Tumor radiofrequency ablation safety boundary prediction method based on finite element simulation

Abstract

The invention relates to the technical field of radiofrequency ablation, in particular to a tumor radiofrequency ablation safety boundary prediction method based on finite element simulation. The method comprises the following steps: acquiring clinical characteristics and tumor image data of a patient, and extracting radiofrequency ablation equipment parameters; performing data processing on the clinical features and the radiofrequency ablation equipment parameters, performing image segmentation on tumor image data, inputting the segmented tumor image data into a finite element model for finite element simulation, and obtaining corresponding clinical feature enhancement tensor, three-dimensional physical field tensor and geometrical relationship tensor; correcting the clinical feature enhancement tensor, the three-dimensional physical field tensor and the geometrical relationship tensor according to the actual measurement data in the historical operation of the patient; and establishing a boundary prediction model, inputting the corrected clinical feature enhancement tensor, the corrected three-dimensional physical field tensor and the corrected geometrical relationship tensor into the boundary prediction model, and obtaining the radiofrequency ablation safety boundary of the patient. According to the invention, a boundary planning scheme highly fitting the actual condition of the patient can be provided, and the limitation of traditional empirical planning is overcome.

Description

Technical Field

[0001] This invention relates to the field of radiofrequency ablation technology, specifically to a method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation. Background Technology

[0002] In the medical field, radiofrequency ablation (RFA) is an effective minimally invasive treatment for tumors. It involves applying radiofrequency energy to tumor tissue via an electrode needle, raising the tissue temperature to a critical level for cell death (typically above 50-60°C), thereby achieving local control or radical cure of the tumor. Due to its advantages of minimal trauma, rapid recovery, and high repeatability, it is widely used in clinical practice. However, the success of RFA treatment highly depends on the safety and precision of the ablation boundary. An ideal ablation boundary needs to ensure complete ablation of the tumor tissue while maximally protecting surrounding normal tissues and adjacent vital organs, avoiding postoperative complications.

[0003] In the field of radiofrequency ablation (RFA) for tumor treatment, existing technical solutions mainly rely on the experience of physicians and manual planning of ablation boundaries. Specifically, physicians manually determine the ablation range and path based on the patient's imaging data (such as CT, MRI, etc.) and preoperative clinical assessment. This method is not only time-consuming and labor-intensive, but also difficult to accurately grasp the boundaries when facing complex tumor locations, multiple blood vessels surrounding the tumor, or proximity to dangerous organs, leading to incomplete ablation or an increased risk of complications. In addition, some advanced technologies combine finite element simulation to predict the temperature field distribution during the ablation process, thereby assisting physicians in treatment planning. However, although finite element simulation can predict the temperature field distribution, its input parameters (such as tissue physical properties) are often based on general models and lack individualized correction, resulting in deviations between simulation results and actual conditions. Furthermore, the results of finite element simulation in existing technologies cannot be directly converted into safety boundaries and still require manual adjustment by physicians. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation. This method acquires the patient's clinical characteristics and tumor imaging data, extracts the parameters of the radiofrequency ablation device, processes the clinical characteristics and radiofrequency ablation device parameters separately, segments the tumor imaging data, and inputs it into a finite element model for finite element simulation to obtain the corresponding clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor. The clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor are corrected based on the patient's historical intraoperative measurement data. A boundary prediction model is established, and the corrected clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor are input into the boundary prediction model to obtain the patient's radiofrequency ablation safety boundary. This invention can provide a boundary planning scheme that highly fits the patient's actual situation, overcoming the limitations of traditional empirical planning. This invention employs the following technical solution: a method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation, comprising: Acquire the patient's clinical characteristics and tumor imaging data, and extract the parameters of the radiofrequency ablation device; The clinical features and the parameters of the radiofrequency ablation device are processed separately, and the tumor imaging data is segmented. The processed clinical features, the parameters of the radiofrequency ablation device, and the segmented tumor image data are input into a finite element model for finite element simulation to obtain the corresponding clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor. The clinical feature enhancement tensor, three-dimensional physics tensor, and geometric relation tensor were corrected based on the patient's historical intraoperative measurement data. A boundary prediction model is established, and the modified clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor are input into the boundary prediction model to obtain the patient's radiofrequency ablation safety boundary.

