS>7, average wall shear stress TAWSS<0.1, and oscillatory shear index OSI>0.25, respectively. WSS SA TAWSS SA OSI ; Step S3: Based on the training set, the deep neural network model is trained using K-fold cross-validation. The number of hidden layer nodes and the Dropout rate are optimized through grid search. The prediction error index is calculated based on the test set. The deep neural network model with the best generalization ability is selected as the final model for predicting the area proportion of unfavorable hemodynamic regions. Step S4: Using a dynamic optimization strategy that combines Bayesian optimization and genetic algorithm, the output parameters of the final model are optimized for multiple objectives to obtain the optimized spiral structure parameters of the artificial blood vessel; Specifically, it includes: Step S4.1: Construct the multi-objective optimization function: ； in, α , β , γ These are the weighting coefficients, and the objective is to minimize them. The weighting coefficient is used to reduce the risk of venous anastomotic stenosis, and the weighting coefficient can be dynamically adjusted according to clinical feedback to adapt to the patient's blood flow environment needs. Step S4.2: Perform dynamic optimization using a combination of Bayesian optimization and genetic algorithm, including: Bayesian optimization phase: Define the parameter space of the helical structure as: Parameter={ P , H ,A}, where, P ∈[ , ]、 H ∈[ , ]、A∈[ , ];in, This is the set minimum pitch value. The maximum set pitch. The minimum ridge height is set. The maximum value of the ridge height is set. This is the minimum set matching angle. The maximum value of the matching angle is set; a Gaussian process surrogate model is constructed, using the Matérn kernel function, and the acquisition function is the desired improved EI, iterating N times; Genetic Algorithm Stage: Individuals with a predetermined fitness level are selected from the Bayesian optimization results as the initial population; offspring are generated through crossover and mutation operations, and the individuals with the highest fitness are retained according to the fitness function. M Individual; Dynamic collaborative strategy: per iteration T After the genetic algorithm, the current optimal parameter set is added to the Bayesian optimized observation dataset to update the surrogate model; the termination condition is the rate of change Δ of the fitness function. f < Or the total number of iterations reaches the maximum number of iterations. This is the threshold parameter.

2. The method for optimizing the spiral structure parameters of artificial blood vessels according to claim 1, characterized in that, Step S1 specifically includes: Step S1.1: Construct an anastomosis model of a spiral artificial blood vessel with different spiral structural parameters and a natural vein using 3D modeling software; Step S1.2: Perform computational fluid dynamics (CFD) calculations on the anastomosis models of each spiral artificial blood vessel and natural vein, and extract the area ratio of the unfavorable hemodynamic region at the anastomosis site of each spiral artificial blood vessel and natural vein. Step S1.3: After normalizing the spiral structure parameters and the corresponding area proportions of unfavorable hemodynamic regions, filtering out noise, and extracting features, the dataset is obtained; Step S1.4: Divide the dataset into training sets according to a preset ratio. With test set .

3. The method for optimizing the spiral structure parameters of artificial blood vessels according to claim 2, characterized in that, The deep neural network model in step S2 includes: an input layer, a feature interaction layer, a self-attention mechanism, n residual block layers, and an output layer. The input layer is used to perform Min-Max standardization on the spiral structure parameters and output a standardized feature vector X. The feature interaction layer is used to expand the standardized feature vector X using a second-order polynomial to generate a feature set Φ(X) containing interaction terms. The self-attention mechanism is used to adjust the feature set based on the extended feature set. We obtain the weighted features by calculating the attention weights in the attention weight matrix; The n residual block layers are used to output a high-level semantic representation of the hemodynamic prediction task based on weighted features; The output layer is used to obtain the final prediction result based on the high-level semantic representation.

4. The method for optimizing the spiral structure parameters of artificial blood vessels according to claim 3, characterized in that, Step S3 specifically includes: Step S3.1: Obtain the training set Divide the data into K=5 mutually exclusive subsets and define the k-th fold verification set. and training set ,satisfy: k=1,2,...,5; Step S3.2: Update the parameters of the deep neural network model using the AdamW optimizer, as shown in the following formula: ； in, The parameters are updated at time t. Here are the parameters before the update at time t-1, and η is the learning rate. and For estimation of first-order and second-order matrices, λ is the weight decay coefficient; Step S3.3: Calculate the difference between the model's predicted value and the true value using the Huber loss function, and output the loss value L; Step S3.4: Based on the loss value L, optimize the hyperparameter combination through grid search to determine the optimal hyperparameters; Step S3.5: Based on the optimal hyperparameters and test set The deep neural network model was tested based on the mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The performance of deep neural network models was evaluated, and the deep neural network model with the best generalization ability was selected as the final model for predicting hemodynamic parameters.

