A deep learning-based microwave ablation power adaptive control method

Abstract

The application discloses a microwave ablation power adaptive control method based on deep learning, relates to the technical field of control and regulation, and comprises the following steps: updating a deep convolution network, re-estimating multi-dimensional ablation dynamic characteristic data based on the updated deep convolution network, obtaining corrected ablation state data of a current ablation cycle, acquiring ablation region temperature field data collected in real time, performing comprehensive state evaluation of microwave energy absorption and heat diffusion by fusing the corrected ablation state data and the ablation region temperature field data, and generating ablation comprehensive state evaluation data; performing microwave power adjustment decision according to the ablation comprehensive state evaluation data, generating a microwave power adaptive adjustment instruction, and performing real-time regulation and control on the transmission power of a microwave ablation system based on the microwave power adaptive adjustment instruction, so as to perform adaptive control on the microwave ablation power. The application has the effect of improving the microwave ablation power control efficiency.

Description

Technical Field

[0001] This application relates to the field of control and regulation technology, and in particular to a microwave ablation power adaptive control method based on deep learning. Background Technology

[0002] The effectiveness of microwave ablation depends directly on the tissue's absorption of microwave energy and is related to its dielectric properties. Intraoperative ultrasound monitoring relies on the tissue's acoustic properties. During ablation, the tissue is heated and dehydrated, proteins denature, and microbubbles are generated and ruptured, which simultaneously cause drastic, nonlinear, and interrelated changes in its dielectric constant and acoustic properties.

[0003] In related technologies, existing adaptive control methods for microwave ablation power often assume a stable and learnable mapping relationship between the ultrasound characteristics used to describe tissue state and the dielectric properties that determine microwave energy deposition, and establish this relationship based on preoperative or in vitro experimental data. However, in real dynamic ablation processes, this mapping relationship is time-varying and non-steady-state. For example, the high echo of tissue at a certain moment may be caused by coagulation necrosis or by instantaneous vaporization. The reflection and absorption characteristics of microwave energy are completely different for the two. Consequently, when a deep learning model trained based on historical static data encounters a new dynamic feature pattern during the operation, the "feature-state" mapping relationship learned by the model will fail, resulting in misjudgment and issuing incorrect power adjustment commands. This leads to a decrease in adaptive control of microwave ablation power, which needs improvement. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides a microwave ablation power adaptive control method based on deep learning.

[0005] In a first aspect, this application provides a microwave ablation power adaptive control method based on deep learning, comprising the following steps: Step S1: Obtain the intraoperative multimodal ultrasound image sequence corresponding to the target tissue and the real-time power parameters of the microwave ablation system. At the same time, obtain the microwave energy absorption feedback data measured by the microwave energy sensor. Based on the intraoperative multimodal ultrasound image sequence, perform multidimensional ablation dynamic feature extraction to obtain multidimensional ablation dynamic feature data. Step S2: Based on the multi-dimensional ablation dynamic feature data and the real-time power parameters, the initial ablation state is estimated through a pre-trained deep convolutional network to obtain the initial ablation state estimation data. The initial ablation state estimation data and the microwave energy absorption feedback data are input into the dynamic feature relationship correction module to perform online correction on the mapping relationship between the ultrasound features and microwave absorption state implied in the deep convolutional network, and generate the corrected network weight parameters. Step S3: Update the deep convolutional network using the corrected network weight parameters, and re-estimate the multi-dimensional ablation dynamic feature data based on the updated deep convolutional network to obtain the corrected ablation state data for the current ablation cycle. Obtain the real-time collected temperature field data of the ablation region, and fuse the corrected ablation state data with the temperature field data of the ablation region to perform a comprehensive state assessment of microwave energy absorption and thermal diffusion, generating comprehensive ablation state assessment data. Step S4: Make microwave power adjustment decisions based on the ablation comprehensive status assessment data, generate microwave power adaptive adjustment instructions, and adjust the transmission power of the microwave ablation system in real time based on the microwave power adaptive adjustment instructions to perform adaptive control of microwave ablation power.