[0005] Furthermore, patient clinical characteristics and tumor imaging data are obtained, and parameters of the radiofrequency ablation device are extracted, including: The clinical characteristics include: age, sex, body mass index, underlying diseases, and history of allergies; The tumor imaging data is a tumor region image, including tumor size, location, density, and information on the distribution of adjacent blood vessels; The parameters of the radiofrequency ablation device include: electrode type, electrode insertion position, output power, ablation time, and cooling mode.

[0006] Furthermore, after acquiring tumor imaging data, the process also includes: Preprocessing operations are performed on the tumor imaging data; Multiple regions of interest were delineated in the preprocessed tumor imaging data; Image features are extracted from each region of interest, and the extracted image features are converted into corresponding tissue physical property parameters.

[0007] Furthermore, the clinical features and the parameters of the radiofrequency ablation device are processed separately, and the tumor imaging data is segmented, including: The clinical features and the radiofrequency ablation device are standardized. The tumor image data was segmented using the U-Net medical image segmentation algorithm.

[0008] Furthermore, the processed clinical features and radiofrequency ablation device parameters, along with the segmented tumor imaging data, are input into a finite element model for finite element simulation, including: The tumor image data after image segmentation is input into the finite element model for unstructured mesh generation; Based on clinical characteristics, corresponding tissue physical property parameters are set for each grid cell after division to obtain a finite element mesh model; Boundary conditions and loads are set in the finite element mesh model according to the parameters of the radio frequency ablation device. Multiphysics coupling simulation was performed on the finite element mesh model to obtain the time-series three-dimensional temperature field; Based on the time-series three-dimensional temperature field, clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relation tensor are obtained.

[0009] Furthermore, the clinical feature enhancement tensor, three-dimensional physics tensor, and geometric relation tensor are corrected based on the patient's historical intraoperative measurement data, including: Obtain the root mean square error between the patient's historical intraoperative measured data and the time-series three-dimensional temperature field; Based on the root mean square error, parameter sensitivity analysis was performed to determine the highly sensitive tissue physical property parameters. An error-back adjustment strategy is used to iteratively correct the physical property parameters of highly sensitive tissues until the root mean square error is less than a preset threshold. Finite element simulation was performed using the tissue physical property parameters at this time to obtain the corrected clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor.

[0010] Furthermore, a boundary prediction model is established, specifically as follows: The boundary prediction model is established based on convolutional neural networks and deep reinforcement learning networks; The boundary prediction model includes an input layer, a feature extraction module, a DRL intervention module, and a boundary generation module; The input layer is used to receive the corrected clinical feature enhancement tensor, the three-dimensional physical field tensor, and the geometric relation tensor, and reconstruct them into a three-dimensional feature map; The feature extraction module employs a three-layer 3D depthwise separable convolutional layer to extract multimodal features from the three-dimensional feature map and output a mid-layer feature map. The DRL intervention module is used to receive the mid-level feature map output by the feature extraction module and perform weighted adjustment on the mid-level feature map based on preset security constraints. The boundary generation module includes a pooling layer, a high-level convolutional layer, and three 3D deconvolutional layers, which are used to perform dimensionality reduction, global feature extraction, and upsampling operations on the weighted and adjusted mid-level feature map in sequence, and output a three-dimensional binary mask.

[0011] Furthermore, the preset security constraints include: The ablation boundary must extend at least 5 mm beyond the tumor imaging margin; The ablation boundary must be at least 3 mm away from adjacent organs of danger; For tumor areas adjacent to large blood vessels, the ablation boundary needs to be extended outward by 10% to 20%.

