A cryoablation temperature prediction method, device, electronic device and storage medium

Abstract

The embodiments of the present invention relate to the field of cryoablation and disclose a cryoablation temperature prediction method, device, electronic device and storage medium, including: constructing a vascular network diagram based on the vascular centerline in the three-dimensional imaging data before the cryoablation operation, the nodes in the vascular network diagram represent the vascular bifurcation points, and the blood edges represent the vascular connection relationship; during the cryoablation process, using a pre-trained temperature prediction model to predict the temperature field of the target sampling point, the temperature prediction model includes a physical information neural network and a graph neural network; the physical information neural network uses the biological heat conduction equation as a constraint condition to predict the temperature field of the target sampling point, and the graph neural network is used to predict the blood perfusion coefficient of each node in the vascular network diagram to determine the blood perfusion coefficient contained in the biological heat conduction equation, which solves the problem of real-time prediction of cryoablation temperature in accordance with physical laws, so that the results conform to physical laws and meet the real-time requirements during the operation.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of cryoablation, and in particular to a cryoablation temperature prediction method, device, electronic device, and storage medium. Background Art

[0002] In cryoablation therapy for tumors, precise control of the ablation area is the core factor that determines efficacy and safety. Traditional methods for controlling the ablation area include: (1) numerical simulation methods based on biological heat conduction equations. This method has problems such as low computational efficiency, poor individual adaptability, and lack of feedback mechanism, making it difficult to form a complete solution that is clinically feasible, computationally efficient, and radiation safe; (2) physical modeling methods. Although this method can accurately simulate the heat conduction process, it has limitations in dealing with the complexity and individual differences of biological tissues. For example, the vascular network structure of biological tissues is complex and has large individual differences, making it difficult to accurately model. In addition, the three-dimensional meshing and iterative solution of the finite element method are time-consuming and cannot meet the real-time requirements of intraoperative surgery. (3) Purely data-driven models, such as traditional neural networks, can learn complex data patterns, but often lack physical interpretability and generalization capabilities. Especially when there is insufficient training data, it is difficult to obtain an accurate model. In addition, purely data-driven models cannot ensure that the model conforms to basic physical laws.

[0003] Therefore, this field urgently needs a cryoablation temperature prediction solution that conforms to physical laws and meets the real-time requirements during surgery. Summary of the Invention

[0004] The purpose of the present invention is to at least provide a method, device, electronic device and storage medium for predicting cryoablation temperature, which can at least solve the problem of real-time prediction of cryoablation temperature in accordance with physical laws, and at least achieve the effect that the prediction results conform to physical laws and meet the real-time requirements during surgery.

[0005] To solve the above technical problems, at least one embodiment of the present application provides a cryoablation temperature prediction method, including: obtaining three-dimensional imaging data before the cryoablation surgery; constructing a vascular network diagram based on the vascular centerlines in the three-dimensional imaging data, the nodes in the vascular network diagram representing vascular bifurcation points, and the edges in the vascular network diagram representing vascular connection relationships; during the cryoablation process, predicting the temperature field of the target sampling point through a pre-trained temperature prediction model, the temperature prediction model includes a physical information neural network and a graph neural network; the physical information neural network uses the biological heat conduction equation as a constraint condition to predict the temperature field of the target sampling point, and the graph neural network is used to predict the blood perfusion coefficient of each node in the vascular network diagram to determine the blood perfusion coefficient contained in the biological heat conduction equation.

[0006] At least one embodiment of the present application also provides a cryoablation temperature prediction device, including: a data input module for acquiring three-dimensional imaging data before the cryoablation surgery; a data processing module for constructing a vascular network diagram based on the vascular centerlines in the three-dimensional imaging data, the nodes in the vascular network diagram representing vascular bifurcation points, and the edges in the vascular network diagram representing vascular connection relationships; a model prediction module for predicting the temperature field of the target sampling point during the cryoablation process through a pre-trained temperature prediction model, the temperature prediction model including a physical information neural network and a graph neural network; the physical information neural network uses the biological heat conduction equation as a constraint condition to predict the temperature field of the target sampling point, and the graph neural network is used to predict the blood perfusion coefficient of each node in the vascular network diagram to determine the blood perfusion coefficient contained in the biological heat conduction equation.

[0007] At least one embodiment of the present application also provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above-mentioned cryoablation temperature prediction method.

[0008] At least one embodiment of the present application further provides a computer-readable storage medium storing a computer program, which implements the above-mentioned cryoablation temperature prediction method when executed by a processor.

[0009] The cryoablation temperature prediction method provided in the embodiments of the present application can achieve safe, real-time, and high-precision temperature field prediction by integrating physical constraints based on the biological heat conduction equation with data-driven cryoablation temperature based on graph neural networks. The physical information neural network is a machine learning model that combines deep learning and physics knowledge. As the backbone network of physical constraints, it can ensure the basic physical rationality of the model. The biological heat conduction equation is used as a constraint condition through the physical information neural network to ensure that the model prediction results conform to the laws of thermodynamics.

[0010] In some optional embodiments, a vascular network graph is constructed based on the vascular centerlines in the three-dimensional image data, including: performing tissue segmentation on the three-dimensional image data to obtain a tissue segmentation result; extracting the vascular centerlines in the tissue segmentation result through skeletonization processing; identifying vascular bifurcation points in the vascular centerlines; constructing a vascular network topology based on the vascular centerlines, and constructing a vascular network graph based on the vascular network topology, wherein the nodes in the vascular network graph represent vascular bifurcation points, and the edges in the vascular network graph represent vascular connection relationships. Using a graph neural network as a data-driven vascular effect subnetwork can learn and compensate for parts of the model that are difficult to accurately model, such as the impact of the vascular network on blood perfusion. The complex relationship between vascular distribution and blood perfusion coefficient can be learned from the data, and an accurate blood perfusion coefficient can be provided to the main network.

[0011] In some optional embodiments, the temperature field of the target sampling point is predicted by a pre-trained temperature prediction model, including: predicting the blood perfusion coefficient of each node in the vascular network graph by the graph neural network; interpolating the blood perfusion coefficient to determine the blood perfusion coefficient of each target sampling point; substituting the blood perfusion coefficient of each target sampling point into the biological heat conduction equation, using the equation as a constraint condition, and predicting the temperature of each target sampling point by the physical information neural network.

[0012] In some optional embodiments, the temperature prediction model further includes a gated fusion network for correcting the blood perfusion coefficient of each target sampling point using the following calculation formula before substituting the blood perfusion coefficient of each target sampling point into the biological heat conduction equation;

[0013]

[0014] Where, represents the corrected blood perfusion coefficient of the target sampling point, represents the weight coefficient, represents the blood perfusion coefficient of the target sampling point before correction, The weight coefficient is dynamically adjusted according to the distance between the current target sampling point and the nearest blood vessel and the type of ablated tissue.

[0015] By introducing a gated fusion mechanism, the weight coefficient is adaptively adjusted according to tissue type and vascular distance (for example, the distance between the target sampling point and the nearest blood vessel). In this way, the contribution ratio of the blood perfusion coefficient output by the GNN sub-network to the cPINN backbone network is dynamically adjusted according to the spatial position in different tissue regions and different vascular distances. This makes the prediction results of the backbone network more consistent with physical laws, thereby improving the robustness and adaptability of the model.

[0016] In some optional embodiments, the cryoablation temperature prediction method further includes: determining target sampling points using the following strategy: for a first region around the ablation tissue boundary, determining target sampling points using a first sampling density; for a second region around the ablation tissue boundary, determining target sampling points using a second sampling density; for a third region around blood vessels with a diameter greater than a preset value and a fourth region around the ablation needle heat exchange zone, increasing the sampling density to determine target sampling points, the sampling density increase in the third region being lower than the sampling density increase in the fourth region; for the remaining regions, determining target sampling points using a third sampling density; the second region being located in an area of ​​the first region away from the ablation tissue boundary, the second sampling density being lower than the first sampling density, and the third sampling density being lower than the second sampling density. Through the above-mentioned partitioned adaptive sampling strategy, temperature information of regions of interest of different clinical importance can be obtained in a targeted manner, thereby providing accurate temperature data support for cryoablation surgery.

[0017] In some optional embodiments, the cryoablation temperature prediction method further includes: constructing and training a temperature prediction model; the training of the temperature prediction model includes: using a synthetic training data set to perform unsupervised training on the constructed temperature prediction model; constructing a finite element model based on the three-dimensional imaging data of a real patient, using the finite element model to solve the biological heat conduction equation to obtain a reference temperature field of the target sampling point, using the unsupervised trained temperature prediction model based on the three-dimensional imaging data of the real patient to obtain a predicted temperature field of the target sampling point, and optimizing the temperature prediction model with the minimum value of the first objective function as the goal, wherein the first objective function is ,in, represents the mean square error between the predicted temperature field and the reference temperature field, λ represents the trade-off coefficient, represents the value of the first objective function; by adjusting the parameters of the physical information neural network, the parameters of the graph neural network, and the parameters of the biological heat conduction equation, the optimized temperature prediction model is adjusted so that the predicted temperature field obtained by the optimized temperature prediction model based on the three-dimensional imaging data of the real patient is close to the reference temperature field. By adjusting the parameters of the physical information neural network, the parameters of the graph neural network, and the parameters of the biological heat conduction equation, the accuracy of the model prediction can be improved.