[0006] Preferably, step S1 includes the following steps: Step S11: Obtain the intraoperative multimodal ultrasound image sequence corresponding to the target tissue, wherein the multimodal ultrasound image sequence includes at least a B-mode image sequence and an ultrasound elastography image sequence; Step S12: Perform image registration and time synchronization processing on the intraoperative multimodal ultrasound image sequence to obtain spatiotemporally registered multimodal image data; Step S13: Based on the spatiotemporally registered multimodal image data, perform multidimensional ablation dynamic feature extraction. The multidimensional ablation dynamic features include the spatiotemporal evolution features of tissue echo intensity, the dynamic change features of tissue elastic modulus, and the heterogeneity features of image texture, to obtain multidimensional ablation dynamic feature data.

[0007] Preferably, step S2 includes the following steps: Step S21: Based on the multi-dimensional ablation dynamic feature data and the real-time power parameters, perform initial ablation state estimation through a pre-trained deep convolutional network to obtain initial ablation state estimation data. The initial ablation state estimation data includes at least the microwave energy deposition distribution state and the tissue degeneration degree state. Step S22: Obtain microwave energy absorption feedback data measured by the microwave energy sensor during the current ablation cycle; Step S23: In the dynamic feature relationship correction module, calculate the state residual between the microwave energy deposition distribution state in the initial ablation state estimation data and the microwave energy absorption feedback data; Step S24: Based on the state residual, derive and adjust the weight parameters of the associated ultrasonic features and microwave absorption state in the deep convolutional network to generate the corrected network weight parameters.

[0008] Preferably, step S24 includes the following steps: Step S241: Calculate the contribution gradient of each feature dimension in the multi-dimensional ablation dynamic feature data to the state residual based on the state residual, and obtain the feature contribution gradient data; Step S242: Adaptively update the feature weight matrix of the fully connected layer in the deep convolutional network based on the feature contribution gradient data to obtain the updated feature weight matrix; Step S243: Use the updated feature weight matrix as the corrected network weight parameters.

[0009] Preferably, step S3 includes the following steps: Step S31: Update the deep convolutional network based on the corrected network weight parameters; Step S32: Input the multi-dimensional ablation dynamic feature data into the updated deep convolutional network to obtain the corrected ablation state data for the current ablation cycle; Step S33: Collect temperature field data of the ablation area in real time using a temperature sensor; Step S34: Perform spatiotemporal alignment and feature-level fusion of the corrected ablation state data and the temperature field data of the ablation region to obtain fused state feature data; Step S35: Based on the fusion state characteristic data, perform a comprehensive quantitative evaluation of microwave energy absorption efficiency and thermal diffusion range to generate ablation comprehensive state evaluation data.

[0010] Preferably, step S35 includes the following steps: Step S351: Perform joint feature reconstruction on the fusion state feature data, extract the features characterizing the core region of energy deposition and the features characterizing the edge region of thermal diffusion, and obtain joint energy-thermal diffusion feature data; Step S352: Based on the energy-thermal diffusion joint characteristic data, evaluate the effective ablation volume and potential overheating risk index under the current ablation state; Step S353: Combining the effective ablation volume and the potential overheating risk index, calculate the comprehensive state score using the state scoring function, and use the comprehensive state score as the comprehensive ablation state assessment data.

[0011] Preferably, step S352 includes the following sub-steps: The energy-thermal diffusion joint feature data is segmented into feature regions to extract feature subsets representing the energy deposition core region and feature subsets representing the thermal diffusion edge region, thereby obtaining energy deposition core feature data and thermal diffusion edge feature data. Based on the core feature data of energy deposition, the continuity analysis and volume quantization of the high absorption stable region are performed, and the total volume of the pixel region that meets the preset energy deposition threshold is calculated to obtain preliminary effective ablation volume data. Based on the thermal diffusion edge feature data, analyze the non-uniformity and abrupt change regions of the thermal diffusion gradient distribution, identify abnormal diffusion regions where the temperature exceeds the safety threshold, and calculate their volume to obtain abnormal thermal diffusion volume data. The preliminary effective ablation volume data is corrected based on the abnormal thermal diffusion volume data, and then the effective ablation volume is obtained based on the correction result. Based on the abnormal thermal diffusion volume data and the temperature gradient change rate in the thermal diffusion edge feature data, the severity and expansion trend of thermal diffusion runaway are assessed, and a potential overheating risk index is obtained through weighted fusion calculation.