5. The method for optimizing the spiral structure parameters of artificial blood vessels according to claim 1, characterized in that, It also includes the following steps: verifying the hemodynamic performance of the three-dimensional artificial blood vessel model generated using the optimized spiral structure parameters through CFD simulation; if the simulation verification is incorrect, the parameter is removed and resampled.

6. A system for optimizing the parameters of an artificial blood vessel spiral structure, characterized in that, It includes the following units: dataset acquisition unit, deep neural network model building unit, model training and testing unit, and multi-objective optimization unit; The dataset acquisition unit is used to construct a dataset with corresponding spiral structure parameters and the area ratio of unfavorable hemodynamic regions, and divide the dataset into a training set and a test set. The deep neural network model building unit is used to build a deep neural network model with helical structure parameters as input and the area ratio of unfavorable hemodynamic regions as output. The helical structure parameters include pitch P, ridge height H, and anastomosis angle A. The area ratio of unfavorable hemodynamic regions includes the area ratios of regions with wall shear stress WSS>7, average wall shear stress TAWSS<0.1, and oscillatory shear index OSI>0.25, respectively. WSS SA TAWSS SA OSI ; The model training and testing unit is used to train the deep neural network model using K-fold cross-validation based on the training set, optimize the number of hidden layer nodes and the Dropout rate through grid search, calculate the prediction error index based on the test set, and select the deep neural network model with the best generalization ability as the final model for predicting the area proportion of unfavorable hemodynamic regions. The multi-objective optimization unit employs a dynamic optimization strategy combining Bayesian optimization and genetic algorithms to perform multi-objective optimization on the output parameters of the final model, obtaining the optimized helical structure parameters of the artificial blood vessel; specifically, it performs the following: Step S4.1: Construct the multi-objective optimization function: ； in, α , β , γ These are the weighting coefficients, and the objective is to minimize them. The weighting coefficient is used to reduce the risk of venous anastomotic stenosis, and the weighting coefficient can be dynamically adjusted according to clinical feedback to adapt to the patient's blood flow environment needs. Step S4.2: Perform dynamic optimization using a combination of Bayesian optimization and genetic algorithm, including: Bayesian optimization phase: Define the parameter space of the helical structure as: Parameter={ P , H ,A}, where, P ∈[ , ]、 H ∈[ , ]、A∈[ , ];in, This is the set minimum pitch value. The maximum set pitch. The minimum ridge height is set. The maximum value of the ridge height is set. This is the minimum set matching angle. The maximum value of the matching angle is set; a Gaussian process surrogate model is constructed, using the Matérn kernel function, and the acquisition function is the desired improved EI, iterating N times; Genetic Algorithm Stage: Individuals with a predetermined fitness level are selected from the Bayesian optimization results as the initial population; offspring are generated through crossover and mutation operations, and the individuals with the highest fitness are retained according to the fitness function. M Individual; Dynamic collaborative strategy: per iteration T After the genetic algorithm, the current optimal parameter set is added to the Bayesian optimized observation dataset to update the surrogate model; the termination condition is the rate of change Δ of the fitness function. f < Or the total number of iterations reaches the maximum number of iterations. This is the threshold parameter.

7. An artificial blood vessel, characterized in that, Its spiral structure parameters are obtained by the method for optimizing the spiral structure parameters of artificial blood vessels according to any one of claims 1-5.

8. The artificial blood vessel according to claim 7, characterized in that, The artificial blood vessel has a diameter of 6 mm and spiral structure parameters of 29.2 mm pitch, 1.75 mm ridge height, and 38.6° anastomosis angle.

9. The artificial blood vessel according to claim 7, characterized in that, The fabrication process of the artificial blood vessel includes the following: First, a sugar helical structure is generated by melt spinning maltitol. Then, using electrospinning technology, an optimized spiral artificial blood vessel is prepared by spinning an electrospun fiber skeleton using biopolymer materials as raw materials; wherein, the biopolymer materials are one or a mixture of several of them in any proportion, including polycaprolactone PCL, poly-L-lactide-caprolactone PLCL, polylactic acid PLA, polyglycolic acid PGA, polylactic acid-glycolic acid copolymer PLGA, polyurethane PU and its modified copolymers.

Images