Intelligent planning method and system for arteriovenous fistula based on hemodynamic simulation

Abstract

This invention relates to the fields of medical image processing and medical auxiliary decision-making technology, specifically to an intelligent planning method and system for arteriovenous fistulas (AVFs) based on hemodynamic simulation. The method first acquires three-dimensional spatial anatomical data of the patient's arteries and veins, constructs a three-dimensional vascular tree model, and inputs vascular parameter data and hemodynamic parameter data under different blood flow velocities into a deep learning model to obtain dynamic shear force data at the anastomosis site. This dynamic shear force data is then compared with preset optimal anastomosis angle data to determine the optimal anastomosis angle. Combining the optimal anastomosis angle, a three-dimensional model of the AVF surgical procedure is created using the vascular tree model, resulting in a three-dimensional AVF model. A virtual surgery is then performed based on this model, obtaining virtual surgical process data and postoperative hemodynamic status assessment results. This method is expected to increase the success rate of a single AVF surgery from the current 60%–65% to 85%–90%.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and medical auxiliary decision-making technology, specifically to an intelligent planning method and system for arteriovenous fistulas based on hemodynamic simulation, which is particularly suitable for planning the establishment of vascular access for hemodialysis patients. Background Technology

[0002] Arteriovenous fistulas (AVFs) are the preferred form of long-term vascular access for hemodialysis patients. The procedure involves directly anastomosing the patient's own artery and vein, causing the vein to dilate and arterialize under high-pressure blood flow, thus creating a vascular access suitable for repeated punctures. However, in clinical practice, the success rate of a single AVF procedure is approximately 60%–65%, with an early failure rate as high as 30%, and about 40% of AVFs requiring repeat surgery within one year to maintain function.

[0003] Currently, arteriovenous fistula (AVF) surgery planning relies primarily on the surgeon's experience and simple vascular measurements (such as vessel diameter and blood flow velocity), lacking a systematic consideration of hemodynamic factors. Although studies have shown that the hemodynamic characteristics at the anastomosis site, particularly the distribution of wall shear forces, are closely related to AVF maturity and long-term patency, existing techniques have limitations in the following aspects:

[0004] 1. Although traditional computational fluid dynamics (CFD) software can perform blood flow simulation, the calculation time is long, making it difficult to apply to real-time clinical decision-making;

[0005] 2. Existing methods have difficulty accurately capturing the dynamic process of blood-vessel wall interaction, especially the differences at different anastomosis angles;

[0006] 3. There is a lack of effective methods to combine changes in microscopic shear force with long-term prediction of arteriovenous fistula function;

[0007] 4. It is impossible to achieve individualized arteriovenous fistula planning and optimization, making it difficult to meet the special needs of different patients.

[0008] Therefore, there is an urgent need to develop a fistula planning method that combines hemodynamic simulation and intelligent prediction to improve the success rate of fistula surgery and long-term patency. Summary of the Invention

[0009] The purpose of this invention is to provide an intelligent planning method and system for arteriovenous fistulas based on hemodynamic simulation, aiming to solve the problems of lack of individualization in fistula planning and difficulty in accurately predicting long-term function in the prior art.

[0010] This invention proposes an intelligent planning method for arteriovenous fistulas based on hemodynamic simulation, including:

[0011] Obtain three-dimensional spatial anatomical data of the patient's arteries and veins, and construct a three-dimensional model of the vascular tree;

[0012] The vascular parameter data of the three-dimensional vascular tree model and the hemodynamic parameter data under different blood flow velocities are combined and input into the deep learning model to obtain the dynamic shear force data of the anastomosis.

[0013] The dynamic shear force data of the anastomosis joint is compared with the preset optimal anastomosis angle data to determine the optimal anastomosis angle;

[0014] The optimal anastomosis angle is used to perform a three-dimensional modeling of the arteriovenous fistula procedure on the three-dimensional model of the vascular tree, resulting in a three-dimensional arteriovenous fistula model.

[0015] Based on the three-dimensional arteriovenous fistula model, a virtual surgery was performed to obtain virtual surgical process data and postoperative hemodynamic status assessment results.

[0016] Preferably, the three-dimensional spatial anatomical data of the patient's arteries and veins are obtained, and the three-dimensional model of the vascular tree is constructed, specifically including:

[0017] Acquire raw 3D vascular data from high-resolution ultrasound or computed tomography angiography.

[0018] The original three-dimensional blood vessel data is input into a two-stage deformable 3D convolutional neural network based on a self-attention mechanism for processing to obtain the three-dimensional model of the blood vessel tree.

[0019] The self-attention mechanism is used to enhance the accuracy of identifying key branch points of blood vessels.

[0020] Preferably, the construction process of the two-stage deformable 3D convolutional neural network based on the self-attention mechanism includes:

[0021] A two-stage deformable 3D convolutional neural network architecture based on self-attention mechanism is constructed, wherein the first stage obtains the coarse segmentation results of blood vessels, and the second stage refines the processing of blood vessel boundaries and branch regions.

[0022] Obtain a training dataset containing three-dimensional blood vessel data and corresponding labels, wherein the labels include information on blood vessel boundaries, centerlines, and branch point locations;

[0023] The training dataset is used to train the two-stage deformable 3D convolutional neural network based on the self-attention mechanism to obtain the trained neural network model.

[0024] Preferably, the vascular parameter data of the 3D vascular tree model and the hemodynamic parameter data under different blood flow velocities are combined and input into a deep learning model to obtain the dynamic shear force data of the anastomosis, specifically including:

[0025] The three-dimensional model of the vascular tree is divided into different regions, including the anastomosis core region, the anastomosis edge region, the transition region and the distal region, and different mesh densities are assigned to each region.

[0026] The pressure field, velocity field, and turbulence field of the vascular tree three-dimensional model were calculated using computational fluid dynamics software, and hemodynamic parameters were set for different blood flow velocities.

[0027] The hemodynamic parameters at different blood flow rates were input into an improved smooth particle hydrodynamics method for simulation to obtain shear force data of the anastomosis region at different blood flow rates.

[0028] The vascular parameter data is used as input to the deep learning model to obtain the dynamic shear force data of the anastomosis.

[0029] Preferably, the improved smooth particle hydrodynamics method includes:

[0030] Based on the morphological characteristics of the vascular tree 3D model, an adaptive particle distribution is automatically generated;

[0031] Increasing particle density in the anastomosis region improves simulation accuracy;

[0032] An adaptive kernel function is used, and the kernel radius is dynamically adjusted according to the local particle density.

[0033] An improved virtual particle method is used to handle the blood vessel wall boundary, improving the stability of the calculation at the boundary.

[0034] Considering the elastic deformation response of the blood vessel wall, a fluid-structure interaction simulation is achieved.

[0035] Preferably, the dynamic shear force data of the anastomosis site is compared with the preset optimal anastomosis angle data to determine the optimal anastomosis angle, specifically including:

[0036] A deep correlation model between dynamic time-varying shear force and blood vessel wall response was constructed, including a dual-flow feature extraction network and a multi-scale fusion module;

[0037] The dynamic shear force data of the anastomosis is input into the depth correlation model to obtain the long-term blood vessel wall response prediction results under different anastomosis angles;

[0038] The Bayesian optimization framework based on physical information enhancement evaluates each candidate anastomosis angle scheme and calculates a multi-objective score including surgical success rate, fistula maturation time and long-term patency rate.

[0039] The optimal anastomosis angle is determined based on the multi-objective score and the patient's specific needs.

[0040] Preferably, the Bayesian optimization framework based on physical information enhancement includes:

[0041] By integrating hemodynamic theory and clinical experience, a theoretical correlation model between anastomosis angle and arteriovenous fistula function was established.

[0042] Define physical feasibility constraints, including angular range and spatial location limitations;

[0043] Construct a multi-objective function and dynamically adjust the weight coefficients of short-term, medium-term, and long-term objectives;

[0044] Initialize candidate matching angle schemes and use a Gaussian process regression model to predict the performance of each scheme;

[0045] Based on the expected improvement and the trade-off between uncertainty, iterative optimization is performed until the optimal solution is converged.

[0046] Preferably, based on the three-dimensional arteriovenous fistula model, a virtual surgery is performed to obtain the virtual surgical process data and the postoperative hemodynamic status assessment results, specifically including:

[0047] A virtual surgical scene is constructed based on the optimal anastomosis angle and the three-dimensional model of the vascular tree.