[0018] In some optional embodiments, the boundary conditions of the physical information neural network include the ablation needle surface temperature, far-field temperature and initial temperature; the loss function used in unsupervised training is as follows:

[0019]

[0020] Where, are respectively the first weight coefficient, the second weight coefficient, the third weight coefficient, the fourth weight coefficient and the fifth weight coefficient; Represents the residual loss of the biological heat conduction equation, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the biological heat conduction equation; Represents the far-field temperature loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the far-field temperature; Represents the initial condition loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the initial temperature; represents the data matching loss, which is used to measure the gap between the temperature field predicted by the physical information neural network and the actual data; The loss function represents the variational energy conservation loss, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the law of conservation of energy. This loss function not only measures the difference between the prediction and the data (if data exists), but also the degree of conformity of the prediction with the physical laws (defined by partial differential equations (physical constraints - the biological heat conduction equation) and boundary conditions). This can make the predictions of the trained model more consistent with reality.

[0021] In some optional embodiments, the cryoablation temperature prediction method further includes constructing thermophysical parameter sets corresponding to different ablated tissues in different patients to solve the bioheat conduction equation when predicting the temperature field at the target sampling point, wherein the thermophysical parameter set includes the thermophysical parameters included in the bioheat conduction equation. When predicting the cryoablation temperature of a specific organ, data from the thermophysical parameter set for the corresponding organ can be used to solve the bioheat conduction equation to improve the model's prediction accuracy, organ adaptability, and specificity.

[0022] In some optional embodiments, the cryoablation temperature prediction method further includes: during the cryoablation process, collecting and measuring the refrigerant medium temperature at a first target sampling point located in the heat exchange area of ​​the ablation needle; using a temperature prediction model to obtain a predicted temperature of the first target sampling point under the currently set blood flow velocity and needle tip heat transfer efficiency factor; based on the refrigerant medium temperature and the predicted temperature at the first target sampling point, using an inversion calculation formula to inversely calculate the blood flow velocity and needle tip heat transfer efficiency factor; smoothing the inversion calculation results to obtain updated blood flow velocity and needle tip heat transfer efficiency factor; updating the biological heat conduction equation based on the updated blood flow velocity and needle tip heat transfer efficiency factor to update the temperature prediction model; the inversion calculation formula is as follows:

[0023]

[0024] Where, represents the inverted blood flow velocity vector, represents the tip heat transfer efficiency factor after inversion, Indicates the refrigerant temperature of the first target sampling point measured, Indicates the predicted temperature of the first target sampling point.

[0025] By monitoring the refrigerant medium and inverting its parameters, temperature prediction can take into account both computational efficiency and individual adaptability, thereby improving the safety of cryoablation surgery.

[0026] In some optional embodiments, the cryoablation temperature prediction method further includes: obtaining a calibration image after ablation is completed; segmenting the calibration image to obtain a target isothermal surface contour segmentation result; determining a target isothermal surface prediction result based on the temperature field of the target sampling point predicted by the temperature prediction model; aligning the target isothermal surface contour segmentation result with the target isothermal surface prediction result and calculating a similarity index; calculating a difference loss function value based on the calculated similarity index; adjusting a thermophysical parameter set with the minimum difference loss function value as the optimization goal, the thermophysical parameter set including the thermophysical parameters included in the biological heat conduction equation; updating the biological heat conduction equation based on the adjusted thermophysical parameter set; and predicting the temperature field of the target sampling point using the updated biological heat conduction equation as a constraint condition. Feedback correction is performed based on the refrigerant temperature feedback and the final image calibration to form a closed-loop control system, which can achieve accurate temperature field prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] One or more embodiments are exemplarily described by the figures in the corresponding drawings, and these exemplified descriptions do not constitute limitations on the embodiments.

[0028] Figure 1 is a flow chart of a cryoablation temperature prediction method provided in an embodiment of the present application;

[0029] Figure 2 is a schematic diagram of a cryoablation temperature prediction process provided in an embodiment of the present application;

[0030] Figure 3 is a schematic diagram of the architecture of the temperature prediction model provided in an embodiment of the present application;

[0031] Figure 4 This is a schematic diagram of the phased training process provided by the embodiment of the present application;

[0032] Figure 5 is a schematic diagram of a cryoablation temperature prediction device provided in an embodiment of the present application;

[0033] Figure 6 It is a schematic diagram of an example of hardware connection and software architecture provided by an embodiment of the present application. DETAILED DESCRIPTION

[0034] To facilitate understanding of the embodiments of the present application, relevant content regarding ablation area control is first introduced here.

[0035] In cryoablation therapy for tumors, precise control of the ablation area is the core factor that determines efficacy and safety. Traditional methods for controlling the ablation area include: (1) numerical simulation methods based on biological heat conduction equations. This method has problems such as low computational efficiency, poor individual adaptability, and lack of feedback mechanism, making it difficult to form a complete solution that is clinically feasible, computationally efficient, and radiation safe; (2) physical modeling methods. Although this method can accurately simulate the heat conduction process, it has limitations in dealing with the complexity and individual differences of biological tissues. For example, the vascular network structure of biological tissues is complex and has large individual differences, making it difficult to accurately model. In addition, the three-dimensional meshing and iterative solution of the finite element method are time-consuming and cannot meet the real-time requirements of intraoperative surgery. (3) Purely data-driven models, such as traditional neural networks, can learn complex data patterns, but often lack physical interpretability and generalization capabilities. Especially when there is insufficient training data, it is difficult to obtain an accurate model. In addition, purely data-driven models cannot ensure that the model conforms to basic physical laws.

[0036] Therefore, this field urgently needs a cryoablation temperature prediction solution that conforms to physical laws and meets the real-time requirements during surgery.

[0037] In order to solve the above-mentioned technical problem of real-time prediction of cryoablation temperature in accordance with physical laws, the present invention proposes a cryoablation temperature prediction method. The implementation details of the cryoablation temperature prediction method of this embodiment are specifically described below. The following content is only for the convenience of understanding the implementation details provided by the present invention.

[0038] Example 1:

[0039] The cryoablation temperature prediction method of this embodiment can be applied to electronic devices with communication, computing and data storage capabilities. The specific process can be as follows: Figure 1 Shown, including:

[0040] Step 101: Acquire three-dimensional image data before cryoablation surgery.

[0041] Specifically, the three-dimensional imaging data before the cryoablation surgery may be the patient's pre-operative CT (Computed Tomography) / MRI (Magnetic Resonance Imaging) three-dimensional imaging data.

[0042] Step 102 : constructing a vascular network graph based on the vascular centerlines in the three-dimensional image data. The nodes in the vascular network graph represent vascular bifurcation points, and the edges in the vascular network graph represent vascular connection relationships.

[0043] Step 103: During the cryoablation process, the temperature field of the target sampling point is predicted using a pre-trained temperature prediction model. The temperature prediction model includes a physical information neural network and a graph neural network. The physical information neural network uses the biological heat conduction equation as a constraint condition to predict the temperature field of the target sampling point. The graph neural network is used to predict the blood perfusion coefficient of each node in the vascular network diagram to determine the blood perfusion coefficient contained in the biological heat conduction equation.

[0044] By integrating physical constraints based on the biological heat conduction equation with data-driven cryoablation temperature prediction based on graph neural networks, safe, real-time, and highly accurate temperature field predictions can be achieved. The physical information neural network is a machine learning model that combines deep learning and physics knowledge. As the backbone network for physical constraints, it ensures the basic physical rationality of the model. Using the biological heat conduction equation as a constraint, the physical information neural network ensures that the model's predictions conform to the laws of thermodynamics. Furthermore, as a data-driven vascular effect subnetwork, the graph neural network can learn and compensate for parts of the model that are difficult to accurately model, such as the impact of the vascular network on blood perfusion. It can learn the complex relationship between vascular distribution and blood perfusion coefficient from data and provide accurate blood perfusion coefficients to the backbone network.

[0045] In specific implementations, Continuity-based Physics-Informed Neural Networks (cPINN) can be used as PINNs for physics-informed neural networks (PINNs). Nodes in graph neural networks (GNNs) can also include key points selected at regular intervals between adjacent vascular bifurcations along the centerline of the vessel.

[0046] In some embodiments, constructing a vascular network map based on vascular centerlines in three-dimensional image data includes:

[0047] Step 1021 , performing tissue segmentation on the three-dimensional image data to obtain a tissue segmentation result.

[0048] In specific implementation, a deep learning segmentation network (such as U-Net) can be used to extract the tumor area, vascular network and surrounding tissue of the tumor as the tissue segmentation result, which is used to extract the vascular centerline and three-dimensional reconstruction.

[0049] Furthermore, a 3D reconstruction model is constructed based on the tissue segmentation results to generate a 3D point cloud representation. ,in, Indicates tissue attribute markers, such as liver; locates the ablation needle placement position and heat exchange zone parameters based on tissue segmentation results. Heat exchange zone parameters include needle tip position , effective heat exchange area length and needle diameter .

[0050] Step 1022: Extract the blood vessel centerline from the tissue segmentation result through skeletonization.

[0051] Step 1023 , identifying the vessel bifurcation point in the vessel centerline.

[0052] In the specific implementation, the vascular bifurcation points are identified by analyzing the vascular intersections and endpoints on the vascular centerline, and then the key bifurcation points are identified as the identification results of this step using the vascular diameter change rate. The key bifurcation points are, for example, the root nodes of the root blood vessels. In actual applications, the key bifurcation points selected for different organs may be different, and this is not limited here.

[0053] Step 1024 : construct a vascular network topology structure based on the vascular centerline, and construct a vascular network graph based on the vascular network topology structure. The nodes in the vascular network graph represent vascular bifurcation points, and the edges in the vascular network graph represent vascular connection relationships.