[0012] The beneficial effects of this invention are as follows: First, by integrating patient clinical characteristics, imaging data, and equipment parameters, and based on finite element simulation for individualized physical field prediction, this invention provides a boundary planning scheme that closely matches the actual situation of patients, overcoming the limitations of traditional experience-based planning. Second, it innovatively introduces a CNN-DRL model to quantify clinical safety rules such as tumor edge constraints, protection of dangerous organs, and vascular thermal sink compensation, driving the model to automatically generate optimized safety boundaries that simultaneously satisfy ablation completeness and tissue protection, significantly reducing the risk of postoperative complications. The entire process is automated and standardized, significantly shortening preoperative planning time, reducing the workload of doctors, and possessing good clinical applicability. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of a tumor radiofrequency ablation safety boundary prediction method based on finite element simulation according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a boundary prediction model structure according to an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] A schematic diagram of the tumor radiofrequency ablation safety boundary prediction method based on finite element simulation according to an embodiment of the present invention is shown below. Figure 1 As shown, it includes: Acquire the patient's clinical characteristics and tumor imaging data, and extract the parameters of the radiofrequency ablation device; In this embodiment of the invention, the acquired clinical characteristics include: age, gender, body mass index, underlying diseases and allergy history. The patient's clinical records within 3 months prior to surgery are retrieved through the information system of the hospital where the patient was treated, and key fields are screened for missing information.

[0017] Tumor imaging data consists of images of the tumor region, which can be any one or more of enhanced CT images, MRI images, and ultrasound images. The images include information on tumor size, location, density, and distribution of adjacent blood vessels. At the same time, the tumor region images within one week before the patient's surgery are obtained through the image archiving and communication system of the hospital where the patient is treated to ensure that there are no motion artifacts.

[0018] The parameters of a radiofrequency ablation device include: electrode type, electrode insertion position, output power, ablation time, and cooling mode.

[0019] In this embodiment of the invention, after acquiring tumor imaging data, the tumor imaging data is further preprocessed by denoising, normalization, and slice thickness calibration to eliminate artifact interference and ensure the accuracy of feature extraction. Then, multiple regions of interest are divided in the preprocessed tumor imaging data. The division can be done manually or automatically to distinguish tumors, surrounding normal tissues, blood vessels, and dangerous organs, thereby highlighting the differences in image features of different regions. Next, image features are extracted from each divided region of interest, and the extracted image features are converted into corresponding tissue physical property parameters. When extracting image features by region, for the tumor region, the CT mean, MRI apparent diffusion coefficient, arterial phase enhancement amplitude, extracted diameter of blood vessel region, blood flow signal intensity, and distance between blood vessel and tumor are obtained; for the normal tissue region, the standard deviation of CT value and T2 signal intensity are considered. Subsequently, a correlation model is established based on physical laws or literature conclusions, and finally, the tissue physical property parameters divided by region are output.

[0020] Clinical features and radiofrequency ablation equipment parameters are processed separately, and tumor imaging data is segmented. In this embodiment of the invention, the method for processing clinical characteristics and radiofrequency ablation device parameters includes: based on the acquired age, gender, body mass index, underlying diseases, and allergy history; electrode type, electrode insertion position, output power, ablation time, and cooling mode, corresponding standardization methods are selected for standardization. For example, for numerical data such as age, body mass index, and output power, the Z-score standardization method can be used; for discrete data such as gender, underlying diseases, and electrode type, one-heat encoding is performed, and missing value imputation and outlier removal are performed simultaneously; for special data such as ablation time, in this embodiment of the invention, a regression equation can be established based on literature (e.g., ablation time = 5 × maximum tumor diameter + 2 minutes), and "outliers" deviating from the equation by ±3 minutes are corrected to ensure the reliability of the data.