[0048] Simulate the surgical procedure and generate postoperative vascular morphology;

[0049] Predict the status of the arteriovenous fistula at different time points, including one week, one month, and six months later;

[0050] Calculate the predicted values ​​for the maturation time and long-term patency rate of the arteriovenous fistula;

[0051] Generate a visual surgical guidance plan, including anastomosis angles, location markers, and key operational points.

[0052] As a preferred option, an adaptive closed-loop feedback optimization step is also included:

[0053] Record actual surgical parameters and postoperative follow-up data;

[0054] Compare the predicted results with the actual results to quantify the accuracy of the prediction;

[0055] Update the model parameters based on the difference between the prediction and the actual results;

[0056] Add new case data to the knowledge base and optimize sampling strategies;

[0057] Anomaly detection and analysis are performed on cases where predictions fail to improve system robustness.

[0058] An intelligent arteriovenous fistula planning system based on hemodynamic simulation includes:

[0059] The 3D model building module for the vascular tree is used to acquire the 3D spatial anatomical data of the patient's arteries and veins and build a 3D model of the vascular tree.

[0060] The hemodynamic simulation module is used to combine the vascular parameter data of the three-dimensional vascular tree model with the hemodynamic parameter data under different blood flow velocities, and input them into the deep learning model to obtain the dynamic shear force data of the anastomosis.

[0061] The optimal anastomosis angle determination module is used to compare the dynamic shear force data of the anastomosis with the preset optimal anastomosis angle data to determine the optimal anastomosis angle.

[0062] The arteriovenous fistula 3D modeling module is used to perform arteriovenous fistula 3D modeling on the vascular tree 3D model in combination with the optimal anastomosis angle, so as to obtain a 3D arteriovenous fistula model.

[0063] The virtual surgery module is used to perform virtual surgery based on the three-dimensional arteriovenous fistula model, and obtain virtual surgical process data and postoperative hemodynamic status assessment results.

[0064] The adaptive closed-loop feedback module is used to record actual surgical parameters and postoperative follow-up data, compare the predicted results with the actual results, update the model parameters, and improve the prediction accuracy of the system.

[0065] This invention innovatively integrates hemodynamic simulation, deep learning, and multi-objective optimization technologies to achieve intelligent management of the entire process, from acquiring individual patient vascular data to generating the optimal surgical plan. Particularly in the prediction of anastomotic shear force and determination of the optimal anastomosis angle, this invention proposes a series of technological innovations, making arteriovenous fistula planning more precise and efficient.

[0066] The present invention has the following beneficial effects:

[0067] 1. Significantly improve the success rate of arteriovenous fistula surgery in one procedure, expected to increase from the current 60%–65% to 85%–90%, and reduce additional surgical trauma for patients;

[0068] 2. Reduce the early failure rate of arteriovenous fistulas from 30% to below 15%, thereby improving the efficiency of medical resource utilization;

[0069] 3. Extends the lifespan of arteriovenous fistulas, increasing the one-year patency rate from 60% to over 80%, and reducing the need for maintenance interventions;

[0070] 4. Shorten arteriovenous fistula planning time from 2-3 hours using traditional methods to within 30 minutes, improving clinical work efficiency;

[0071] 5. Reduce reliance on expert doctors, enabling primary hospitals to provide high-quality arteriovenous fistula creation services and promoting a more balanced distribution of medical resources. Attached Figure Description

[0072] Figure 1 This is a flowchart of the intelligent planning method for arteriovenous fistulas based on hemodynamic simulation of the present invention;

[0073] Figure 2 This is a structural diagram of the vascular tree 3D model construction module of the present invention;

[0074] Figure 3 This is a schematic diagram of the multi-scale hemodynamic parameter adaptive acquisition system of the present invention;

[0075] Figure 4 This is a framework diagram of the improved smooth particle hydrodynamics method of the present invention;

[0076] Figure 5 This is a diagram of the deep correlation model between dynamic time-varying shear force and blood vessel wall response of the present invention;

[0077] Figure 6 This is a flowchart of the Bayesian optimization framework based on physical information enhancement of this invention.

[0078] Figure 7 This is the overall system architecture diagram of the present invention;

[0079] Figure 8 This is an example diagram of the three-dimensional arteriovenous fistula model of the present invention;

[0080] Figure 9 This is a comparison chart of the virtual surgical planning and actual surgical results of this invention;

[0081] Figure 10 This is a schematic diagram of the adaptive closed-loop feedback optimization system of the present invention. Detailed Implementation

[0082] Please refer to the attached document. Figure 1-10 The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0083] like Figure 1 As shown, the intelligent arteriovenous fistula planning method based on hemodynamic simulation provided by this invention includes the following steps:

[0084] Step S1: Acquire three-dimensional spatial anatomical data of the patient's arteries and veins, and construct a three-dimensional model of the vascular tree. This step first acquires the patient's original three-dimensional vascular data using medical imaging techniques such as high-resolution ultrasound or computed tomography angiography (CTA). Preferably, for ultrasound imaging, a high-frequency probe of 7-12MHz is used, with a scanning resolution of 0.1mm to ensure the capture of minute vascular structures; for CTA imaging, spiral CT is preferred, with a slice thickness not exceeding 1mm, to obtain sufficiently detailed vascular anatomy information.

[0085] After acquiring the raw data, it is input into a two-stage deformable 3D convolutional neural network based on a self-attention mechanism for processing, resulting in a 3D model of the vascular tree. For example... Figure 2 As shown, the network has a two-stage structure: the first stage acquires coarse segmentation results of blood vessels, and the second stage refines the processing of vessel boundaries and branching regions. A self-attention mechanism is used to enhance the accuracy of identifying key vascular branching points, especially in the arteriovenous crossing region, which is crucial for determining the location of subsequent arteriovenous fistula anastomoses.

[0086] In a preferred embodiment of the present invention, the construction process of the two-stage deformable 3D convolutional neural network based on the self-attention mechanism includes:

[0087] First, a two-stage deformable 3D convolutional neural network architecture based on a self-attention mechanism is constructed. This network architecture employs an encoder-decoder structure. The encoder consists of four downsampling layers, each containing two 3D convolutional layers (kernel size 3×3×3), a batch normalization layer, and a ReLU activation function. Deformable convolutions are introduced in the second, third, and fourth downsampling layers to enhance the network's adaptability to irregular blood vessel shapes. The self-attention module is located in the deepest layer of the encoder and is used to capture global vascular structure information. The decoder consists of four upsampling layers, employing transposed convolutions to amplify the feature maps and fusing features from the corresponding encoder layers through skip connections.

[0088] Secondly, a training dataset containing 3D vascular data and corresponding labels was obtained. The training dataset includes 3D vascular data from 500 different patients and their corresponding vascular segmentation labels. The labels include information on vascular boundaries, centerlines, and branch point locations, and were jointly annotated and cross-validated by three experienced vascular surgeons to ensure label quality. In this dataset, 70% was used for training, 15% for validation, and 15% for testing.

[0089] Finally, the self-attention-based two-stage deformable 3D convolutional neural network was trained using the training dataset. The training process employed a combined loss function, including Dice loss and weighted cross-entropy loss, to balance the segmentation accuracy of blood vessels and background regions. The loss function can be expressed as:

[0090] L total =α·L Dice +(1-α)·L WCE ,

[0091] Where α is the weighting coefficient, with a value of 0.7, used to balance the importance of the two losses; L Dice The Dice loss measures the overlap between the predicted segmentation result and the true label, and is calculated using the following formula:

[0092]

[0093] Where, p i Let g be the predicted probability of the i-th voxel. i L is the true label (0 or 1) of the i-th voxel, and N is the total number of primes; WCE The weighted cross-entropy loss is used to address the class imbalance between blood vessels and the background. The calculation formula is as follows:

[0094]

[0095] Among them, w pos The weight for positive samples (vascular regions) is set to the ratio of background to vascular voxels, typically around 4.0; w neg The weight for negative samples (background area) is set to 1.0.