[0054] In the specific implementation, construct a vascular network diagram G =( V , E ), node set V Represents the identified blood vessel bifurcation points, edge set E Indicates the vascular connection relationship. Furthermore, physical feature vectors are assigned to the nodes of the vascular network graph as node features in the subsequent graph neural network. The physical feature vectors include:

[0055] Blood vessel type : Such as hepatic vein (0), hepatic artery (1), portal vein (2);

[0056] Blood vessel diameter : The measurement value obtained from the three-dimensional image, unit: mm;

[0057] Blood flow velocity : The initial value estimated based on the blood vessel type, in cm / s, can be determined by searching existing literature;

[0058] Blood flow direction : three-dimensional unit vector (dx, dy, dz)

[0059] In some embodiments, the above-mentioned cryoablation temperature prediction method further includes: using the following strategy to determine target sampling points to form a target sampling point set, so as to perform temperature field solution for these target sampling points in subsequent steps.

[0060] Among them, different strategies include:

[0061] For a first area around the boundary of the ablated tissue, a first sampling density is used to determine target sampling points;

[0062] For a second area around the boundary of the ablated tissue, a second sampling density is used to determine target sampling points;

[0063] For a third region around a blood vessel with a diameter greater than a preset value and a fourth region around a heat exchange zone of the ablation needle, the sampling density is increased to determine a target sampling point, and the sampling density increase in the third region is lower than that in the fourth region;

[0064] For the remaining areas, the third sampling density is used to determine the target sampling points;

[0065] The second region is located in a region of the first region far away from the boundary of the ablated tissue. The second sampling density is smaller than the first sampling density, and the third sampling density is smaller than the second sampling density.

[0066] Through the above-mentioned partitioned adaptive sampling strategy, temperature information of regions of interest with different clinical importance can be obtained in a targeted manner to provide accurate temperature data support for cryoablation surgery.

[0067] In one example, the ablation tissue is a tumor, and the target sampling points are spatial sampling points in the area around the ablation needle. , used for subsequent temperature field solution, a partitioned adaptive sampling strategy was adopted, and the sampling density was adjusted according to clinical importance: 5mm area around the tumor boundary: high-density sampling (1.0mm spacing); 5-15mm area around the tumor boundary: medium-density sampling (1.5mm spacing); area around large blood vessels (>3mm): 50% density; around the ablation needle heat exchange area: 100% density; remaining areas: low-density sampling (2.0-3.0mm).

[0068] The method of this embodiment constructs a vascular network diagram based on the three-dimensional imaging data before the cryoablation surgery, and maps the vascular bifurcation points and vascular connection relationships into nodes and edges in the vascular network diagram, respectively. During the cryoablation process, the blood perfusion coefficient of each node in the vascular network diagram is predicted through a graph neural network to determine the blood perfusion coefficient contained in the biological heat conduction equation, and the temperature field of the target sampling point is predicted using the biological heat conduction equation as a constraint condition. This can accurately simulate the influence of blood vessels on heat dissipation, making the temperature prediction conform to physical laws while realizing real-time prediction during the operation.

[0069] In a specific example, the flow chart of temperature prediction is as follows Figure 2 As shown in Figure 2, the architecture of the temperature prediction model is as follows: Figure 3 As shown. Figure 3 In the hybrid architecture shown, which combines a physical constraint-based backbone network with a data-driven vascular effect sub-network, cPINN serves as the physical constraint backbone network, and its network structure includes:

[0070] Input layer: 5-dimensional tensor ( x , y , z , t , M ),in,( x , y , z ) represents the spatial coordinates of the target sampling point, t Indicates time, M Represents the blood vessel mask (0-1 value, 0-not a blood vessel, 1-a blood vessel). The blood vessel mask value can be determined based on the tissue segmentation result.

[0071] Hidden layer: 8 fully connected layers, 256 neurons in each layer, using SiLU activation function.

[0072] Skip connection: A residual module is added between every two layers to achieve skip connection.

[0073] Output layer: temperature value T ( x , y , z , t ).

[0074] The physical constraint uses the Pennes-based biological heat conduction equation. The original Pennes equation has limitations when simulating processes involving phase changes, such as tissue freezing / thawing. Therefore, in this example, the biological heat conduction equation with a phase change latent heat correction term is used:

[0075]

[0076] Where, Indicates tissue density (kg / m³), represents the temperature-dependent specific heat capacity ; represents the thermal conductivity term, represents the divergence operator, Represents the temperature field T( x , y , z , t )’s gradient; represents the temperature-dependent thermal conductivity ; Represents blood perfusion coefficient , provided by graph neural networks; Indicates blood density (kg / m³); Represents the specific heat capacity of blood ; represents blood flow temperature (°C); represents metabolic heat generation rate (W / m³); represents the correction term for the latent heat of phase change (W / m³).

[0077] Tissue density ρ , temperature-dependent specific heat capacity c , temperature-dependent thermal conductivity λ , as thermophysical properties of tissue, average values ​​or ranges can be obtained from published literature for specific organs (e.g., liver, kidney, lung). These parameters may also vary with temperature.

[0078] Blood density ρb Specific heat capacity of blood cb , blood flow temperature Tb , as a standard physiological parameter, recognized values ​​can be obtained from medical literature, blood flow temperature Tb This is usually assumed to be core body temperature (e.g. 37°C).

[0079] The metabolic heat generation rate, Qm, can be obtained from the literature for some tissues.

[0080] The latent heat of phase change (Qphase) is treated by the effective specific heat method, in which the latent heat of phase change L and the midpoint temperature of phase change Tm are also known physical properties of the material (water / tissue).

[0081] The effective specific heat method is an approximate method for dealing with the latent heat effect of phase change. That is, the influence of the latent heat of phase change is equivalently incorporated into the specific heat capacity, so that the traditional heat conduction equation based on specific heat capacity can still be used for calculation. Specifically, the effective specific heat capacity is defined as the basic specific heat capacity plus a term related to the phase change:

[0082]

[0083] Where, represents the effective specific heat capacity, represents the basic specific heat capacity, is the latent heat of phase change (J / kg), For phase transition completion, smooth Type Function express:

[0084]

[0085] Where, represents the smoothing coefficient, represents the phase transition midpoint temperature, Used to generate smooth functions.

[0086] cPINN boundary conditions include Dirichlet boundary conditions (temperature in the heat exchange zone of the ablation needle), Neumann boundary conditions (far-field temperature), and initial conditions (initial temperature). These conditions ensure the existence and uniqueness of the model solution and accurately describe the thermal behavior of biological tissue during ablation therapy:

[0087] The temperature of the heat exchange zone of the ablation needle is set to the temperature of the surface of the ablation needle, that is:

[0088]

[0089] Where, Indicates the surface temperature of the ablation needle, which is determined by the internal cooling medium temperature The probe measuring the temperature of the refrigerant medium has an inherent thermal resistance and is higher by a temperature difference ΔT. ΔT can be measured experimentally and is an empirical constant for a specific probe model and flow rate.

[0090] Far-field temperature: In order to simplify the calculation and focus on the thermal effect of the local area around the ablation needle, it is assumed that the model calculation area is large enough and the heat exchange away from the boundary of the ablation needle can be ignored. Set to human physiological temperature ,Right now:

[0091]

[0092] Initial temperature: The temperature of all spatial positions (x, y, z) at the initial moment All are set to physiological body temperature (e.g. 37°C), the initial temperature is:

[0093]

[0094] In the case where the boundary conditions of the physical information neural network include the surface temperature of the ablation needle, the far-field temperature, and the initial temperature, the loss function should not only measure the gap between the prediction result and the data (if data exists), but more importantly, it should measure the degree of conformity of the prediction result with the physical laws (defined by the partial differential equation (physical constraints - biological heat conduction equation) and the boundary conditions. Therefore, the loss function used in training is is a weighted sum consisting of multiple loss terms. Each loss term corresponds to a specific constraint, including PDE residual, far-field temperature, initial temperature, data matching, and variational energy conservation.

[0095]

[0096] Where, are respectively the first weight coefficient, the second weight coefficient, the third weight coefficient, the fourth weight coefficient and the fifth weight coefficient; Represents the residual loss of the biological heat conduction equation, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the biological heat conduction equation; Represents the far-field temperature loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the far-field temperature; Represents the initial condition loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the initial temperature; represents the data matching loss, which is used to measure the gap between the temperature field predicted by the physical information neural network and the actual data; Represents the variational energy conservation loss, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the law of conservation of energy.

[0097] exist Figure 3 In the architecture shown, the graph neural network is used as a subnetwork, and its graph construction principle is as follows:

[0098] node : Blood vessel bifurcation points and key sampling points. For example, key sampling points can be selected along the center line of the blood vessel at a certain distance interval between the key bifurcation points.

[0099] side : Physical adjacency, that is, vascular connection, for example, there is a directly connected vascular segment between two vascular branch points.

[0100] Node Features : For nodes of vascular network:

[0101]

[0102] in, represents the spatial coordinates of the target sampling point, represents the blood vessel diameter, Indicates the type of blood vessels, Indicates the blood vessel density, which refers to the density of blood vessels in the tissue segmentation result within a 3cm radius of the spatial coordinates of the current target sampling point. The direction of blood flow;

[0103] Edge Features :[distance , relative direction ]

[0104] The structure of the graph neural network: The Graph Attention Network (GAT) is the core architecture of the graph neural network. It uses a three-layer GAT, with a feature dimension of 64 per layer. The node feature update formula is:

[0105]

[0106] Where, Indicates the l The features of layer node v, represents the neighbor set of node v, represents the attention coefficient, represents the weight matrix, Represents the nonlinear activation function, attention coefficient:

[0107]

[0108] Where, represents a learnable parameter vector, || represents a concatenation operation, Represents the influence coefficient.