[0021] The method for segmenting tumor imaging data includes: using the U-Net medical image segmentation algorithm to automatically segment the tumor imaging data. In this embodiment of the invention, the tumor imaging data is taken as a fusion feature map of CT and MRI. The weights of different modalities are dynamically allocated through an attention mechanism. For example, MRI is given a higher weight in the tumor boundary region, and CT is given a higher weight in the vascular region. The segmentation verification uses the Dice coefficient and 95% Hausdorff distance to evaluate the boundary localization accuracy, ensuring that the spatial relationship between tumor, blood vessels and dangerous organs is calculated accurately, thereby obtaining the three-dimensional boundary coordinates of the tumor region, the three-dimensional course and diameter of key blood vessels, and the three-dimensional contour of dangerous organs.

[0022] In one specific embodiment of the present invention, the acquired tissue physical property parameters are further adjusted according to the correlation between the patient's clinical characteristics and the tissue physical property parameters. For example, diabetic patients may have reduced blood perfusion rate due to tumor microcirculation disorders, and elderly patients may need to have their conductivity lowered using a correction model trained with clinical data due to decreased elasticity of normal tissue. If there is significant heterogeneity in the CT value within the tumor (e.g., some areas have a CT value of 40 HU, while others have a CT value of 60 HU), the thermal conductivity needs to be corrected separately for each area. For example, 0.55 W / (m·K) is used for areas with a CT value of 40 HU, and 0.65 W / (m·K) is used for areas with a CT value of 60 HU. For the tumor edge near large blood vessels, the blood perfusion rate correction coefficient in that area is additionally increased based on the distance between the extracted blood vessels and the tumor to reflect the local differences in vascular thermal settling. After these processes, standardized clinical feature data, segmented tumor, blood vessel, and dangerous organ 3D coordinates or mesh models and other 3D structured data are output, as well as final tissue physical property parameters that are strongly correlated with the patient's clinical characteristics and have higher accuracy. This provides more accurate data support for subsequent steps such as finite element simulation that are more in line with the individual patient situation.

[0023] The processed clinical features, the parameters of the radiofrequency ablation device, and the segmented tumor image data are input into the finite element model for finite element simulation to obtain the corresponding clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor. In this embodiment of the invention, the tumor image data after image segmentation is first input into a finite element model for unstructured mesh generation; based on clinical characteristics, corresponding tissue physical property parameters are set in each mesh unit after division to obtain a finite element mesh model; boundary conditions and loads are set in the finite element mesh model according to the parameters of the radiofrequency ablation device; multiphysics coupling simulation calculation is performed on the finite element mesh model to obtain a time-series three-dimensional temperature field; based on the time-series three-dimensional temperature field, the clinical feature enhancement tensor, the three-dimensional physical field tensor, and the geometric relationship tensor are obtained. In one specific embodiment of the present invention, the tumor image data after image segmentation is a three-dimensional structured data, which includes a three-dimensional mesh model of the tumor, blood vessels, and dangerous organs; the clinical features after data processing are Z-score standardized structured data and uniquely thermally encoded data; the parameters of the radiofrequency ablation device after data processing include standardized output power, ablation time, uniquely thermally encoded electrode type, and three-dimensional coordinates of electrode insertion depth / position; the tissue physical property parameters include a corrected sub-regional parameter matrix, such as thermal conductivity, electrical conductivity, specific heat capacity, and blood perfusion rate of tumor, normal tissue, and blood vessels.

[0024] The finite element model first converts the segmented tumor image data into a geometric model that can be recognized by the finite element analysis software. Then, an unstructured mesh generation method is used to obtain a finite element mesh model containing multiple mesh elements. In this embodiment of the invention, the mesh element size of the tumor and perivascular area is 0.5-1 mm, and the mesh element size of the normal tissue area is 2-3 mm, which is used to balance the calculation accuracy and efficiency. Then, each mesh element is assigned a corresponding tissue physical property parameter, such as the thermal conductivity assigned to the tumor area element and the blood perfusion rate assigned to the vascular element.

[0025] When setting boundary conditions and loads in the finite element mesh model according to the parameters of the radiofrequency ablation device, the embodiments of the present invention mainly target the electrode boundary, thermal boundary, and vascular boundary in the mesh. Among them, the electrode boundary sets the electrode surface as a heat source according to the device parameters. For example, when the power is 50W, the heat flux density of the electrode surface = power / electrode surface area, and an electric field excitation is defined. The thermal boundary is set for convective heat dissipation on the skin surface, while the internal parts are adiabatic by default. The vascular boundary sets convective heat transfer conditions based on blood flow velocity, corresponding to the corresponding convective coefficient.