[0096] In arteriovenous fistula (AVF) planning scenarios, accurate vessel segmentation is crucial for subsequent hemodynamic analysis. Traditional segmentation methods often struggle to accurately identify small vessels and bifurcation areas. For example, the radial and ulnar arteries and cephalic veins, with diameters less than 3 mm, are commonly used for forearm AVF creation, and their accurate segmentation directly impacts the selection of the fistula anastomosis location. The two-stage deformable network structure of this invention effectively solves this problem. Through deformable convolution, it adapts to irregular vessel shapes, achieving an accuracy 10%–15% higher than traditional convolutional networks, especially for pathological changes such as stenosis, tortuosity, or twisting.

[0097] Training employed the Adam optimizer with an initial learning rate of 0.001 and a learning rate decay strategy, decreasing the rate by 0.1 times every 50 epochs. To prevent overfitting, weight decay (with a coefficient of 0.0001) and an early stopping strategy were used (training stopped if there was no improvement for 20 consecutive epochs on the validation set). Training was performed on a workstation equipped with an NVIDIA RTX 3090 GPU, with a batch size of 4, for a total of 300 epochs.

[0098] Through the above training process, a neural network model with high-precision vessel segmentation capability was obtained. This model achieved a Dice coefficient of over 95% for vessel segmentation on the test set, with a particularly high accuracy of 92% in identifying vessel branch regions, significantly outperforming traditional segmentation methods. In practical applications, this model can complete 3D reconstruction of the vascular tree for a patient within 2-3 minutes, providing a foundation for subsequent arteriovenous fistula planning.

[0099] Step S2: Combine the vascular parameter data of the 3D vascular tree model with the hemodynamic parameter data under different blood flow velocities, and input them into the deep learning model to obtain the anastomotic dynamic shear force data. In this step, the 3D vascular tree model is first divided into different regions, including the anastomotic core region, the anastomotic edge region, the transition region, and the distal region, and different mesh densities are assigned to each region. Figure 3 As shown, the multi-scale hemodynamic parameter adaptive acquisition system automatically allocates grid density based on regional importance: 10 μm for the core anastomosis region, 25 μm for the anastomosis edge region, 50 μm for the transition region, and 100 μm for the distal region. This adaptive grid density allocation strategy can significantly reduce the overall computational complexity while ensuring the computational accuracy of key regions.

[0100] Next, computational fluid dynamics software (such as ANSYS CX) is used to calculate the pressure, velocity, and turbulence fields of the vascular tree 3D model, setting hemodynamic parameters for different blood flow velocities. In practical applications, typically 5-7 different flow velocity values ​​are set, covering the range of blood flow changes from resting to active states (e.g., 20 cm / s to 100 cm / s). For each flow velocity value, corresponding boundary conditions and physical parameters are set, including inlet flow rate, outlet pressure, and blood density (approximately 1060 kg / m³). 3 ) and viscosity (approximately 0.0035 Pa·s).

[0101] Then, hemodynamic parameters at different blood flow velocities were input into an improved smoothed particle hydrodynamics (SPH) method for simulation to obtain shear force data of the anastomotic region at different blood flow velocities. For example... Figure 4 As shown, the improved SPH method of the present invention includes the following key technical features:

[0102] (1) Based on the morphological characteristics of the vascular tree 3D model, an adaptive particle distribution is automatically generated. The particle spacing is set to approximately 10 μm in the core area of ​​the anastomosis and gradually transitions to 100 μm towards the distal end to ensure sufficient computational accuracy in critical areas;

[0103] (2) Increase particle density in the anastomosis region to improve simulation accuracy. The particle number density at the anastomosis region is 2-3 times higher than that of the standard SPH method to capture the complex flow characteristics of this region;

[0104] (3) An adaptive kernel function is used, and the kernel function radius is dynamically adjusted according to the local particle density. The kernel function radius h is calculated using the following formula:

[0105]

[0106] Where h is the kernel function radius at the current position, in μm; h0 is the reference kernel function radius, usually taken as 1.2 times the initial particle spacing, in μm; and ρ0 is the reference density, taken as the particle density when initially uniformly distributed, in particles / mm². 3 ρ represents the local particle density, expressed in particles per mm. 3 The exponent of 1 / 3 originates from the inverse relationship between density and distance in three-dimensional space. This adaptive kernel function can expand its effective range in sparse particle regions and improve local accuracy in dense particle regions.

[0107] (4) An improved virtual particle method is used to handle the vessel wall boundary, improving the stability of the calculation at the boundary. Based on the traditional virtual particle method, a wall normal gradient correction term C is added. grad :

[0108]

[0109] Among them, C grad d is the gradient correction coefficient, dimensionless; d is the distance from the particle to the wall, in μm; d0 is the characteristic distance, which is the average particle spacing, in μm. This correction term gradually reduces the gradient calculation weight near the wall, avoiding the instability problem of the traditional SPH method at the solid-liquid interface.

[0110] (5) Considering the elastic deformation response of the blood vessel wall, a fluid-structure interaction simulation is achieved. The blood vessel wall adopts an elastic thin film model, and its deformation follows the following relationship:

[0111] σ=E·ε,

[0112] Where σ represents stress in Pa; E represents Young's modulus, ranging from 0.5 to 1.5 MPa for arterial walls and 0.2 to 0.6 MPa for venous walls; and ε represents strain, which is dimensionless. In actual calculations, Young's modulus is dynamically adjusted based on the patient's age and vessel type (artery or vein). For example, for patients over 60 years of age, the upper limit of Young's modulus for arterial walls is typically used (approximately 1.5 MPa), and for venous walls, it is 0.4 to 0.6 MPa; while for younger patients under 40 years of age, it is 0.5 to 0.8 MPa for arterial walls and 0.2 to 0.3 MPa for venous walls. This difference reflects age-related decrease in vascular elasticity, which has a significant impact on the maturation process of arteriovenous fistulas.

[0113] The improved SPH method described above allows for the acquisition of detailed shear force distribution in the anastomotic region under different blood flow velocities. Wall shear force (WSS) is calculated using the following formula:

[0114]

[0115] Where, τ w The wall shear force is expressed in Pa; μ is the blood dynamic viscosity, expressed in Pa. This represents the velocity gradient at the wall (y=0), in seconds. -1 Within the SPH framework, the velocity gradient is calculated using the velocity difference between adjacent particles:

[0116]

[0117] Where, m j ρ is the mass of the j-th particle, in kg. j The density at the j-th particle is expressed in kg / m³. 3 u j and u i These are the velocities of the j-th particle and the i-th particle near the wall, respectively, in m / s; The gradient of the kernel function, in units of m. -4 ;r ij The distance between the two particles is expressed in meters (m).

[0118] In arteriovenous fistula planning, wall shear force is a key indicator for evaluating the rationality of anastomosis design. Clinical studies have shown that excessively low shear force (<1.5 Pa) easily leads to intimal hyperplasia and stenosis, while excessively high shear force (>40 Pa) may cause vascular injury and thrombosis. An ideal anastomosis design should ensure uniform shear force distribution within a moderate range (approximately 4–10 Pa). The SPH method of this invention can accurately capture the shear force distribution characteristics at different anastomosis angles, providing a quantitative basis for optimal angle selection.

[0119] Finally, the vascular parameter data was used as input to the deep learning model to obtain the dynamic shear force data of the anastomosis. The vascular parameter data included 27 feature parameters such as vessel diameter, wall thickness, elastic modulus, and branch angle. The deep learning model employed a Temporal Convolutional Network (TCN) structure, which effectively captured the temporal pattern of time-varying shear force. The mathematical expression of the TCN model is:

[0120] F(x) = ReLU(W) f *x+b f ),

[0121] Where F(x) is the feature map output; x is the input sequence with dimensions (batch size, sequence length, number of features); W fThis is the convolution kernel weight matrix; * indicates the convolution operation; b f is the bias term; ReLU is the activation function, defined as ReLU(z) = max(0,z).

[0122] To capture long-term dependencies, TCN employs a dilated convolutional structure:

[0123] F d (x)=ReLU(W d * d x+b d ),

[0124] Among them, F d (x) represents the output of the dilated convolution; * d This represents a convolution operation with a dilation factor of d. The dilation factor controls the size of the receptive field and increases exponentially with the number of layers (e.g., 1, 2, 4, 8).

[0125] The model was trained using paired data consisting of 1000 sets of different vascular parameters and corresponding SPH simulation results. The loss function employed was a combination of mean squared error and smoothed L1 loss.

[0126] L=β·MSE+(1-β)·SmoothL1,

[0127] Where β is the weighting coefficient with a value of 0.6; MSE is the mean squared error; and SmoothL1 is the smoothing L1 loss, which is insensitive to outliers.