[0109] The first layer of the graph neural network receives the initial node feature vector h0 of all nodes as its input. After passing through multiple layers of GAT, the node feature vector The structural information of the surrounding blood vessels is integrated. The last layer of the graph neural network is a fully connected layer, also known as the mapping function, which transforms this high-dimensional feature vector Mapping (or decoding) these values ​​into numerical values, the blood perfusion coefficients for each node. To more accurately model the differential impact of vessels of different calibers, graph neural networks employ a multi-head attention mechanism to capture the varying influences of different types of neighboring nodes on the central node (the node currently being calculated). Specifically, multiple attention heads (for example, three) can be configured, each learning the influence coefficient for a specific vessel caliber (for example, large vessels (>5mm), medium vessels (2-5mm), and microvessels (<2mm)).

[0110] Vascular networks are naturally graph structures. Vascular bifurcations can be considered nodes, and vascular segments can be considered edges. By constructing a vascular network graph and using graph neural networks to learn its features, we can effectively extract structural information from the vascular network and convert it into blood perfusion coefficients, providing more accurate blood perfusion parameters for the backbone network.

[0111] In some embodiments, predicting the temperature field of a target sampling point using a pre-trained temperature prediction model includes:

[0112] Step 1031 , predicting the blood perfusion coefficient of each node in the vascular network graph through a graph neural network.

[0113] Step 1032: Perform interpolation processing on the blood perfusion coefficient to determine the blood perfusion coefficient of each target sampling point.

[0114] In the specific implementation, since the target sampling point may be a node in the non-vascular network graph (the mask value is 0, that is, non-vascular), and the blood perfusion coefficient of each node in the vascular network graph is predicted by the graph neural network, in order to obtain the blood perfusion coefficient of all target sampling points and solve the biological heat conduction equation, the predicted blood perfusion coefficient of each node in the vascular network graph can be extended to the entire spatial domain through interpolation processing (for example, nearest neighbor interpolation) as the blood perfusion coefficient of each spatial point in cPINN. enter.

[0115] Step 1033: Substitute the blood perfusion coefficient of each target sampling point into the biological heat conduction equation, use the equation as a constraint condition, and predict the temperature of each target sampling point through a physical information neural network.

[0116] In some embodiments, as Figure 3 As shown, the temperature prediction model also includes a gated fusion network for correcting the blood perfusion coefficient of each target sampling point using the following calculation formula before substituting the blood perfusion coefficient of each target sampling point into the biological heat conduction equation;

[0117]

[0118] Where, represents the corrected blood perfusion coefficient of the target sampling point, Represents the weight coefficient, which can be dynamically adjusted based on the distance between the current target sampling point and the nearest blood vessel and the ablation tissue type. represents the blood perfusion coefficient of the target sampling point before correction, that is, the blood perfusion coefficient obtained after the above interpolation processing, It represents a pre-set baseline blood perfusion coefficient, for example, the basal perfusion coefficient value applicable to normal tissues. Depending on the target organ to be treated (such as the liver), a recognized average value is selected from authoritative literature.

[0119] To improve the robustness and adaptability of the model, this embodiment introduces a gated fusion mechanism to adaptively adjust the weight coefficient according to tissue type and vascular distance (for example, the distance between the target sampling point and the nearest blood vessel). This allows the contribution ratio of the blood perfusion coefficient output by the GNN sub-network to the cPINN backbone network to be dynamically adjusted based on spatial position in different tissue regions and at different vascular distances, making the prediction results of the backbone network more consistent with physical laws.

[0120] The gated fusion mechanism can use a small neural network (e.g., a fully connected network) as the gated fusion network. The input of the gated fusion network includes tissue type information and vascular distance information, and the output is the weight coefficient And the corrected blood perfusion coefficient of the target sampling point. Weight coefficient Between 0 and 1, it is used to control the contribution ratio of the blood perfusion coefficient obtained by GNN prediction and interpolation processing to the final blood perfusion coefficient.

[0121] In some embodiments, the above-mentioned cryoablation temperature prediction method further includes:

[0122] Step 100, constructing and training a temperature prediction model;

[0123] Training the temperature prediction model further includes:

[0124] Step 1001: Use the synthetic training data set to perform unsupervised training on the constructed temperature prediction model.

[0125] In the specific implementation, a simplified geometric model (spherical tumor + cylindrical blood vessels) is constructed, and a simplified biological heat conduction equation is solved (ignoring some complex physical effects, such as metabolic heat generation, and simplifying the blood perfusion term). These data constitute a synthetic training dataset for the unsupervised training stage, and a virtual vascular topology structure is constructed, including vascular networks of different diameters and densities. The constructed temperature prediction model is used to obtain the temperature field distribution in different scenarios for unsupervised training.

[0126] In the unsupervised training process, the loss function used is as follows:

[0127]

[0128] Where, are respectively the first weight coefficient, the second weight coefficient, the third weight coefficient, the fourth weight coefficient and the fifth weight coefficient; Represents the residual loss of the biological heat conduction equation, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the biological heat conduction equation; Represents the far-field temperature loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the far-field temperature; Represents the initial condition loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the initial temperature; represents the data matching loss, which is used to measure the gap between the temperature field predicted by the physical information neural network and the actual data; Represents the variational energy conservation loss, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the law of conservation of energy.

[0129] The initial learning rate for training is 10 -3 , cosine annealing strategy, batch size 128, number of training phase iterations: 200000 times, by optimizing the loss function (as the objective function) , and obtain the temperature prediction model after unsupervised training.

[0130] Step 1002: construct a finite element model based on the three-dimensional imaging data of a real patient, and use the finite element model to solve the biological heat conduction equation to obtain the reference temperature field of the target sampling point; use the temperature prediction model after unsupervised training based on the three-dimensional imaging data of the real patient to obtain the predicted temperature field of the target sampling point, and optimize the temperature prediction model with the minimum value of the first objective function as the goal. The first objective function is ,in, represents the mean square error between the predicted temperature field and the reference temperature field, λ represents the trade-off coefficient, Represents the first objective function value.

[0131] In one example, a detailed finite element model was constructed based on 3D medical images (CT / MRI) of 5-10 real patients, and the complete (non-simplified) biological heat conduction equation was solved to generate a reference temperature field. The combined scenarios of different tissue parameters (such as tumor density, specific heat capacity, etc.), vascular layout (vascular map structure) and ablation parameters (including the heat exchange area length of the ablation needle, the diameter of the ablation needle, etc.) were simulated. The predicted temperature field of the target sampling point was obtained by using the temperature prediction model after unsupervised training based on the 3D imaging data of real patients. The temperature prediction model was optimized with the goal of minimizing the value of the first objective function, and supervised fine-tuning based on finite elements was achieved. The initial value of the trade-off coefficient was set to 0.5 and gradually increased to 2.0. A low learning rate (10 -4 ) and early stopping strategy.

[0132] Step 1003: Adjust the optimized temperature prediction model by adjusting the parameters of the physical information neural network, the parameters of the graph neural network, and the parameters of the biological heat conduction equation so that the predicted temperature field obtained by the optimized temperature prediction model based on the three-dimensional imaging data of the real patient is close to the reference temperature field.

[0133] In one example, the parameters of the graph neural network, the parameters related to the latent heat of phase change in the biological heat conduction equation (phase transition midpoint temperature, smoothing coefficient), and the last two layers of cPINN ( Figure 3 The weights of the hidden layer 8 and the output layer in the model are fine-tuned to further improve the generalization of the temperature prediction model and make it closer to the finite element results (reference temperature field).

[0134] In some embodiments, the above-mentioned cryoablation temperature prediction method also includes: constructing a set of thermophysical parameters corresponding to different ablation tissues of different patients to solve the biological heat conduction equation when predicting the temperature field of the target sampling point, and the thermophysical parameter set includes the thermophysical parameters contained in the biological heat conduction equation.

[0135] For the same organ, thermophysical parameters may vary between patients. Therefore, in practical applications, the distribution range of thermophysical parameters for the same organ in different patients can be analyzed based on clinical case databases. Blood perfusion characteristics can vary significantly for different organs (such as the liver, kidneys, and lungs). To improve the organ specificity of the model, different thermophysical parameters can be set for different organs. Thermophysical parameters for different organs, such as the liver, kidneys, and lungs, can be collected from literature or experimental measurements to construct organ-specific parameter sets. When predicting the cryoablation temperature for a specific organ, the corresponding organ's thermophysical parameter set can be used to solve the biological heat conduction equation to improve the model's prediction accuracy and organ adaptability.

[0136] Furthermore, the model's generalizability can be evaluated using independent clinical datasets not used in training, employing strategies such as K-fold cross-validation, to comprehensively assess its practical application. Evaluation metrics include temperature prediction accuracy (such as the root mean square error (RMSE) between the predicted and measured temperatures) and the degree of match between key isotherms (such as the Dice coefficient, which measures the similarity between the predicted contours of key isotherms like 0°C and -20°C and the actual ablation boundaries observed in postoperative imaging). Furthermore, the model's performance is compared with existing baseline methods to analyze its robustness and sensitivity to input perturbations and parameter variations. Clinician feedback is also used to assess the clinical relevance and practical value of the predictions.

[0137] In one example, the above stage-by-stage training process is as follows Figure 4 As shown in the figure, the method includes an unsupervised physical pre-training phase, a finite element metadata fine-tuning phase, and a clinical data-adaptive generalization phase. The phased training strategy includes: unsupervised physical pre-training: initial training based on synthetic data and physical constraints; finite element metadata fine-tuning: supervised fine-tuning based on high-fidelity finite element metadata; and clinical data-adaptive generalization: organ-specific parameter optimization based on clinical data. The data flow, optimization objectives, learning rate adjustment strategies, and evaluation metrics for each phase are also shown. This phased training strategy addresses the scarcity of clinical data.

[0138] The existing technology still lacks the ability to dynamically correct the temperature field in real time during cryoablation. Therefore, in some embodiments, the above-mentioned cryoablation temperature prediction method further includes:

[0139] Step 201 : During cryoablation, the temperature of the refrigerant at a first target sampling point located in a heat exchange area of ​​an ablation needle is collected and measured.