[0026] In one specific embodiment of the present invention, the multiphysics coupling simulation calculation of the finite element mesh model includes the following: First, the electromagnetic field is solved based on Maxwell's equations. Then, the radio frequency energy distribution is calculated and converted into Joule heat. Next, heat transfer is calculated based on the biological heat conduction equation, while considering tissue metabolic heat production and blood perfusion heat dissipation. The three-dimensional temperature distribution is recorded every 10 seconds, and the time-series three-dimensional temperature field is output. Finally, the simulation duration is matched with the parameters of the radio frequency ablation device. For example, if the ablation time is 15 minutes, the simulation is extended to 900 seconds. If the device is a cold-circulating electrode, a cooling channel is added to the boundary conditions, and the refrigerant inlet temperature and flow rate are set. Cooling heat dissipation is calculated based on the thermal convection equation. During the simulation, the "cooling mode" is included as a unique thermal encoding parameter in the device parameters, and the output temperature field tensor can be used to label the temperature distribution of the "cooling region".

[0027] Finally, a series of tensor data are generated based on the output time-series three-dimensional temperature field, including clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor. The three-dimensional physical field tensor is obtained by integrating the time-series three-dimensional temperature field; the geometric relationship tensor is obtained by calculating the shortest distance matrix between the tumor and blood vessels and the spatial overlap between the tumor and dangerous organs; and the clinical feature enhancement tensor is obtained by performing sensitivity analysis on the time-series three-dimensional temperature field.

[0028] The clinical feature enhancement tensor, three-dimensional physics tensor, and geometric relation tensor were corrected based on the patient's historical intraoperative measurement data; This invention first obtains the root mean square error between the patient's historical intraoperative measured data and the time-series three-dimensional temperature field; based on the root mean square error, parameter sensitivity analysis is performed to determine the highly sensitive tissue physical property parameters; an error-backward adjustment strategy is used to iteratively correct the highly sensitive tissue physical property parameters until the root mean square error is less than a preset threshold; finite element simulation is performed using the tissue physical property parameters at this time to obtain the corrected clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor.

[0029] In one specific embodiment of the present invention, the patient's historical intraoperative measured data includes the tumor center or edge temperature recorded by intraoperative thermocouples and the actual ablation zone shown by enhanced CT scan 1 week postoperatively; the root mean square error includes temperature field error, ablation zone error, and clinical characteristic error. The temperature field error is calculated by comparing the intraoperative measured temperature with the simulated temperature in the time-series three-dimensional temperature field; the ablation zone error is obtained by calculating the relative error between the volume enclosed by the 60°C isothermal surface in the time-series three-dimensional temperature field and the actual ablation zone volume postoperatively; the clinical characteristic error is first grouped according to the patient's age and underlying diseases, such as whether the patient has diabetes or is over seventy years old; then the temperature field error and ablation zone error of each group are calculated to construct an error matrix, thus obtaining the clinical characteristic error.

[0030] After calculating the root mean square error between the patient's historical intraoperative measured data and the time-series three-dimensional temperature field, parameter sensitivity analysis is used to identify the tissue physical property parameters that have the greatest impact on the error, and high-sensitivity parameters are corrected first. If the temperature field error exceeds the standard during parameter iteration correction, the parameters are adjusted in reverse based on the error: if the measured temperature is lower than the simulated temperature, it indicates that the blood flow heat dissipation is too strong, and the vascular blood flow perfusion rate needs to be reduced by 15%-20%; if the ablation zone volume error exceeds the standard, the tumor tissue thermal conductivity is corrected; for high error groups, targeted correction strategies are adopted, such as increasing the correction range of blood flow perfusion rate from 15%-20% to 20%-25%, and then the finite element simulation is run again, and the error calculation and correction are repeated until all error indicators are lower than the preset threshold. Finally, the corrected three-dimensional physical field tensor, geometric relationship tensor, and clinical feature enhancement tensor are output.