[0128] This deep learning model is of significant value in arteriovenous fistula planning. For example, for patients requiring evaluation of multiple potential anastomosis sites, traditional methods necessitate complete CFD simulations for each site, which can take several hours; while this model can predict the shear force distribution at each site within seconds, enabling physicians to quickly select the optimal approach. The model's prediction error is controlled within 7%, meeting the needs of clinical decision-making.

[0129] After training, this deep learning model can directly predict the dynamic shear force distribution in the anastomosis area based on the input vascular parameters, avoiding the need for time-consuming SPH full flow field simulation each time, reducing the calculation time from hours to seconds, and greatly improving the system response speed.

[0130] Step S3: Compare the dynamic shear force data of the anastomosis with the preset optimal anastomosis angle data to determine the optimal anastomosis angle. In this step, a deep correlation model between dynamic time-varying shear force and vessel wall response is first constructed, including a dual-flow feature extraction network and a multi-scale fusion module. Figure 5As shown, the model features a dual-stream architecture: one stream processes time-varying shear force data, and the other processes vessel wall property data. The time-varying shear force feature extraction branch employs a three-layer temporal convolution module, each layer containing dilated convolution, residual connections, and layer normalization; the vessel wall feature extraction branch uses a multilayer perceptron structure to extract vessel wall thickness, elasticity, and geometric features. The two feature streams are fused through a cross-attention mechanism to form a correlated representation of shear force and vessel wall response.

[0131] The mathematical expression for the cross-attention mechanism is:

[0132]

[0133] Where A(Q,K,V) is the attention output; Q is the query matrix, from the first feature path; K is the key matrix, from the second feature path; V is the value matrix, also from the second feature path; and T is the matrix transpose; d k Let x be the dimension of the key vector; softma x is the normalization function defined as follows:

[0134] By inputting the dynamic shear force data of the anastomosis site into the aforementioned depth correlation model, long-term vessel wall response prediction results under different anastomosis angles were obtained. The model output includes key indicators such as the degree of endothelial cell proliferation, the rate of change of lumen diameter, and the probability of patency. These indicators are directly related to the maturation time and long-term function of the arteriovenous fistula.

[0135] Next, a Bayesian optimization framework based on physical information enhancement was used to evaluate each candidate anastomosis angle scheme, and a multi-objective score was calculated, including surgical success rate, fistula maturation time, and long-term patency rate. For example... Figure 6 As shown, the core of this optimization framework is to integrate hemodynamic theory and clinical experience to establish a theoretical correlation model between anastomosis angle and arteriovenous fistula function.

[0136] In a preferred embodiment of the present invention, the Bayesian optimization framework based on physical information enhancement includes:

[0137] (1) Integrating hemodynamic theory and clinical experience, a theoretical correlation model between anastomosis angle and arteriovenous fistula function was established. This model considers the influence of anastomosis angle on flow field distribution, wall shear force, and thrombosis risk. For example, studies have shown that an end-to-side anastomosis angle of 25°-45° can usually achieve a good balance between flow stability and shear force distribution;

[0138] (2) Define physical feasibility constraints, including angle range and spatial location limitations. Considering surgical operability, the anastomosis angle is usually limited to the range of 15°-60°; the spatial location needs to take into account local anatomical structures and avoid important tissues such as surrounding nerves and tendons;

[0139] (3) Construct a multi-objective function and dynamically adjust the weight coefficients of short-term, medium-term, and long-term objectives. The expression for the multi-objective function is:

[0140] F(θ)=w1·f1(θ)+w2·f2(θ)+w3·f3(θ),

[0141] Where F(θ) is the comprehensive scoring function, with a score range of 0-1, and a higher score indicates a better solution; θ represents the anastomosis angle, in degrees (°); f1(θ) represents the uniformity of shear force at the anastomosis joint, calculated using the following formula:

[0142]

[0143] Where, σ τ The standard deviation of shear force in the anastomosis region is expressed in Pa. The average shear force is expressed in Pa. This indicator reflects the success rate of the surgery; a uniform shear force distribution helps reduce the risk of vascular damage caused by localized high-shear areas.

[0144] f2(θ) represents the predicted maturation time of the arteriovenous fistula, and the calculation formula is:

[0145]

[0146] Among them, T m (θ) represents the predicted maturation time of the arteriovenous fistula, in days; T0 is the reference time, taken as 42 days (6 weeks, the clinically considered ideal maturation time). The closer this index is to 1, the closer the predicted maturation time is to the ideal value;

[0147] f3(θ) represents the long-term accessibility prediction, which directly uses the 6-month accessibility probability value output by the deep correlation model, ranging from 0 to 1;

[0148] w1, w2, and w3 are dynamic weighting coefficients that are automatically adjusted based on individual patient characteristics, ensuring that w1 + w2 + w3 = 1. For elderly patients (>65 years old), w1 = 0.5, w2 = 0.3, and w3 = 0.2 are typically set, with a greater emphasis on short-term surgical success; for younger patients (<45 years old), w1 = 0.3, w2 = 0.2, and w3 = 0.5 are set, with a greater emphasis on long-term patency.

[0149] In arteriovenous fistula planning practice, different patient groups exhibit significant differences in the prioritization of various objectives. For example, diabetic patients, whose blood vessels are often more fragile and take longer to mature, have their weights adjusted to w1=0.4, w2=0.4, w3=0.2, focusing more on the maturation process. Conversely, patients requiring emergency dialysis may prioritize short-term success rates, with weights set to w1=0.6, w2=0.3, w3=0.1. This personalized weighting allows the system to generate the most suitable plan based on the patient's specific circumstances.

[0150] (4) Initialize candidate matching angle schemes and use a Gaussian process regression model to predict the performance of each scheme. The Gaussian process regression model is defined as:

[0151]

[0152] Where f(θ) represents the objective function to be optimized; This represents a Gaussian process; m(θ) is the mean function, initially set to a constant; k(θ,θ′) is the kernel function, using the Matérn5 / 2 kernel:

[0153]

[0154] Where σ is the signal standard deviation, usually initialized to 0.5; l is the length scale parameter, controlling the correlation decay rate, initially set to 5°; |θ-θ′| represents the absolute difference between two angles, in degrees (°). The parameters σ and l are continuously updated during the optimization process using the maximum likelihood estimation method.

[0155] The Matérn5 / 2 kernel excels at describing smooth but non-infinitely differentiable functions common in physical systems, making it particularly suitable for modeling discontinuous biological systems such as vascular responses. In practical applications, compared to the standard square exponential kernel, this kernel function is better able to capture hemodynamic abrupt changes caused by variations in anastomosis angles.

[0156] (5) Based on the trade-off between expected improvement and uncertainty, iterative optimization is performed until convergence to the optimal solution. An acquisition function is used to guide the optimization process; this invention uses an expected improvement (EI) acquisition function:

[0157] EI(θ)=E[max(f(θ)-f(θ + ),0)],

[0158] Where EI(θ) is the expected improvement value; E represents the expectation operation; f(θ) is the predicted value of the objective function at angle θ; f(θ) + The given optimal solution is denoted as . The EI function can be explicitly calculated as:

[0159]

[0160] Where μ(θ) and σ(θ) are the predicted mean and standard deviation of the Gaussian process at θ, respectively; Φ and φ are the cumulative distribution function and probability density function of the standard normal distribution, respectively. The acquisition function achieves larger values ​​in the regions of high predicted values ​​and high uncertainty, balancing the relationship between exploration and utilization.

[0161] During the arteriovenous fistula planning and optimization process, the initial stage involves extensive exploration of different angles within the range of 15°-60°. As iterations proceed, the optimization focus gradually shifts to promising regions. For example, the system may initially try four angle points: 15°, 30°, 45°, and 60°. If the 30°-45° range is found to perform well, more intensive sampling will be performed within this range to ultimately determine the optimal angle.

[0162] Through the Bayesian optimization process described above, the optimal anastomosis angle can usually be converged within 20-30 iterations. Finally, the optimal anastomosis angle is determined based on multi-objective scoring and specific patient needs. In practical applications, the common optimal anastomosis angle range is 30°-40°, but the specific value will be adjusted according to individual differences. For example, for hypertensive patients, the optimal angle is usually smaller (25°-35°) to reduce turbulence and vessel wall stress; while for patients with slower blood flow velocities, a larger angle (35°-45°) may be needed to promote sufficient blood flow mixing and shear force stimulation.