[0140] Specifically, a high-precision (±0.1°C) refrigerant temperature sensor is used to collect the refrigerant temperature at any target sampling point near the heat exchange area inside the ablation needle (for example, near the heat exchange area on the refrigerant return path) in real time at a frequency of 10 times per second. Furthermore, the refrigerant temperature data is preprocessed, specifically using a 5-second moving average method for smoothing to filter out random noise and obtain a more stable temperature measurement value, which is defined as .

[0141] Step 202: Using the temperature prediction model, the predicted temperature of the first target sampling point is obtained under the currently set blood flow velocity and needle tip heat transfer efficiency factor.

[0142] In a specific implementation, the blood flow velocity and the needle tip heat transfer efficiency factor are given parameters when the temperature prediction model is used to predict the temperature of the target sampling point.

[0143] Step 203 : Based on the refrigerant temperature and the predicted temperature at the first target sampling point, an inverse calculation formula is used to perform an inverse calculation on the blood flow velocity and the needle tip heat transfer efficiency factor.

[0144] In one example, during a cryoablation procedure, a calibration process for inverse calculations is initiated every 5 seconds, using the latest measured temperature. To reversely infer and optimize the key dynamic parameters in the temperature prediction model that have a significant impact on the temperature of the heat exchange area: blood flow velocity vector ( v ) and tip heat transfer efficiency factor ( η ).

[0145] The inversion calculation formula is as follows:

[0146]

[0147] Where, represents the inverted blood flow velocity vector, represents the tip heat transfer efficiency factor after inversion, Indicates the refrigerant temperature (actual temperature) of the first target sampling point. Indicates the predicted temperature of the first target sampling point.

[0148] Through inversion calculation, physiological parameters that are difficult to measure directly and have a significant impact on the local temperature field (such as the cooling effect brought by local blood flow) and device interaction parameters (such as the actual heat transfer efficiency between the needle tip and the tissue) can be inferred in real time from the easily accessible refrigerant medium temperature through a computationally efficient inversion calculation model, thereby dynamically correcting the parts of the physical temperature prediction model that are difficult to accurately model, thereby improving the accuracy of temperature prediction.

[0149] Step 204 : Smoothing the inversion calculation results to obtain updated blood flow velocity and needle tip heat transfer efficiency factor.

[0150] In order to avoid model oscillation caused by excessive parameter updates, the Exponential Moving Average (EMA) method is used to update the parameters obtained by inversion calculation. and After smoothing, update the model.

[0151] Updated parameters and The calculation is as follows:

[0152]

[0153]

[0154] Where, represents a smoothing coefficient, for example, 0.7, which ensures the stability of parameter updates. Key parameters in the physical model (v and η) are updated every 15 seconds based on the latest smoothing results. In subsequent temperature predictions, cPINN uses the updated v and η as parameters in its physical model to predict the new temperature field.

[0155] Step 205 : updating the biological heat conduction equation according to the updated blood flow velocity and needle tip heat transfer efficiency factor to update the temperature prediction model.

[0156] By monitoring the refrigerant medium and inverting its parameters, temperature prediction can take into account both computational efficiency and individual adaptability, thereby improving the safety of cryoablation surgery.

[0157] In some embodiments, the above-mentioned cryoablation temperature prediction method further includes:

[0158] Step 104: Acquire a calibration image after ablation is completed.

[0159] Specifically, final calibration images were obtained after cryoablation was completed, using low-dose CT or MRI scans.

[0160] Step 105 , segmenting the calibration image to obtain the target isothermal surface contour segmentation result; determining the target isothermal surface prediction result based on the temperature field of the target sampling point predicted by the temperature prediction model; registering the target isothermal surface contour segmentation result with the target isothermal surface prediction result, and calculating the similarity index.

[0161] Specifically, the 0°C isothermal surface contour is segmented and extracted from the calibration image. , the predicted 0°C isothermal surface It is aligned with the measured 0°C isothermal surface and shape similarity metrics are calculated, including the Dice coefficient and Hausdorff distance.

[0162] Step 106: Calculate the difference loss function value based on the calculated similarity index, and adjust the thermophysical parameter set with the minimum difference loss function value as the optimization goal. The thermophysical parameter set includes the thermophysical parameters included in the biological heat conduction equation.

[0163] Specifically, the Dice coefficient ( ) and Hausdorff distance ( ) measure the difference and construct the difference loss function.

[0164] In one example, based on the predicted 0°C isotherm Compared with the measured 0°C isothermal surface extracted from the final calibration image The difference between them is used to construct the following difference loss function :

[0165]

[0166] Where, Represents the weight coefficient of Hausdorff distance.

[0167] Based on the thermophysical parameters at the minimum of this difference loss function, the tissue thermophysical parameter set involved in the biological heat conduction equation is updated , λ(T) represents the temperature-dependent thermal conductivity, c(T) represents the temperature-dependent specific heat capacity, represents the blood perfusion coefficient, and a Bayesian optimization framework is used to generate posterior probability distributions for these thermophysical parameters. Then, in subsequent surgeries or new predictions, the cPINN network uses these optimized tissue thermophysical parameters Θ that are closer to the actual situation to perform calculations to obtain more accurate prediction results.

[0168] Through the final image calibration, the static or quasi-static physical parameters describing the inherent thermal properties of tissues in the Pennes biological heat conduction equation are inversely optimized. , in order to make the entire physical model closer to the actual situation of a specific patient or a specific organ, thereby improving the accuracy of future predictions and achieving personalized calibration.

[0169] Step 107 : Based on the adjusted set of thermophysical parameters, the biological heat conduction equation is updated, and the temperature field of the target sampling point is predicted using the updated biological heat conduction equation as a constraint condition.

[0170] Through the correction of cooling medium and terminal image feedback, multimodal feedback is achieved, which avoids frequent radiation exposure during surgery, enables temperature prediction to take into account both computational efficiency and individual adaptability, and improves the safety of cryoablation surgery.

[0171] according to Figure 2 The figure shows the cryoablation temperature prediction process. The prediction process includes data input, model prediction and feedback correction. By processing preoperative CT / MRI images and real-time refrigerant temperature monitoring data, the cPINN backbone network and GNN vascular sub-network are used to predict the temperature field. Feedback correction is performed based on the refrigerant temperature feedback and final image calibration to form a closed-loop control system and achieve accurate temperature field prediction.

[0172] Compared with the prior art, the method of this embodiment has the following beneficial effects:

[0173] (1) Significantly improved computational efficiency: Compared with traditional finite element methods, the inference speed is increased by more than 100 times, achieving sub-second temperature field updates. At the same time, it can also meet the real-time visualization requirements during surgery through three-dimensional isothermal surface rendering.

[0174] (2) Enhanced individual adaptability: It has the ability to automatically integrate patient-specific anatomical structures and physiological parameters and adapt to the differences in the thermal physical properties of multiple organs (liver, kidney, and lung). It also explicitly models the heat dissipation effect of the vascular network through graph neural networks, improving the prediction accuracy near large blood vessels.

[0175] (3) Closed-loop feedback mechanism: Real-time parameter adjustment is achieved through refrigerant temperature feedback without additional radiation exposure; final image calibration provides model self-correction capabilities, improving the accuracy of subsequent surgeries. Uncertainty quantification provides a safety margin reference for doctors' decision-making.

[0176] (4) Clinical feasibility: Only routine preoperative imaging and ablation device temperature data are required, no additional equipment is required; compatible with existing cryoablation workflow, no need to change surgical procedures.

[0177] Example 2:

[0178] Another embodiment of the present application relates to a cryoablation temperature prediction device. The implementation details of the cryoablation temperature prediction device of this embodiment are specifically described below. The following content is only for the convenience of understanding the implementation details provided by this solution and is not necessary for the implementation of this solution. The schematic diagram of the cryoablation temperature prediction device of this embodiment can be as follows: Figure 5 As shown, it includes a data input module 201, a data processing module 202 and a model prediction module 203.

[0179] Data input module 201, used to obtain three-dimensional image data before cryoablation surgery;

[0180] A data processing module 202 is configured to construct a vascular network diagram based on the vascular centerlines in the three-dimensional image data, wherein the nodes in the vascular network diagram represent vascular bifurcation points, and the edges in the vascular network diagram represent vascular connection relationships;

[0181] The model prediction module 203 is used to predict the temperature field of the target sampling point during the cryoablation process through a pre-trained temperature prediction model. The temperature prediction model includes a physical information neural network and a graph neural network. The physical information neural network uses the biological heat conduction equation as a constraint condition to predict the temperature field of the target sampling point. The graph neural network is used to predict the blood perfusion coefficient of each node in the vascular network diagram to determine the blood perfusion coefficient contained in the biological heat conduction equation.

[0182] Specifically, the three-dimensional imaging data before the cryoablation surgery may be the patient's pre-operative CT (Computed Tomography) / MRI (Magnetic Resonance Imaging) three-dimensional imaging data.

[0183] By integrating physical constraints based on the biological heat conduction equation with data-driven cryoablation temperature prediction based on graph neural networks, safe, real-time, and highly accurate temperature field predictions can be achieved. The physical information neural network is a machine learning model that combines deep learning and physics knowledge. As the backbone network for physical constraints, it ensures the basic physical rationality of the model. Using the biological heat conduction equation as a constraint, the physical information neural network ensures that the model's predictions conform to the laws of thermodynamics. Furthermore, as a data-driven vascular effect subnetwork, the graph neural network can learn and compensate for parts of the model that are difficult to accurately model, such as the impact of the vascular network on blood perfusion. It can learn the complex relationship between vascular distribution and blood perfusion coefficient from data and provide accurate blood perfusion coefficients to the backbone network.