[0031] A boundary prediction model is established, and the modified clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor are input into the boundary prediction model to obtain the patient's radiofrequency ablation safety boundary.

[0032] In this embodiment of the invention, the boundary prediction model is built based on a convolutional neural network and a deep reinforcement learning network. It includes an input layer, a feature extraction module, a DRL intervention module, and a boundary generation module. The input layer receives the corrected clinical feature enhancement tensor, the three-dimensional physical field tensor, and the geometric relation tensor, and reconstructs them into a three-dimensional feature map. The feature extraction module uses three 3D deep separable convolutional layers to extract multimodal features from the three-dimensional feature map and outputs a mid-level feature map. The DRL intervention module receives the mid-level feature map output by the feature extraction module and performs weighted adjustment on the mid-level feature map based on preset safety constraints. The boundary generation module includes a pooling layer, a high-level convolutional layer, and three 3D deconvolutional layers, which sequentially perform dimensionality reduction, global feature extraction, and upsampling operations on the weighted mid-level feature map to output a three-dimensional binary mask.

[0033] In one specific embodiment of the present invention, such as Figure 2 As shown, a schematic diagram of a boundary prediction model structure according to an embodiment of the present invention is given. The input layer is used to receive the corrected three-dimensional physical field tensor, geometric relationship tensor and clinical feature enhancement tensor, and normalize them into a 64×64×64×5 multi-channel feature map. Channel 1 is the time series temperature field; channel 2 is the electric field distribution; channel 3 is the tumor-blood vessel distance matrix; channel 4 is the tumor-dangerous organ distance matrix; and channel 5 is the clinical feature enhancement tensor.

[0034] The initial convolutional layer of the CNN is the feature extraction module. It employs three 3D depthwise separable convolutional layers to introduce structural parameters, achieving lightweight model construction. A 3×3×3 convolutional kernel is used to convolve each channel of the input feature map individually; the parameter is calculated as "3×3×3 × number of input channels". Pointwise convolution follows immediately after depthwise convolution, using a 1×1×1 kernel to perform cross-channel feature fusion on the output of the depthwise convolution; the parameter is calculated as "1×1×1 × number of input channels × number of output channels". The specific parameters for each layer are as follows: Layer 1: Input 64×64×64×5, output 64×64×64×64, depthwise convolution parameters: 3×3×3×5=135, pointwise convolution parameters: 1×1×1×5×64=320, total parameters: 135+320=455; the traditional 3D convolution parameters are 3×3×3×5×64=8640, compared to which the parameters of the first layer in this invention are reduced by about 95%; the ReLU activation function is used to extract basic physical and geometric features, reducing the data dimensionality. Layer 2: Input 64×64×64×64, output 64×64×64×128, depthwise convolution parameters: 3×3×3×64=1728, pointwise convolution parameters: 1×1×1×64×128=8192, total parameters: 1728+8192=9920. The traditional convolution parameters are 3×3×3×64×128=221184. In comparison, the parameters of the second layer in this invention are reduced by about 96%. The activation function is ReLU, which integrates multimodal features and captures local physical-geometric relationships. Layer 3: Input 64×64×64×128, output 64×64×64×256, depthwise convolution parameters: 3×3×3×128=3456, pointwise convolution parameters: 1×1×1×128×256=32768, total parameters: 3456+32768=36224. The traditional convolution parameters are 3×3×3×128×256=884736. In comparison, the parameters of the third layer in this invention are reduced by about 96%. The ReLU activation function is used to capture mesoscale spatial correlations, such as the overall distribution relationship between tumor subregions and surrounding blood vessels or dangerous organs, thereby outputting an unpooled mesoscale feature map (size 64×64×64×256) to prepare for DRL intervention.