[0163] Step S4: Using the optimal anastomosis angle, perform 3D modeling of the arteriovenous fistula (AVF) procedure based on the 3D vascular tree model to obtain a 3D AVF model. After determining the optimal anastomosis angle, this step applies it to the 3D vascular tree model to perform 3D AVF procedure modeling. The modeling process considers actual operational factors during surgery, including:

[0164] (1) Shape of blood vessel incision: usually elliptical or rhomboid, with the long axis being about 1.5-2 times the diameter of the blood vessel;

[0165] (2) Anastomosis technique: set as continuous suture or interrupted suture, with suture spacing of approximately 0.5-1mm;

[0166] (3) Spatial location: Consider local anatomical limitations to ensure the feasibility of the anastomosis procedure.

[0167] 3D modeling employs a parametric approach, automatically generating the anastomosis structure based on parameters such as the optimal anastomosis angle, vessel diameter, and anastomosis length. For common end-to-side anastomosis, the modeling process can be represented as follows:

[0168] V AVF =P(θ) opt ,da ,d v ,l a ,S G ),

[0169] Among them, V AVF The generated 3D model of the arteriovenous fistula; P represents the parametric modeling operation; θ opt For the optimal fitting angle; d a and d v These are the diameters of the arteries and veins, respectively; a S is the length of the anastomosis. G For anatomical space geometric constraints.

[0170] For end-to-side anastomosis, the model first creates an elliptical incision on the venous side, then aligns the arterial end with the incision at the optimal anastomosis angle, and finally connects the two with sutures to form a complete anastomosis structure. In practice, the anastomosis length is usually set to 1.5-2 times the arterial diameter; this ratio is derived from hemodynamic optimization results and clinical experience.

[0171] After modeling is completed, the quality of the model is evaluated by checking the anastomosis morphology, lumen continuity, and spatial location rationality.

[0172] Evaluation indicators include:

[0173] Q geom =w a ·A ratio +w c ·C score +w s ·S access ,

[0174] Among them, Q geom For geometric quality rating; A ratio This is the ratio of the anastomosis area to the arterial cross-sectional area, with an ideal range of 1.5-2.0; C score Lumen continuity is scored based on curvature continuity at the anastomosis; S access For surgical accessibility scoring, surrounding tissue limitations should be considered; a w c and w s These are the corresponding weighting coefficients, typically set to 0.4, 0.3, and 0.3.

[0175] If necessary, fine-tuning is made to meet the actual needs of the surgery. For example, if the anastomosis location is limited by local anatomy (such as adjacent nerves or tendons), the position or angle is adjusted appropriately to ensure operative safety. The final three-dimensional arteriovenous fistula model is shown below. Figure 8 As shown, it includes complete postoperative vascular morphology and anastomotic structure details.

[0176] In clinical practice, different types of arteriovenous fistulas (such as radial artery-cephalic vein, brachial artery-cephalic vein, etc.) have their own modeling characteristics. For example, for forearm fistulas, due to the limited local anatomical space, special attention must be paid to the relationship between the anastomosis and surrounding structures during modeling; while for upper arm fistulas, there may be a risk of excessive blood flow, so the size of the anastomosis must be controlled during modeling to avoid excessive blood flow load. The parametric modeling method of this invention can be flexibly adjusted according to different situations to meet diverse clinical needs.

[0177] Step S5: Based on the 3D arteriovenous fistula model, perform virtual surgery to obtain virtual surgical process data and postoperative hemodynamic status assessment results. In this step, a virtual surgical scene is constructed based on the optimal anastomosis angle and the 3D vascular tree model. The virtual surgical scene includes elements such as patient anatomy, surgical instruments, and operational steps, presented through 3D visualization technology. The scene construction formula is:

[0178] S virtual =C(V) AVF A patient ,I surgical ),

[0179] Among them, S virtual For a virtual surgical scene; C represents the scene construction operation; V AVF A three-dimensional arteriovenous fistula model; A patient For the patient's anatomical data; I surgical Information on surgical instruments and procedures.

[0180] The surgical procedure is simulated to generate postoperative vascular morphology. The simulation follows a standard surgical procedure: vascular exposure, vascular incision, anastomosis, and blood flow reconstruction. Key operational parameters are recorded for each step, such as incision location, suture method, and tension control. The generated postoperative vascular morphology takes into account the local deformation and stress state caused by the surgical procedure. The deformation model can be represented as:

[0181] V post =D(V) AVF ,T surgical E tissue ),

[0182] Among them, V post The postoperative vascular morphology after deformation; D represents the deformation calculation operation; T surgical For surgical tension parameters; E tissue For the elastic properties of tissues, the elastic modulus is commonly used to represent the elastic properties of blood vessel walls. The elastic modulus of arterial walls ranges from 0.5 to 1.5 MPa, while that of venous walls ranges from 0.2 to 0.6 MPa.

[0183] Predicting the state of the arteriovenous fistula at different time points, including one week, one month, and six months later. The prediction is based on a physical-biological coupling model, considering the vascular wall remodeling process under hemodynamic stimulation. The remodeling model can be represented as:

[0184] V t =R(V post ,τ w ,t,P patient ),

[0185] Among them, V t The state of the blood vessels after time t; R represents the reconstruction calculation operation; τ w The wall shear force distribution is shown; t is time, in days; P patient Individual patient parameters, including age and underlying diseases, are used. The rate of vascular dilation during remodeling exhibits a non-linear relationship with the wall shear force.

[0186]

[0187] in, τ is the rate of change in vessel diameter, expressed in mm / day; k is the proportionality coefficient, adjusted according to the patient's age: 0.03-0.05 for younger patients (<40 years old) and 0.01-0.02 for older patients (>65 years old); w τ0 is the wall shear force, in Pa; τ0 is the reference shear force, usually taken as 1.5 Pa; α is the nonlinear exponent, with a value of approximately 0.4; f(D,D) max () is a limiting function to ensure that the expansion does not exceed the physiological limit D. max .

[0188] Vascular dilation is a key indicator during the maturation of an arteriovenous fistula. The formula above reflects clinical observations: moderate shear force promotes vascular dilation, while the rate of dilation slows down and eventually reaches equilibrium after a certain degree of dilation. For example, in a typical cephalic vein fistula, the initial diameter is approximately 3-4 mm, reaching 6-8 mm after maturation, but almost never exceeding 10 mm. This upper limit is determined by D. max control.

[0189] Calculate the predicted maturation time and long-term patency rate of the arteriovenous fistula (AVF). The maturation criteria are: vessel diameter ≥ 6 mm, blood flow ≥ 600 ml / min, and depth from the skin ≤ 6 mm. Based on the predicted vascular dilation rate and blood flow growth curve, calculate the time required to reach the maturation criteria.

[0190] T mature =min{t|D(t)≥6mm∧Q(t)≥600ml / min∧Depth(t)≤6mm},

[0191] Among them, Tmature The maturation time is expressed in days; D(t), Q(t), and Depth(t) represent the vessel diameter, blood flow, and skin depth at time t, respectively.

[0192] Long-term patency prediction is based on a comprehensive assessment of various risk factors, including shear stress distribution, vascular elasticity, and underlying patient conditions. The patency probability model can be expressed as:

[0193]

[0194] Among them, P patency (t) represents the patency probability at time t; h(s,X) is the risk function, which is related to time s and the risk factor vector X. Risk factors include wall shear force nonuniformity, vessel tortuosity, patient age, and comorbidities.

[0195] In arteriovenous fistula (AVF) planning practice, maturation time and patency rate are two core indicators for evaluating the quality of a treatment plan. For example, for patients requiring emergency dialysis, a shorter maturation time is more advantageous; while for younger, long-term dialysis patients, a high patency rate may be more important. This system's predictive model can provide quantitative estimates of these key indicators, helping physicians make more informed decisions.

[0196] Finally, a visual surgical guidance plan is generated, including anastomosis angles, location markers, and key operational points. The guidance plan is presented as an intuitive 3D image, annotating key anatomical landmarks and surgical reference points. Figure 9 The comparison between virtual surgical planning and actual surgical results was demonstrated, verifying the accuracy and reliability of the proposed method.