[0184] In specific implementations, Continuity-based Physics-Informed Neural Networks (cPINN) can be used as PINNs for physics-informed neural networks (PINNs). Nodes in graph neural networks (GNNs) can also include key points selected at regular intervals between adjacent vascular bifurcations along the centerline of the vessel.

[0185] In some embodiments, a vascular network diagram is constructed using vascular centerlines in three-dimensional image data, including: performing tissue segmentation on the three-dimensional image data to obtain tissue segmentation results; extracting vascular centerlines in the tissue segmentation results through skeletonization processing; identifying vascular bifurcation points in the vascular centerlines; constructing a vascular network topology structure based on the vascular centerlines, and constructing a vascular network diagram based on the vascular network topology structure, wherein nodes in the vascular network diagram represent vascular bifurcation points, and edges in the vascular network diagram represent vascular connection relationships.

[0186] In specific implementation, a deep learning segmentation network (such as U-Net) can be used to extract the tumor area, vascular network and surrounding tissue as the tissue segmentation result, which is used to extract the vascular centerline and 3D reconstruction. Furthermore, a 3D reconstruction model is constructed based on the tissue segmentation result to generate a 3D point cloud representation. ,in, Indicates tissue attribute markers, such as liver; locates the ablation needle placement position and heat exchange zone parameters based on tissue segmentation results. Heat exchange zone parameters include needle tip position , effective heat exchange area length and needle diameter .

[0187] In the specific implementation, the vascular bifurcation points are identified by analyzing the vascular intersections and endpoints on the vascular centerline, and then the key bifurcation points are identified as the identification results of this step using the vascular diameter change rate. The key bifurcation points are, for example, the root nodes of the root blood vessels. In actual applications, the key bifurcation points selected for different organs may be different, and this is not limited here.

[0188] In the specific implementation, construct a vascular network diagram G =( V , E ), node set V Represents the identified blood vessel bifurcation points, edge set E Indicates the vascular connection relationship. Furthermore, physical feature vectors are assigned to the nodes of the vascular network graph as node features in the subsequent graph neural network. The physical feature vectors include:

[0189] Blood vessel type : Such as hepatic vein (0), hepatic artery (1), portal vein (2);

[0190] Blood vessel diameter : The measurement value obtained from the three-dimensional image, unit: mm;

[0191] Blood flow velocity : The initial value estimated based on the blood vessel type, in cm / s, can be determined by searching existing literature;

[0192] Blood flow direction : three-dimensional unit vector (dx, dy, dz)

[0193] In some embodiments, the data processing module 202 is further configured to: adopt the following strategy to determine target sampling points to form a target sampling point set, so as to perform temperature field solution for these target sampling points in subsequent steps.

[0194] Among them, different strategies include:

[0195] For a first area around the boundary of the ablated tissue, a first sampling density is used to determine target sampling points;

[0196] For a second area around the boundary of the ablated tissue, a second sampling density is used to determine target sampling points;

[0197] For a third region around a blood vessel with a diameter greater than a preset value and a fourth region around a heat exchange zone of the ablation needle, the sampling density is increased to determine a target sampling point, and the sampling density increase in the third region is lower than that in the fourth region;

[0198] For the remaining areas, the third sampling density is used to determine the target sampling points;

[0199] The second region is located in a region of the first region far away from the boundary of the ablated tissue. The second sampling density is smaller than the first sampling density, and the third sampling density is smaller than the second sampling density.

[0200] Through the above-mentioned partitioned adaptive sampling strategy, temperature information of regions of interest with different clinical importance can be obtained in a targeted manner to provide accurate temperature data support for cryoablation surgery.

[0201] This embodiment constructs a vascular network diagram based on the three-dimensional imaging data before the cryoablation surgery, and maps the vascular bifurcation points and vascular connection relationships into nodes and edges in the vascular network diagram, respectively. During the cryoablation process, the blood perfusion coefficient of each node in the vascular network diagram is predicted through a graph neural network to determine the blood perfusion coefficient contained in the biological heat conduction equation, and the temperature field of the target sampling point is predicted using the biological heat conduction equation as a constraint condition. This can accurately simulate the influence of blood vessels on heat dissipation, making the temperature prediction conform to physical laws while realizing real-time prediction during the operation.

[0202] In some embodiments, a pre-trained temperature prediction model is used to predict the temperature field of the target sampling point, including: using a graph neural network to predict the blood perfusion coefficient of each node in the vascular network diagram; interpolating the blood perfusion coefficient of each node in the vascular network diagram predicted by the graph neural network to determine the blood perfusion coefficient of each target sampling point; substituting the blood perfusion coefficient of each target sampling point into the biological heat conduction equation, using the biological heat conduction equation as a constraint condition, and using a physical information neural network to predict the temperature of each target sampling point.

[0203] In the specific implementation, since the target sampling point may be a node in the non-vascular network graph (the mask value is 0, that is, non-vascular), and the blood perfusion coefficient of each node in the vascular network graph is predicted by the graph neural network, in order to obtain the blood perfusion coefficient of all target sampling points and solve the biological heat conduction equation, the predicted blood perfusion coefficient of each node in the vascular network graph can be extended to the entire spatial domain through interpolation processing (for example, nearest neighbor interpolation) as the blood perfusion coefficient of each spatial point in cPINN. enter.

[0204] In some embodiments, as Figure 3 As shown, the temperature prediction model also includes a gated fusion network for correcting the blood perfusion coefficient of each target sampling point using the following calculation formula before substituting the blood perfusion coefficient of each target sampling point into the biological heat conduction equation;

[0205]

[0206] Where, represents the corrected blood perfusion coefficient of the target sampling point, Represents the weight coefficient, which can be dynamically adjusted based on the distance between the current target sampling point and the nearest blood vessel and the ablation tissue type. represents the blood perfusion coefficient of the target sampling point before correction, that is, the blood perfusion coefficient obtained after the above interpolation processing, Represents a pre-set baseline blood perfusion coefficient. A base perfusion coefficient value applicable to normal tissues, selected from authoritative literature as a recognized average value based on the target organ to be treated (e.g., liver).

[0207] To improve the robustness and adaptability of the model, this embodiment introduces a gated fusion mechanism to adaptively adjust the weight coefficient according to tissue type and vascular distance (for example, the distance between the target sampling point and the nearest blood vessel). This allows the contribution ratio of the blood perfusion coefficient output by the GNN sub-network to the cPINN backbone network to be dynamically adjusted based on spatial position in different tissue regions and at different vascular distances, making the prediction results of the backbone network more consistent with physical laws.

[0208] The gated fusion mechanism can use a small neural network (e.g., a fully connected network) as the gated fusion network. The input of the gated fusion network includes tissue type information and vascular distance information, and the output is the weight coefficient And the corrected blood perfusion coefficient of the target sampling point. Weight coefficient Between 0 and 1, it is used to control the contribution ratio of the blood perfusion coefficient obtained by GNN prediction and interpolation processing to the final blood perfusion coefficient.

[0209] In some embodiments, the above-mentioned cryoablation temperature prediction device further includes a model training module for constructing and training a temperature prediction model; training the temperature prediction model further includes: using a synthetic training data set to perform unsupervised training on the constructed temperature prediction model; constructing a finite element model based on the three-dimensional imaging data of a real patient, and using the finite element model to solve the biological heat conduction equation to obtain a reference temperature field of the target sampling point; using the temperature prediction model after unsupervised training based on the three-dimensional imaging data of the real patient to obtain a predicted temperature field of the target sampling point, and optimizing the temperature prediction model with the minimum value of the first objective function as the goal, and the first objective function is ,in, represents the mean square error between the predicted temperature field and the reference temperature field, λ represents the trade-off coefficient, represents the value of the first objective function; by adjusting the parameters of the physical information neural network, the parameters of the graph neural network, and the parameters of the biological heat conduction equation, the optimized temperature prediction model is adjusted so that the predicted temperature field obtained by the optimized temperature prediction model based on the three-dimensional imaging data of the real patient is close to the reference temperature field.

[0210] In the specific implementation, a simplified geometric model (spherical tumor + cylindrical blood vessels) is constructed, and a simplified biological heat conduction equation is solved (ignoring some complex physical effects, such as metabolic heat generation, and simplifying the blood perfusion term). These data constitute a synthetic training dataset for the unsupervised training stage, and a virtual vascular topology structure is constructed, including vascular networks of different diameters and densities. The constructed temperature prediction model is used to obtain the temperature field distribution in different scenarios for unsupervised training.

[0211] In the unsupervised training process, the loss function used is as follows:

[0212]

[0213] Where, are respectively the first weight coefficient, the second weight coefficient, the third weight coefficient, the fourth weight coefficient and the fifth weight coefficient; Represents the residual loss of the biological heat conduction equation, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the biological heat conduction equation; Represents the far-field temperature loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the far-field temperature; Represents the initial condition loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the initial temperature; represents the data matching loss, which is used to measure the gap between the temperature field predicted by the physical information neural network and the actual data; Represents the variational energy conservation loss, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the law of conservation of energy.

[0214] The initial learning rate for training is 10 -3 , cosine annealing strategy, batch size 128, number of training phase iterations: 200000 times, by optimizing the loss function (as the objective function) , and obtain the temperature prediction model after unsupervised training.

[0215] In one example, a detailed finite element model was constructed based on 3D medical images (CT / MRI) of 5-10 real patients, and the complete (non-simplified) biological heat conduction equation was solved to generate a reference temperature field. The combined scenarios of different tissue parameters (such as tumor density, specific heat capacity, etc.), vascular layout (vascular map structure) and ablation parameters (including the heat exchange area length of the ablation needle, the diameter of the ablation needle, etc.) were simulated. The predicted temperature field of the target sampling point was obtained by using the temperature prediction model after unsupervised training based on the 3D imaging data of real patients. The temperature prediction model was optimized with the goal of minimizing the value of the first objective function, and supervised fine-tuning based on finite elements was achieved. The initial value of the trade-off coefficient was set to 0.5 and gradually increased to 2.0. A low learning rate (10-4 ) and early stopping strategy.