[0035] The DRL module, or DRL intervention module, is embedded between the third convolutional layer and the first pooling layer in this embodiment of the invention to constrain the mid-layer feature map. The core is to apply set safety constraints between convolution and pooling to ensure that subsequent feature processing complies with safety rules. The safety constraints set in this embodiment of the invention include: the ablation boundary must extend at least 5 mm beyond the edge of the tumor imaging; the ablation boundary must maintain a distance of at least 3 mm from adjacent dangerous organs; and for tumor areas adjacent to large blood vessels, the ablation boundary must be extended outward by 10% to 20%.

[0036] Therefore, the state space in the DRL network is defined as follows: based on the feature values ​​of each spatial location in the feature map, safety constraint parameters are fused, and the state vector is: s = [mean of regional features (tumor or blood vessel or dangerous organ sub-region), tumor edge distance, dangerous organ distance, blood vessel diameter, safety constraint threshold]; where, the tumor edge distance refers to the shortest distance from the ablation boundary to the tumor boundary, the dangerous organ distance refers to the distance from the ablation boundary to the nearest dangerous organ, and the blood vessel diameter is used to dynamically determine the compensation coefficient.

[0037] The action space is defined as a weighted adjustment of the feature values ​​at each location in the feature map, and the action vector can be represented as: ,in, The weighting coefficients for each sub-region of the tumor are represented (e.g., 1.2 for sub-regions within 5mm of the edge and 1.0 for the interior). The weighting coefficients for each sub-region of the dangerous organ are represented (e.g., 0.3 for sub-regions less than 3 mm from the tumor and 0.8 for sub-regions greater than 3 mm). The weighting coefficients for each sub-region of the blood vessel are represented (based on diameter classification, with 1.2 for sub-regions greater than or equal to 5 mm and 1.1 for sub-regions between 3 and 5 mm).

[0038] In this embodiment of the invention, the network architecture of DRL adopts the "Actor-Critic" framework. In the Actor network, the input state s and the output action a (i.e., three weighted coefficients) are output through a fully connected layer to ensure that the action meets the constraints. The Critic network evaluates the value Q of the action by inputting the state s and the action a, and guides the Actor optimization. The specific structure can refer to any of the existing technologies.

[0039] After the DRL module performs weighted adjustments on the mid-layer feature maps Figure 2 The pooling layer, high-level feature processing, and deconvolution output together constitute the boundary generation module. Its specific implementation process is as follows: First, the pooling layer uses 3D max pooling to reduce the dimensionality of the 128×128×128×256 feature map output by the DRL, compressing the spatial resolution to 64×64×64 while retaining key features enhanced by constraints, such as tumor edge enhancement and suppression near dangerous organs. Then, two high-level convolutional layers in the high-level feature processing further extract global features, focusing on capturing the matching degree between the overall tumor morphology and the safety boundary, such as whether the boundary completely covers the tumor and avoids dangerous organs, and strengthening the consistency of constraints across regions. Finally, three 3D deconvolutional layers in the deconvolution output gradually restore the spatial resolution, decoding the 64×64×64 feature map into a 128×128×128 three-dimensional binary mask. Regions with a value of 1 represent regions within the ablation boundary that satisfy all safety constraints, while regions with a value of 0 represent dangerous regions that need to be avoided. This completes the mapping transformation from constrained features to executable safety boundaries, outputting a three-dimensional safety boundary coordinate point cloud.

[0040] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation, characterized in that, include: Acquire the patient's clinical characteristics and tumor imaging data, and extract the parameters of the radiofrequency ablation device; The clinical features and the parameters of the radiofrequency ablation device are processed separately, and the tumor imaging data is segmented. The processed clinical features, the parameters of the radiofrequency ablation device, and the segmented tumor image data are input into a finite element model for finite element simulation to obtain the corresponding clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor. The clinical feature enhancement tensor, three-dimensional physics tensor, and geometric relation tensor were corrected based on the patient's historical intraoperative measurement data. A boundary prediction model is established, and the modified clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor are input into the boundary prediction model to obtain the patient's radiofrequency ablation safety boundary.