[0197] In clinical applications, virtual surgical guidance significantly improves surgical accuracy. For example, with traditional methods, doctors often rely on experience to estimate the anastomosis angle, resulting in an error of ±10° between the actual and target angles; however, with the visualization guidance of this system, the error can be controlled within ±3°, significantly improving the repeatability and success rate of the surgery.

[0198] Step S6: Adaptive closed-loop feedback optimization, such as... Figure 10 As shown, the present invention also includes an adaptive closed-loop feedback optimization step, which is crucial to the performance of the entire system. Specifically, it includes:

[0199] First, record the actual surgical parameters and postoperative follow-up data. Surgical parameters include the actual anastomosis angle, location, and technical details; follow-up data includes fistula maturation time, blood flow changes, and complication status. Data collection uses standardized forms to ensure consistency and comparability across centers. The data structure can be represented as follows:

[0200] D feedback ={(P i ,Si ,F i |i=1,2,...,N},

[0201] Among them, D feedback For the feedback dataset; P i S represents the characteristic data of the i-th patient; i For the corresponding surgical parameters; F i This represents the follow-up results; N represents the total number of cases.

[0202] Secondly, compare the predicted results with the actual results to quantify the prediction accuracy. For quantitative indicators (such as blood flow), calculate the relative error between the predicted and actual values; for qualitative indicators (such as patency), calculate the prediction accuracy, sensitivity, and specificity. Evaluation indicators can be expressed as:

[0203]

[0204] Where Acc represents accuracy; Sen represents sensitivity; Spe represents specificity; and TP, TN, FP, and FN represent the number of true positive, true negative, false positive, and false negative samples, respectively. For predicting patency status, patency at 6 months is typically considered a positive event.

[0205] Next, the model parameters are updated based on the difference between the prediction and the actual results. For deep learning models, an incremental learning method is used, which retains existing knowledge while integrating new data. The loss function for incremental learning is:

[0206] L inc =L new +λ·L dist ,

[0207] Among them, L inc L represents the total loss from incremental learning. new For the prediction loss of new data; L dist To mitigate knowledge distillation loss and prevent catastrophic forgetting, λ is the balance coefficient, typically set to 0.5. For the physical model, relevant physical parameters are adjusted to reduce systematic errors. The model update cycle is set every 3 months or when 50 new cases have accumulated.

[0208] Then, new case data are added to the knowledge base to optimize the sampling strategy. The knowledge base uses a hierarchical storage structure, categorized and archived according to patient characteristics, surgical procedures, and outcome evaluations. The sampling strategy prioritizes data from high-error regions, concentrating resources on improving the aspects most in need of improvement. The sampling probability formula is:

[0209] P(sample i )∝exp(γ·err i ),

[0210] Where, P(sample)i ) represents the probability of selecting the i-th sample; err i γ represents the prediction error for the corresponding sample; γ is a temperature parameter that controls the degree of sampling bias, usually set to 2.0.

[0211] Finally, anomaly detection and analysis are performed on cases of prediction failure to improve system robustness. For cases where predictions significantly deviate from reality, the reasons are analyzed in depth to identify potential special factors or model blind spots. The anomaly judgment criteria are as follows:

[0212] |pred i -actual i |>μ err +2σ err ,

[0213] Among them, pred i and actual i Let μ be the predicted value and the actual value of the i-th example, respectively; err and σ err These are the mean and standard deviation of the historical error, respectively.

[0214] In practical applications, closed-loop feedback mechanisms are crucial for the continuous improvement of system performance. For example, the system may find that the predicted maturation time of arteriovenous fistulas in diabetic patients is generally shorter. By analyzing such abnormal patterns, the system can automatically adjust the weights of diabetes-related parameters or add specific diabetes-influencing factors to the model, thereby improving the prediction accuracy for this special population.

[0215] Through the aforementioned closed-loop feedback mechanism, the system can continuously learn and optimize, with prediction accuracy steadily improving over time. In practical applications, after learning from approximately 500 cases, the overall prediction accuracy reaches over 85%, providing reliable support for clinical decision-making. Compared to traditional methods that rely on physician experience (with a success rate of approximately 60%–65%), this system significantly improves the success rate and long-term outcomes of arteriovenous fistula surgery.

[0216] like Figure 7 As shown, the intelligent arteriovenous fistula planning system based on hemodynamic simulation provided by this invention includes the following modules:

[0217] The 3D vascular tree model construction module 10 is used to acquire 3D spatial anatomical data of the patient's arteries and veins and construct a 3D vascular tree model. This module includes a data acquisition unit 11 and a model construction unit 12.

[0218] Data acquisition unit 11 is responsible for acquiring raw 3D vascular data from high-resolution ultrasound or computed tomography (CT) angiography. In practical applications, this unit can connect to existing hospital imaging equipment (such as ultrasound, CT, or MRI) via a DICOM standard interface to automatically acquire patient image data. To ensure data quality, ultrasound imaging uses a 7-12MHz high-frequency probe with a scanning resolution of 0.1mm; CTA imaging uses spiral CT with a slice thickness not exceeding 1mm.

[0219] The model building unit 12 inputs the acquired raw 3D vascular data into a two-stage deformable 3D convolutional neural network based on a self-attention mechanism for processing, resulting in a 3D vascular tree model. This unit adopts the neural network architecture detailed in Example 1, achieving efficient processing through GPU acceleration. For standard resolution vascular data (e.g., 512×512×200 voxels), model building can typically be completed within 2–3 minutes, achieving a Dice coefficient accuracy of over 95%.

[0220] In clinical applications, this module addresses the issue of insufficient accuracy in traditional vessel segmentation methods when processing images with uneven quality and low contrast. For example, traditional methods often result in discontinuous segmentation or missed detections for tiny vessels with a diameter of only 2 mm or complex bifurcation regions; however, this module, through deformable convolution and self-attention mechanisms, can accurately identify these key structures, providing a reliable foundation for subsequent analysis.

[0221] The hemodynamic simulation module 20 combines vascular parameter data from the 3D vascular tree model with hemodynamic parameter data under different blood flow velocities, inputting them into the deep learning model to obtain anastomotic dynamic shear force data. This module includes a region partitioning unit 21, a fluid calculation unit 22, an SPH simulation unit 23, and a deep learning prediction unit 24.

[0222] Region partitioning unit 21 is responsible for dividing the 3D vascular tree model into different regions, including the anastomotic core region, anastomotic edge region, transition region, and distal region, and assigning different mesh densities to each region. The partitioning algorithm, based on vascular anatomy features and clinical experience, automatically identifies potential anastomotic locations and labels the regions. A multi-scale meshing strategy is adopted: the mesh density is 10 μm for the anastomotic core region, 25 μm for the anastomotic edge region, 50 μm for the transition region, and 100 μm for the distal region.

[0223] The fluid computation unit 22 uses computational fluid dynamics software to calculate the pressure field, velocity field, and turbulence field of the vascular tree 3D model, and sets hemodynamic parameters for different blood flow velocities. This unit dynamically allocates computational resources according to computational needs, and for complex vascular structures, distributed computing is used to accelerate processing. Typically, 5-7 different flow velocity values ​​(e.g., from 20 cm / s to 100 cm / s) are set to cover the range of blood flow changes from resting to movement in patients.

[0224] SPH simulation unit 23 inputs hemodynamic parameters at different blood flow velocities into an improved smoothed particle hydrodynamics method for simulation, obtaining shear force data of the anastomosis region under different blood flow velocities. This unit implements the improved SPH method detailed in Example 1, significantly improving simulation speed through GPU parallel computing. Compared with traditional mesh-based CFD, the SPH method has a natural advantage in handling large deformations and fluid-structure interaction problems, and is particularly suitable for simulating complex flow fields at arteriovenous fistula anastomoses.

[0225] The deep learning prediction unit 24 uses vascular parameter data as input to the deep learning model to obtain anastomotic dynamic shear force data. This unit employs a pre-trained temporal convolutional network model, which can complete predictions within seconds, greatly improving the system response speed. The model input dimension is 27 (number of vascular feature parameters), and the output is temporal shear force data, with a typical size of 32×32×100 (32×32 spatial grid, 100 time steps).