[0216] In one example, the parameters of the graph neural network, the parameters related to the latent heat of phase change in the biological heat conduction equation (phase transition midpoint temperature, smoothing coefficient), and the last two layers of cPINN ( Figure 3 The weights of the hidden layer 8 and the output layer in the model are fine-tuned to further improve the generalization of the temperature prediction model and make it closer to the finite element results (reference temperature field).

[0217] In some embodiments, the model training module is also used to: construct a set of thermophysical parameters corresponding to different ablation tissues of different patients to solve the biological heat conduction equation when predicting the temperature field of the target sampling point, and the thermophysical parameter set includes the thermophysical parameters contained in the biological heat conduction equation.

[0218] For the same organ, thermophysical parameters may vary between patients. Therefore, in practical applications, the distribution range of thermophysical parameters for the same organ in different patients can be analyzed based on clinical case databases. Blood perfusion characteristics can vary significantly for different organs (such as the liver, kidneys, and lungs). To improve the organ specificity of the model, different thermophysical parameters can be set for different organs. Thermophysical parameters for different organs, such as the liver, kidneys, and lungs, can be collected from literature or experimental measurements to construct organ-specific parameter sets. When predicting the cryoablation temperature for a specific organ, the corresponding organ's thermophysical parameter set can be used to solve the biological heat conduction equation to improve the model's prediction accuracy and organ adaptability.

[0219] Furthermore, the model's generalizability can be evaluated using independent clinical datasets not used in training, employing strategies such as K-fold cross-validation, to comprehensively assess its practical application. Evaluation metrics include temperature prediction accuracy (such as the root mean square error (RMSE) between the predicted and measured temperatures) and the degree of match between key isotherms (such as the Dice coefficient, which measures the similarity between the predicted contours of key isotherms like 0°C and -20°C and the actual ablation boundaries observed in postoperative imaging). Furthermore, the model's performance is compared with existing baseline methods to analyze its robustness and sensitivity to input perturbations and parameter variations. Clinician feedback is also used to assess the clinical relevance and practical value of the predictions.

[0220] In some embodiments, the above-mentioned cryoablation temperature prediction device also includes a feedback correction module, which is used to: collect and measure the refrigerant medium temperature of the first target sampling point located in the heat exchange area of ​​the ablation needle during the cryoablation process; use the temperature prediction model to obtain the predicted temperature of the first target sampling point under the currently set blood flow velocity and needle tip heat transfer efficiency factor; based on the refrigerant medium temperature and the predicted temperature of the first target sampling point, use the inversion calculation formula to inversely calculate the blood flow velocity and needle tip heat transfer efficiency factor; smooth the inversion calculation results to obtain updated blood flow velocity and needle tip heat transfer efficiency factor; update the biological heat conduction equation according to the updated blood flow velocity and needle tip heat transfer efficiency factor to update the temperature prediction model.

[0221] Specifically, a high-precision (±0.1°C) refrigerant temperature sensor is used to collect the refrigerant temperature at any target sampling point near the heat exchange area inside the ablation needle (for example, near the heat exchange area on the refrigerant return path) in real time at a frequency of 10 times per second. Furthermore, the refrigerant temperature data is preprocessed, specifically using a 5-second moving average method for smoothing to filter out random noise and obtain a more stable temperature measurement value, which is defined as .

[0222] In one example, during a cryoablation procedure, a calibration process for inverse calculations is initiated every 5 seconds, using the latest measured temperature. To reversely infer and optimize the key dynamic parameters in the temperature prediction model that have a significant impact on the temperature of the heat exchange area: blood flow velocity vector ( v ) and tip heat transfer efficiency factor ( η ).

[0223] The inversion calculation formula is as follows:

[0224]

[0225] Where, represents the inverted blood flow velocity vector, represents the tip heat transfer efficiency factor after inversion, Indicates the refrigerant temperature (actual temperature) of the first target sampling point. Indicates the predicted temperature of the first target sampling point.

[0226] Through inversion calculation, physiological parameters that are difficult to measure directly and have a significant impact on the local temperature field (such as the cooling effect brought by local blood flow) and device interaction parameters (such as the actual heat transfer efficiency between the needle tip and the tissue) can be inferred in real time from the easily accessible refrigerant medium temperature through a computationally efficient inversion calculation model, thereby dynamically correcting the parts of the physical temperature prediction model that are difficult to accurately model, thereby improving the accuracy of temperature prediction.

[0227] In order to avoid model oscillation caused by excessive parameter updates, the Exponential Moving Average (EMA) method is used to update the parameters obtained by inversion calculation. and After smoothing, update the model.

[0228] Updated parameters and The calculation is as follows:

[0229]

[0230]

[0231] Where, represents a smoothing coefficient, for example, 0.7, which ensures the stability of parameter updates. Key parameters in the physical model (v and η) are updated every 15 seconds based on the latest smoothing results. In subsequent temperature predictions, cPINN uses the updated v and η as parameters in its physical model to predict the new temperature field.

[0232] In some embodiments, the feedback correction module is also used to: obtain a calibration image after ablation is completed; segment the calibration image to obtain a target isothermal surface contour segmentation result; determine the target isothermal surface prediction result based on the temperature field of the target sampling point predicted by the temperature prediction model, align the target isothermal surface contour segmentation result with the target isothermal surface prediction result, and calculate a similarity index; calculate the difference loss function value based on the calculated similarity index, and adjust the thermophysical parameter set with the minimum difference loss function value as the optimization goal, wherein the thermophysical parameter set includes the thermophysical parameters included in the biological heat conduction equation; based on the adjusted thermophysical parameter set, update the biological heat conduction equation, and use the updated biological heat conduction equation as a constraint condition to predict the temperature field of the target sampling point.

[0233] Specifically, after cryoablation is completed, a final calibration image is obtained using low-dose CT or MRI. The 0°C isothermal surface contour is segmented and extracted from the calibration image. , the predicted 0°C isothermal surface It is aligned with the measured 0°C isothermal surface and the shape similarity index is calculated, including the Dice coefficient and Hausdorff distance. ) and Hausdorff distance ( ) measure the difference and construct the difference loss function.

[0234] In one example, based on the predicted 0°C isotherm Compared with the measured 0°C isothermal surface extracted from the final calibration image The difference between them is used to construct the following difference loss function :

[0235]

[0236] Where, Represents the weight coefficient of Hausdorff distance.

[0237] Based on the thermophysical parameters at the minimum of this difference loss function, the tissue thermophysical parameter set involved in the biological heat conduction equation is updated , λ(T) represents the temperature-dependent thermal conductivity, c(T) represents the temperature-dependent specific heat capacity, represents the blood perfusion coefficient, and a Bayesian optimization framework is used to generate posterior probability distributions for these thermophysical parameters. Then, in subsequent surgeries or new predictions, the cPINN network uses these optimized tissue thermophysical parameters Θ that are closer to the actual situation to perform calculations to obtain more accurate prediction results.

[0238] Through the final image calibration, the static or quasi-static physical parameters describing the inherent thermal properties of tissues in the Pennes biological heat conduction equation are inversely optimized. , in order to make the entire physical model closer to the actual situation of a specific patient or a specific organ, thereby improving the accuracy of future predictions and achieving personalized calibration.

[0239] In one example, the hardware connection and software architecture of the cryoablation temperature prediction device and the workstation layout in the clinical application scenario are implemented as follows: Figure 6 As shown, where:

[0240] Hardware components: computing server (implementing data input module, data processing module, model prediction module and feedback correction module), temperature monitoring module (monitoring the temperature of the refrigerant medium), display terminal (displaying three-dimensional isotherms, etc.)

[0241] Software modules: Image processing engine (acquisition of 3D image data, preprocessing, etc.), neural network prediction engine (temperature prediction), parameter inversion module (parameter inversion), visualization module (providing 3D isotherms, etc.)

[0242] User interface: 3D isothermal surface visualization, prediction uncertainty display, parameter adjustment control.

[0243] Among them, prediction uncertainty is achieved through uncertainty quantification, providing support for improving prediction reliability and clinical decision-making.

[0244] This embodiment has at least all the beneficial effects of the first embodiment, which will not be described in detail here.

[0245] It is worth mentioning that all modules involved in this embodiment are logical modules. In actual applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovation of this application, this embodiment does not include units that are not closely related to solving the technical problem proposed by this application. However, this does not mean that other units do not exist in this embodiment.

[0246] Example 3:

[0247] Another embodiment of the present application relates to an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the cryoablation temperature prediction method in the above-mentioned embodiments.

[0248] The memory and processor are connected using a bus, which can include any number of interconnected buses and bridges. The bus connects various circuits of one or more processors and memories. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits. These are all well known in the art and are therefore not described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over a wireless medium via an antenna. Furthermore, the antenna receives data and transmits it to the processor.

[0249] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory can be used to store data used by the processor when performing operations.

[0250] Example 4:

[0251] Another embodiment of the present application relates to a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the cryoablation temperature prediction method of the above embodiment.

[0252] That is, those skilled in the art will understand that all or part of the steps in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a program. The program is stored in a storage medium and includes a number of instructions for causing a device (which may be a single-chip microcomputer, chip, etc.) or a processor to execute all or part of the steps in the methods described in the various embodiments of this application. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0253] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application, and that in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present application.