2. The method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation according to claim 1, characterized in that: Obtain the patient's clinical characteristics and tumor imaging data, and extract the parameters of the radiofrequency ablation device, including: The clinical characteristics include: age, sex, body mass index, underlying diseases, and history of allergies; The tumor imaging data is a tumor region image, including tumor size, location, density, and information on the distribution of adjacent blood vessels; The parameters of the radiofrequency ablation device include: electrode type, electrode insertion position, output power, ablation time, and cooling mode.

3. The method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation according to claim 1, characterized in that: After acquiring tumor imaging data, the following is also included: Preprocessing operations are performed on the tumor imaging data; Multiple regions of interest were delineated in the preprocessed tumor imaging data; Image features are extracted from each region of interest, and the extracted image features are converted into corresponding tissue physical property parameters.

4. The method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation according to claim 1, characterized in that: Data processing is performed on the clinical features and the parameters of the radiofrequency ablation device, and image segmentation is performed on the tumor imaging data, including: The clinical features and the radiofrequency ablation device are standardized. The tumor image data was segmented using the U-Net medical image segmentation algorithm.

5. The method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation according to claim 3, characterized in that: The processed clinical features, radiofrequency ablation device parameters, and segmented tumor imaging data are input into a finite element model for finite element simulation, including: The tumor image data after image segmentation is input into the finite element model for unstructured mesh generation; Based on clinical characteristics, corresponding tissue physical property parameters are set for each grid cell after division to obtain a finite element mesh model; Boundary conditions and loads are set in the finite element mesh model according to the parameters of the radio frequency ablation device. Multiphysics coupling simulation was performed on the finite element mesh model to obtain the time-series three-dimensional temperature field; Based on the time-series three-dimensional temperature field, clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relation tensor are obtained.

6. The method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation according to claim 1, characterized in that: The clinical feature enhancement tensor, three-dimensional physics tensor, and geometric relation tensor are corrected based on the patient's historical intraoperative measurement data, including: Obtain the root mean square error between the patient's historical intraoperative measured data and the time-series three-dimensional temperature field; Based on the root mean square error, parameter sensitivity analysis was performed to determine the highly sensitive tissue physical property parameters. An error-back adjustment strategy is used to iteratively correct the physical property parameters of highly sensitive tissues until the root mean square error is less than a preset threshold. Finite element simulation was performed using the tissue physical property parameters at this time to obtain the corrected clinical feature enhancement tensor, three-dimensional physical field tensor, and geometric relationship tensor.

7. The method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation according to claim 1, characterized in that: Establish a boundary prediction model, specifically as follows: The boundary prediction model is established based on convolutional neural networks and deep reinforcement learning networks; The boundary prediction model includes an input layer, a feature extraction module, a DRL intervention module, and a boundary generation module; The input layer is used to receive the corrected clinical feature enhancement tensor, the three-dimensional physical field tensor, and the geometric relation tensor, and reconstruct them into a three-dimensional feature map; The feature extraction module employs a three-layer 3D depthwise separable convolutional layer to extract multimodal features from the three-dimensional feature map and output a mid-layer feature map. The DRL intervention module is used to receive the mid-level feature map output by the feature extraction module and perform weighted adjustment on the mid-level feature map based on preset security constraints. The boundary generation module includes a pooling layer, a high-level convolutional layer, and three 3D deconvolutional layers, which are used to perform dimensionality reduction, global feature extraction, and upsampling operations on the weighted and adjusted mid-level feature map in sequence, and output a three-dimensional binary mask.

8. The method for predicting the safety boundary of tumor radiofrequency ablation based on finite element simulation according to claim 7, characterized in that: The preset security constraints include: The ablation boundary must extend at least 5 mm beyond the tumor imaging margin; The ablation boundary must be at least 3 mm away from adjacent organs of danger; For tumor areas adjacent to large blood vessels, the ablation boundary needs to be extended outward by 10% to 20%.

Images