[0226] The core innovation of this module lies in combining traditional fluid dynamics calculations with deep learning predictions, ensuring both physical accuracy and overcoming computational efficiency bottlenecks. For example, in situations requiring comparison of multiple scenarios, traditional methods often necessitate a complete CFD simulation for each scenario, taking several hours; while this system, through deep learning predictions, can reduce the computation time to a few seconds while maintaining an error margin of less than 7%, making real-time clinical decision-making possible.

[0227] The optimal anastomosis angle determination module 30 compares the dynamic shear force data of the anastomosis joint with the preset optimal anastomosis angle data to determine the optimal anastomosis angle. This module includes an association modeling unit 31, a multi-objective evaluation unit 32, and a scheme generation unit 33.

[0228] The correlation modeling unit 31 constructs a deep correlation model between dynamic time-varying shear force and vessel wall response, including a dual-stream feature extraction network and a multi-scale fusion module. This unit processes shear force data from the hemodynamic simulation module and establishes the correlation between shear force patterns and long-term vessel wall response. The model adopts a dual-stream architecture, processing time-varying shear force data and vessel wall characteristic data separately, and fusing the two feature streams through a cross-attention mechanism.

[0229] The multi-objective assessment unit 32 evaluates each candidate anastomosis angle scheme based on a physically-infused Bayesian optimization framework, calculating a multi-objective score that includes surgical success rate, fistula maturation time, and long-term patency rate. The assessment process fully considers individual patient characteristics and dynamically adjusts the weights of each objective. The optimization framework employs a Matérn 5 / 2 kernel Gaussian process regression model, combined with an expected improvement acquisition function to guide the search process.

[0230] The protocol generation unit 33 determines the optimal anastomosis angle based on multi-objective scoring and specific patient needs, and generates a detailed protocol description. The description includes angle values ​​(usually accurate to 1°), location descriptions, and special considerations. Typically, the optimal anastomosis angle is within the range of 30°–40°, but this can be individually adjusted based on patient characteristics (such as age and underlying diseases) and vascular features.

[0231] The unique value of this module lies in its intelligent transformation from physical simulation to clinical decision-making. Traditional arteriovenous fistula (AVF) planning relies heavily on physician experience and lacks quantifiable standards; however, this module, through a multi-objective optimization framework, integrates hemodynamic parameters, surgical technique requirements, and long-term outcome prediction into a unified decision-making system, making AVF planning a quantifiable and optimizable process. In clinical validation, the system-recommended anastomosis angle scheme improved the success rate by 25%–30% compared to the traditional experience-based method.

[0232] The 3D modeling module 40 for arteriovenous fistula (AVF) surgery is used to perform 3D modeling of the vascular tree 3D model in conjunction with the optimal anastomosis angle, resulting in a 3D AVF model. This module employs a parametric modeling method, automatically generating the anastomosis structure based on parameters such as the optimal anastomosis angle, vessel diameter, and anastomosis length.

[0233] For common end-to-side anastomosis, the modeling process first creates an elliptical incision on the venous side, with the long axis typically 1.5-2 times the vein diameter; then, the arterial end is aligned with the incision at the optimal anastomosis angle; finally, the two are connected using simulated sutures to form a complete anastomosis structure. The model also considers factors such as suture tension and vascular deformation to more realistically reflect the postoperative morphology.

[0234] After modeling is completed, the system automatically evaluates the model quality, checking the anastomosis morphology, lumen continuity, and spatial positioning rationality. If necessary, fine-tuning is made to meet the actual surgical requirements. For example, if the anastomosis location is limited by local anatomy, the position or angle is adjusted appropriately to ensure operative safety.

[0235] The advantage of this module lies in combining virtual surgical planning with actual anatomical constraints to generate arteriovenous fistula models that meet both hemodynamic optimization requirements and surgical feasibility. This balance between ideal and reality is a key factor for clinical success. In practical applications, this module can flexibly adjust modeling parameters according to different fistula types (such as radial artery-cephalic vein, brachial artery-cephalic vein, etc.) to adapt to diverse clinical needs.

[0236] The virtual surgery module 50 is used to perform virtual surgery based on a three-dimensional arteriovenous fistula model, and obtain virtual surgical process data and postoperative hemodynamic status assessment results. This module includes a scene construction unit 51, a process simulation unit 52, a status prediction unit 53, and a guidance generation unit 54.

[0237] Scene construction unit 51 constructs a virtual surgical scene based on the optimal anastomosis angle and a 3D vascular tree model, including elements such as patient anatomy, surgical instruments, and operational procedures. Scene construction utilizes 3D visualization technology to provide an immersive surgical environment. In practical applications, the surgical scene also includes surrounding important anatomical structures (such as nerves and tendons) to help surgeons avoid potential risks.

[0238] The process simulation unit 52 simulates the surgical procedure and generates postoperative vascular morphology. The simulation follows the standard surgical procedure: vascular exposure, vascular incision, anastomosis, and blood flow reconstruction. Key operational parameters are recorded for each step, taking into account local deformation and stress states caused by the surgical procedure. The simulation results show the immediate vascular status after surgical intervention, including anastomosis morphology and initial blood flow distribution.

[0239] The status prediction unit 53 predicts the fistula status at different stages and calculates predicted values ​​for fistula maturation time and long-term patency rate. The prediction is based on a physical-biological coupling model, considering the vascular wall remodeling process under hemodynamic stimulation. This unit can simulate the entire process of fistula establishment to maturation, predicting changes in vascular morphology and function one week, one month, and six months later. Under standard conditions, the prediction error for fistula maturation time is controlled within ±10 days, and the accuracy of patency rate prediction reaches 80%.

[0240] The guidance generation unit 54 generates a visual surgical guidance plan, including anastomosis angles, location markers, and key operational points. The guidance plan is presented as an intuitive 3D image, annotating key anatomical landmarks and surgical reference points. For example, the system clearly marks the ideal vascular incision location, incision length, suture point distribution, and areas requiring special attention, enabling surgeons to execute the surgery precisely according to the plan.

[0241] This module transforms static planning into dynamic surgical guidance through virtual surgery and predictive simulation. It not only tells doctors what the optimal approach is, but also demonstrates how to implement this approach and what its effects will be, greatly improving the operability and predictability of the planning results. Clinical practice shows that after adopting virtual surgical guidance, surgical deviations are significantly reduced; the error between the actual anastomosis angle and the planned angle is reduced from the traditional ±10° to ±3°, and surgical accuracy is improved by more than 70%.

[0242] The adaptive closed-loop feedback module 60 is used to record actual surgical parameters and postoperative follow-up data, compare the predicted results with the actual results, update the model parameters, and improve the system's prediction accuracy. This module includes a data collection unit 61, a result comparison unit 62, a model update unit 63, and an anomaly analysis unit 64.

[0243] Data collection unit 61 is responsible for recording actual surgical parameters and postoperative follow-up data. Standardized forms are used for data collection to ensure consistency and comparability across centers. The collected data includes surgical details (such as actual anastomosis angle, location, and technique) and follow-up results (such as fistula maturation time, blood flow changes, and complication status), forming a structured database to support subsequent analysis.

[0244] The results comparison unit 62 compares the predicted results with the actual results, quantifying the prediction accuracy. The comparison employs multiple evaluation metrics, such as relative error, accuracy, sensitivity, and specificity, to comprehensively assess the model's predictive performance. The evaluation results not only reflect the overall prediction level but also identify prediction biases for specific patient groups or arteriovenous fistula types, providing direction for targeted model optimization.

[0245] The model update unit 63 updates the model parameters based on the difference between the predicted and actual results, and adds new case data to the knowledge base. The model update adopts an incremental learning method, retaining existing knowledge while integrating new data. The update strategy includes parameter fine-tuning, feature weight adjustment, and model structure optimization to ensure that the system can adapt to constantly changing clinical situations.

[0246] The anomaly analysis unit 64 performs anomaly detection and analysis on cases where predictions fail, improving system robustness. The analysis results are used to adjust the model structure or add special case handling mechanisms, continuously improving the system's adaptability to diverse situations. For example, if it is found that predictions for a specific type of patient (such as elderly diabetic patients) generally have large deviations, the system will automatically increase the weight of relevant features or add specific adjustment factors to improve the prediction accuracy for such patients.

[0247] This module is a crucial component for the continuous improvement of the entire system. Through a closed-loop mechanism of planning-execution-feedback-optimization, it achieves self-learning and evolutionary capabilities. Clinical data shows that after learning from approximately 500 cases, the system's predictive accuracy improved from around 70% initially to over 85%, far exceeding the 60%–65% of traditional empirical methods. This continuous learning ability allows the system to constantly adapt to new patient groups and clinical situations, maintaining long-term effectiveness.