[0012] Preferably, step S4 includes the following steps: Step S41: Based on the ablation comprehensive state assessment data, a decision is made through a power control model driven by a strategy gradient algorithm. The power control model aims to maximize the expected comprehensive state score for several future cycles, outputs the optimal power adjustment action, and generates an adaptive microwave power adjustment command. Step S42: Send the microwave power adaptive adjustment command to the power controller of the microwave ablation system; Step S43: The power controller adjusts the output power of the microwave transmitter in real time according to the microwave power adaptive adjustment command, thus completing the power adaptive control closed loop.

[0013] Preferably, the training method for the power control model driven by the policy gradient algorithm in step S41 includes the following steps: Step S411: In the offline stage, a state-action-reward sample pool is constructed using historical ablation surgery data. The state is the comprehensive ablation state evaluation data, the action is the power adjustment amount, and the reward is calculated based on the degree of consistency between the ablation effect and the expected target. Step S412: Initialize the policy network parameters, wherein the policy network is used to output the probability distribution of actions based on the current state; Step S413: Using the policy gradient algorithm, perform multiple rounds of iterative training with the data in the state-action-reward sample pool to update the policy network parameters, so that the policy network outputs a power adjustment policy that maximizes the cumulative expected reward.

[0014] Secondly, this application provides a microwave ablation power adaptive control system based on deep learning, comprising: The data acquisition module is used to acquire the intraoperative multimodal ultrasound image sequence corresponding to the target tissue and the real-time power parameters of the microwave ablation system. At the same time, it acquires the microwave energy absorption feedback data measured by the microwave energy sensor. Based on the intraoperative multimodal ultrasound image sequence, it performs multidimensional ablation dynamic feature extraction to obtain multidimensional ablation dynamic feature data. The calibration module is used to estimate the initial ablation state through a pre-trained deep convolutional network based on the multi-dimensional ablation dynamic feature data and the real-time power parameters, to obtain the initial ablation state estimation data. The initial ablation state estimation data and the microwave energy absorption feedback data are then input into the dynamic feature relationship correction module to perform online correction of the mapping relationship between the ultrasound features and microwave absorption state implicit in the deep convolutional network, and to generate the corrected network weight parameters. The evaluation module is used to update the deep convolutional network using the corrected network weight parameters, and to re-estimate the multi-dimensional ablation dynamic feature data based on the updated deep convolutional network to obtain the corrected ablation state data for the current ablation cycle, acquire the real-time collected temperature field data of the ablation region, and fuse the corrected ablation state data with the temperature field data of the ablation region to perform a comprehensive state evaluation of microwave energy absorption and thermal diffusion, thereby generating comprehensive ablation state evaluation data. The control module is used to make microwave power adjustment decisions based on the comprehensive ablation status assessment data, generate microwave power adaptive adjustment commands, and adjust the transmission power of the microwave ablation system in real time based on the microwave power adaptive adjustment commands to execute adaptive control of microwave ablation power.

[0015] In summary, this application includes at least one of the following beneficial technical effects: 1. This application provides a deep learning-based adaptive control method for microwave ablation power. After initial estimation through a deep convolutional network, a dynamic feature relationship correction module based on measured data from a microwave energy sensor is introduced to calculate the residual between the estimated state and the actual feedback state online and to correct the network weights in reverse. This effectively solves the model failure problem caused by the time-varying and non-steady state of feature-state mapping during the operation, enabling the system to adapt to the dynamic evolution of tissue characteristics during the ablation process in real time, and greatly reducing the risk of misjudgment caused by inaccurate mapping relationship. 2. By utilizing multimodal ultrasound sequences to extract dynamic features of tissue structure and elasticity, and simultaneously integrating microwave absorption feedback data that directly reflects energy deposition and real-time temperature field data, a more comprehensive and accurate ablation state profile is constructed through multi-level and multi-physical quantity information fusion and comprehensive evaluation, which overcomes the possible interpretations that may occur based solely on ultrasound image features, thus providing a more reliable basis for power decision-making. 3. By using a power control model driven by a strategy gradient algorithm based on refined comprehensive state assessment data, and with the goal of optimizing the expected future ablation effect, the power control model can generate more scientific and forward-looking power adjustment instructions. This not only improves ablation efficiency, but also actively identifies and suppresses overheating risk areas, realizing the upgrade of power regulation from "passive response" to "active optimization". While improving treatment efficiency, it also enhances the safety of the operation and reduces the possibility of accidental damage to surrounding healthy tissues. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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.