Claims

1. A cryoablation temperature prediction device, characterized in that: include: A data input module is used to obtain three-dimensional imaging data before cryoablation surgery; a data processing module, configured to construct a vascular network diagram based on the vascular centerlines in the three-dimensional image data, wherein nodes in the vascular network diagram represent vascular bifurcation points, and edges in the vascular network diagram represent vascular connection relationships; A model prediction module is used to predict the temperature field of the target sampling point during the cryoablation process using a pre-trained temperature prediction model. The temperature prediction model includes a physical information neural network and a graph neural network. The physical information neural network uses the biological heat conduction equation as a constraint condition to predict the temperature field of the target sampling point. The graph neural network is used to predict the blood perfusion coefficient of each node in the vascular network diagram to determine the blood perfusion coefficient contained in the biological heat conduction equation.

2. The cryoablation temperature prediction device according to claim 1, characterized in that: Constructing a vascular network diagram based on the vascular centerlines in the three-dimensional image data includes: performing tissue segmentation on the three-dimensional image data to obtain a tissue segmentation result; extracting the blood vessel centerline from the tissue segmentation result through skeletonization processing; identifying a vessel bifurcation point in the vessel centerline; A vascular network topology is constructed based on the vascular centerline, and a vascular network graph is constructed based on the vascular network topology. Nodes in the vascular network graph represent vascular bifurcation points, and edges in the vascular network graph represent vascular connection relationships.

3. The cryoablation temperature prediction device according to claim 1, characterized in that: The temperature field of the target sampling point is predicted using a pre-trained temperature prediction model, including: predicting the blood perfusion coefficient of each node in the vascular network graph by using the graph neural network; performing interpolation processing on the blood perfusion coefficient to determine the blood perfusion coefficient of each target sampling point; The blood perfusion coefficient of each target sampling point is substituted into the biological heat conduction equation, and the equation is used as a constraint condition to predict the temperature of each target sampling point through the physical information neural network.

4. The cryoablation temperature prediction device according to claim 3, characterized in that: The temperature prediction model further includes a gated fusion network for correcting the blood perfusion coefficient of each target sampling point using the following calculation formula before substituting the blood perfusion coefficient of each target sampling point into the biological heat conduction equation; Where, represents the corrected blood perfusion coefficient of the target sampling point, represents the weight coefficient, represents the blood perfusion coefficient of the target sampling point before correction, The weight coefficient is dynamically adjusted according to the distance between the current target sampling point and the nearest blood vessel and the type of ablated tissue.

5. The cryoablation temperature prediction device according to claim 1, characterized in that: It also includes: determining target sampling points using the following strategy; For a first area around the boundary of the ablated tissue, a first sampling density is used to determine target sampling points; For a second area around the boundary of the ablated tissue, a second sampling density is used to determine target sampling points; For a third region around a blood vessel having a diameter greater than a preset value and a fourth region around a heat exchange zone of the ablation needle, the sampling density is increased to determine a target sampling point, wherein the sampling density increase in the third region is lower than the sampling density increase in the fourth region; For the remaining areas, the third sampling density is used to determine the target sampling points; The second region is located in a region of the first region far away from the ablation tissue boundary, the second sampling density is smaller than the first sampling density, and the third sampling density is smaller than the second sampling density.

6. The cryoablation temperature prediction device according to claim 1, characterized in that: Also includes: Build and train a temperature prediction model; The training temperature prediction model includes: The constructed temperature prediction model is trained unsupervised using a synthetic training dataset; A finite element model is constructed based on the three-dimensional imaging data of a real patient, and the biological heat conduction equation is solved using the finite element model to obtain a reference temperature field of the target sampling point. The temperature prediction model after unsupervised training is used based on the three-dimensional imaging data of the real patient to obtain a predicted temperature field of the target sampling point. The temperature prediction model is optimized with the minimum value of the first objective function as the goal. The first objective function is ,in, represents the mean square error between the predicted temperature field and the reference temperature field, λ represents the trade-off coefficient, represents the first objective function value, represents the loss function; By adjusting the parameters of the physical information neural network, the parameters of the graph neural network and the parameters of the biological heat conduction equation, the optimized temperature prediction model is adjusted so that the predicted temperature field obtained by the optimized temperature prediction model based on the three-dimensional imaging data of the real patient is close to the reference temperature field.

7. The cryoablation temperature prediction device according to claim 6, characterized in that: The boundary conditions of the physical information neural network include the ablation needle surface temperature, far-field temperature and initial temperature; Loss function used in unsupervised training as follows: Where, are respectively the first weight coefficient, the second weight coefficient, the third weight coefficient, the fourth weight coefficient and the fifth weight coefficient; Represents the residual loss of the biological heat conduction equation, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the biological heat conduction equation; Represents the far-field temperature loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the far-field temperature; Represents the initial condition loss, which is used to measure whether the temperature field predicted by the physical information neural network meets the initial temperature; represents the data matching loss, which is used to measure the gap between the temperature field predicted by the physical information neural network and the actual data; Represents the variational energy conservation loss, which is used to measure whether the temperature field predicted by the physical information neural network satisfies the law of conservation of energy.

8. The cryoablation temperature prediction device according to claim 1, characterized in that: Also includes: A thermophysical parameter set corresponding to different ablated tissues of different patients is constructed to solve the biological heat conduction equation when predicting the temperature field of the target sampling point. The thermophysical parameter set includes the thermophysical parameters included in the biological heat conduction equation.

9. The cryoablation temperature prediction device according to claim 1, characterized in that: Also includes: During cryoablation, collecting and measuring the temperature of the refrigerant at a first target sampling point located in a heat exchange area of ​​the ablation needle; Using a temperature prediction model, a predicted temperature of the first target sampling point is obtained under the currently set blood flow velocity and needle tip heat transfer efficiency factor; Based on the refrigerant temperature and the predicted temperature at the first target sampling point, performing an inverse calculation on the blood flow velocity and the needle tip heat transfer efficiency factor using an inverse calculation formula; The inversion calculation results are smoothed to obtain updated blood flow velocity and needle tip heat transfer efficiency factor; The biological heat conduction equation is updated according to the updated blood flow velocity and needle tip heat transfer efficiency factor to update the temperature prediction model; The inversion calculation formula is as follows: Where, represents the inverted blood flow velocity vector, represents the tip heat transfer efficiency factor after inversion, Indicates the refrigerant temperature of the first target sampling point measured, represents the predicted temperature of the first target sampling point, v represents the blood flow velocity vector, η represents the tip heat transfer efficiency factor.

10. The cryoablation temperature prediction device according to claim 1, characterized in that: Also includes: After ablation is complete, calibration images are obtained; Segmenting the calibration image to obtain a target isothermal surface contour segmentation result; Determine the prediction result of the target isothermal surface based on the temperature field of the target sampling point predicted by the temperature prediction model; The target isothermal surface contour segmentation result is registered with the target isothermal surface prediction result, and the similarity index is calculated; Calculate the difference loss function value based on the calculated similarity index; Taking the minimum difference loss function value as the optimization goal, adjusting the thermophysical parameter set, wherein the thermophysical parameter set includes thermophysical parameters included in the biological heat conduction equation; Based on the adjusted thermophysical parameter set, the bioheat conduction equation is updated; The temperature field of the target sampling point is predicted using the updated biological heat conduction equation as a constraint.

11. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform: obtaining three-dimensional imaging data before the cryoablation surgery; constructing a vascular network diagram based on the vascular centerlines in the three-dimensional imaging data, the nodes in the vascular network diagram representing vascular bifurcation points, and the edges in the vascular network diagram representing vascular connection relationships; during the cryoablation process, predicting the temperature field of the target sampling point through a pre-trained temperature prediction model, the temperature prediction model includes a physical information neural network and a graph neural network; the physical information neural network predicts the temperature field of the target sampling point using the biological heat conduction equation as a constraint condition, and the graph neural network is used to predict the blood perfusion coefficient of each node in the vascular network diagram to determine the blood perfusion coefficient contained in the biological heat conduction equation.

12. The electronic device according to claim 11, wherein: The at least one processor is further capable of performing the following operations: during the cryoablation process, collecting and measuring the refrigerant temperature at a first target sampling point located in a heat exchange zone of the ablation needle; using a temperature prediction model to obtain a predicted temperature of the first target sampling point under a currently set blood flow velocity and needle tip heat transfer efficiency factor; performing an inverse calculation on the blood flow velocity and needle tip heat transfer efficiency factor based on the refrigerant temperature and the predicted temperature at the first target sampling point; smoothing the inverse calculation results to obtain updated blood flow velocity and needle tip heat transfer efficiency factor; and updating the bioheat conduction equation based on the updated blood flow velocity and needle tip heat transfer efficiency factor to update the temperature prediction model; the inverse calculation formula is as follows: Where, represents the inverted blood flow velocity vector, represents the tip heat transfer efficiency factor after inversion, Indicates the refrigerant temperature of the first target sampling point measured, represents the predicted temperature of the first target sampling point, v represents the blood flow velocity vector, η represents the tip heat transfer efficiency factor.

13. A computer-readable storage medium storing a computer program, characterized in that: When executed by a processor, the computer program achieves the following: acquiring three-dimensional imaging data before a cryoablation procedure; constructing a vascular network diagram based on the vascular centerlines in the three-dimensional imaging data, wherein the nodes in the vascular network diagram represent vascular bifurcation points, and the edges in the vascular network diagram represent vascular connection relationships; during the cryoablation process, predicting the temperature field of a target sampling point using a pre-trained temperature prediction model, wherein the temperature prediction model includes a physical information neural network and a graph neural network; the physical information neural network predicts the temperature field of a target sampling point using a biological heat conduction equation as a constraint condition, and the graph neural network is used to predict the blood perfusion coefficient of each node in the vascular network diagram to determine the blood perfusion coefficient contained in the biological heat conduction equation.

Images