[0248] Through the aforementioned system structure, this invention achieves intelligent automation of the entire process, from acquiring individual patient vascular data to generating the optimal surgical plan, providing an efficient and precise solution for arteriovenous fistula (AVF) planning. Compared to traditional methods, this system not only significantly improves the success rate and long-term outcomes of AVF surgery but also reduces reliance on senior specialists through standardization and intelligentization, enabling primary hospitals to provide high-quality AVF creation services and promoting the balanced utilization of medical resources.

[0249] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent planning method for arteriovenous fistulas based on hemodynamic simulation, characterized in that, include: Obtain three-dimensional spatial anatomical data of the patient's arteries and veins, and construct a three-dimensional model of the vascular tree; The vascular parameter data of the three-dimensional vascular tree model and the hemodynamic parameter data under different blood flow velocities are combined and input into the deep learning model to obtain the dynamic shear force data of the anastomosis. The dynamic shear force data of the anastomosis joint is compared with the preset optimal anastomosis angle data to determine the optimal anastomosis angle; The optimal anastomosis angle is used to perform a three-dimensional modeling of the arteriovenous fistula procedure on the three-dimensional model of the vascular tree, resulting in a three-dimensional arteriovenous fistula model. Based on the three-dimensional arteriovenous fistula model, a virtual surgery was performed to obtain virtual surgical process data and postoperative hemodynamic status assessment results.

2. The intelligent planning method for arteriovenous fistulas based on hemodynamic simulation according to claim 1, characterized in that, Obtaining the three-dimensional spatial anatomical data of the patient's arteries and veins, and constructing the three-dimensional model of the vascular tree, specifically includes: Acquire raw 3D vascular data from high-resolution ultrasound or computed tomography angiography. The original three-dimensional blood vessel data is input into a two-stage deformable 3D convolutional neural network based on a self-attention mechanism for processing to obtain the three-dimensional model of the blood vessel tree. The self-attention mechanism is used to enhance the accuracy of identifying key branch points of blood vessels.

3. The intelligent planning method for arteriovenous fistulas based on hemodynamic simulation according to claim 2, characterized in that, The construction process of the self-attention-based two-stage deformable 3D convolutional neural network includes: A two-stage deformable 3D convolutional neural network architecture based on self-attention mechanism is constructed, wherein the first stage obtains the coarse segmentation results of blood vessels, and the second stage refines the processing of blood vessel boundaries and branch regions. Obtain a training dataset containing three-dimensional blood vessel data and corresponding labels, wherein the labels include information on blood vessel boundaries, centerlines, and branch point locations; The training dataset is used to train the two-stage deformable 3D convolutional neural network based on the self-attention mechanism to obtain the trained neural network model.

4. The intelligent planning method for arteriovenous fistulas based on hemodynamic simulation according to claim 1, characterized in that, The vascular parameter data of the 3D vascular tree model and the hemodynamic parameter data under different blood flow velocities are combined and input into a deep learning model to obtain the dynamic shear force data of the anastomosis, specifically including: The three-dimensional model of the vascular tree is divided into different regions, including the anastomosis core region, the anastomosis edge region, the transition region and the distal region, and different mesh densities are assigned to each region. The pressure field, velocity field, and turbulence field of the vascular tree three-dimensional model were calculated using computational fluid dynamics software, and hemodynamic parameters were set for different blood flow velocities. The hemodynamic parameters at different blood flow rates were input into an improved smooth particle hydrodynamics method for simulation to obtain shear force data of the anastomosis region at different blood flow rates. The vascular parameter data is used as input to the deep learning model to obtain the dynamic shear force data of the anastomosis.

5. The intelligent planning method for arteriovenous fistulas based on hemodynamic simulation according to claim 4, characterized in that, The improved smooth particle hydrodynamics method includes: Based on the morphological characteristics of the vascular tree 3D model, an adaptive particle distribution is automatically generated; Increasing particle density in the anastomosis region improves simulation accuracy; An adaptive kernel function is used, and the kernel radius is dynamically adjusted according to the local particle density. An improved virtual particle method is used to handle the blood vessel wall boundary, improving the stability of the calculation at the boundary. Considering the elastic deformation response of the blood vessel wall, a fluid-structure interaction simulation is achieved.

6. The intelligent planning method for arteriovenous fistulas based on hemodynamic simulation according to claim 1, characterized in that, The dynamic shear force data of the anastomosis site is compared with the preset optimal anastomosis angle data to determine the optimal anastomosis angle, specifically including: A deep correlation model between dynamic time-varying shear force and blood vessel wall response was constructed, including a dual-flow feature extraction network and a multi-scale fusion module; The dynamic shear force data of the anastomosis is input into the depth correlation model to obtain the long-term blood vessel wall response prediction results under different anastomosis angles; The Bayesian optimization framework based on physical information enhancement evaluates each candidate anastomosis angle scheme and calculates a multi-objective score including surgical success rate, fistula maturation time and long-term patency rate. The optimal anastomosis angle is determined based on the multi-objective score and the patient's specific needs.

7. The intelligent arteriovenous fistula planning method based on hemodynamic simulation according to claim 6, characterized in that, The Bayesian optimization framework based on physical information enhancement includes: By integrating hemodynamic theory and clinical experience, a theoretical correlation model between anastomosis angle and arteriovenous fistula function was established. Define physical feasibility constraints, including angular range and spatial location limitations; Construct a multi-objective function and dynamically adjust the weight coefficients of short-term, medium-term, and long-term objectives; Initialize candidate matching angle schemes and use a Gaussian process regression model to predict the performance of each scheme; Based on the expected improvement and the trade-off between uncertainty, iterative optimization is performed until the optimal solution is converged.

8. The intelligent planning method for arteriovenous fistulas based on hemodynamic simulation according to claim 1, characterized in that, Based on the aforementioned three-dimensional arteriovenous fistula model, a virtual surgery is performed to obtain the virtual surgical process data and the postoperative hemodynamic status assessment results, specifically including: A virtual surgical scene is constructed based on the optimal anastomosis angle and the three-dimensional model of the vascular tree. Simulate the surgical procedure and generate postoperative vascular morphology; Predict the status of the arteriovenous fistula at different time points, including one week, one month, and six months later; Calculate the predicted values ​​for the maturation time and long-term patency rate of the arteriovenous fistula; Generate a visual surgical guidance plan, including anastomosis angles, location markers, and key operational points.

9. The intelligent planning method for arteriovenous fistulas based on hemodynamic simulation according to claim 1, characterized in that, It also includes an adaptive closed-loop feedback optimization step: Record actual surgical parameters and postoperative follow-up data; Compare the predicted results with the actual results to quantify the accuracy of the prediction; Update the model parameters based on the difference between the prediction and the actual results; Add new case data to the knowledge base and optimize sampling strategies; Anomaly detection and analysis are performed on cases where predictions fail to improve system robustness.

10. An intelligent arteriovenous fistula planning system based on hemodynamic simulation, used to execute the intelligent arteriovenous fistula planning method based on hemodynamic simulation as described in any one of claims 1-9, characterized in that, include: The 3D model building module for the vascular tree is used to acquire the 3D spatial anatomical data of the patient's arteries and veins and build a 3D model of the vascular tree. The hemodynamic simulation module is used to combine the vascular parameter data of the three-dimensional vascular tree model with the hemodynamic parameter data under different blood flow velocities, and input them into the deep learning model to obtain the dynamic shear force data of the anastomosis. The optimal anastomosis angle determination module is used to compare the dynamic shear force data of the anastomosis with the preset optimal anastomosis angle data to determine the optimal anastomosis angle. The arteriovenous fistula 3D modeling module is used to perform arteriovenous fistula 3D modeling on the vascular tree 3D model in combination with the optimal anastomosis angle, so as to obtain a 3D arteriovenous fistula model. The virtual surgery module is used to perform virtual surgery based on the three-dimensional arteriovenous fistula model, and obtain virtual surgical process data and postoperative hemodynamic status assessment results. The adaptive closed-loop feedback module is used to record actual surgical parameters and postoperative follow-up data, compare the predicted results with the actual results, update the model parameters, and improve the system's prediction accuracy.

Images