[0017] Figure 1 This is a flowchart of a microwave ablation power adaptive control method based on deep learning, according to an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of a microwave ablation power adaptive control system based on deep learning, according to an embodiment of this application. Detailed Implementation

[0019] The following is in conjunction with the appendix Figure 1-2 This application will be described in further detail.

[0020] Example 1 This application discloses a microwave ablation power adaptive control method based on deep learning.

[0021] Reference Figure 1 A deep learning-based adaptive control method for microwave ablation power includes the following steps: Step S1: Obtain the intraoperative multimodal ultrasound image sequence corresponding to the target tissue and the real-time power parameters of the microwave ablation system. At the same time, obtain the microwave energy absorption feedback data measured by the microwave energy sensor. Based on the intraoperative multimodal ultrasound image sequence, perform multidimensional ablation dynamic feature extraction to obtain multidimensional ablation dynamic feature data. For example, step S1 includes the following steps: Step S11: Obtain the intraoperative multimodal ultrasound image sequence corresponding to the target tissue, wherein the multimodal ultrasound image sequence includes at least a B-mode image sequence and an ultrasound elastography image sequence; Step S12: Perform image registration and time synchronization processing on the intraoperative multimodal ultrasound image sequence to obtain spatiotemporally registered multimodal image data; Step S13: Based on the spatiotemporally registered multimodal image data, perform multidimensional ablation dynamic feature extraction. The multidimensional ablation dynamic features include the spatiotemporal evolution features of tissue echo intensity, the dynamic change features of tissue elastic modulus, and the heterogeneity features of image texture, to obtain multidimensional ablation dynamic feature data.

[0022] Specifically, firstly, an intraoperative multimodal ultrasound image sequence corresponding to the target tissue is acquired using an ultrasound imaging system. This sequence includes at least a B-mode image sequence and an ultrasound elastography image sequence. The B-mode image sequence provides grayscale information about the tissue structure, while the ultrasound elastography image sequence reflects the elastic properties of the tissue through shear wave or strain imaging techniques. A uniform sampling frequency and spatial resolution are used during data acquisition to ensure sequence integrity. Next, image registration and time synchronization processing are performed on the multimodal ultrasound image sequence. Specifically, a spatial registration method based on feature point matching is used, by extracting key points from the B-mode image and the ultrasound elastography image and calculating their descriptors. After removing mismatched points using the RANSAC algorithm, the optimal spatial transformation matrix is ​​solved to achieve pixel-level alignment between multimodal images. Simultaneously, the sequence is synchronized in time based on timestamp information or a dynamic time warping algorithm. Interpolation or resampling ensures that each modal image is aligned in the time dimension, thus outputting spatiotemporally registered multimodal image data. Subsequently, multidimensional ablation dynamic feature extraction is performed based on the spatiotemporally registered multimodal image data. The spatiotemporal evolution feature of tissue echo intensity is calculated by altering the intensity value change of each pixel over time. A sliding window analysis (with the window size adaptively adjusted according to the sampling frequency) combined with a temporal gradient operator is used to quantify the intensity evolution rate, expressed as: in, It is represented as the echo intensity value of the pixel at the image spatial location (x, y) and time point t. Represented as echo intensity The partial derivative with respect to time point t is used to calculate the instantaneous rate of change of intensity. This represents the echo intensity breakdown between two adjacent sampling time points, where k represents the number corresponding to the sampling time point, and n represents the number of sampling time points within the sliding window. This is represented as an adjustment parameter, used to obtain data through fitting based on historical data. It is represented as the quantized value of the spatiotemporal evolution characteristics of the tissue echo intensity, and the larger the quantized value of the spatiotemporal evolution characteristics of the tissue echo intensity, the more significant the dynamic change of the location at the corresponding time point. The dynamic variation characteristics of tissue elastic modulus are obtained by extracting the elastic modulus value from the registered ultrasound elastography images, calculating the modulus variation curve over time using the local region averaging method, and fitting the modulus variation trend using a nonlinear regression model (such as exponential decay or polynomial fitting). The formula is expressed as: in, This represents the time since the start of the ablation process. The corresponding tissue elastic modulus value, Let denoted as the initial elastic modulus, β be the amplitude coefficient, which determines the magnitude of the modulus change curve relative to the initial elastic modulus, γ be the decay rate constant and a positive value, which controls how fast the modulus changes with time. That is, the larger the value of γ, the faster the change process tends to stabilize, and ϵ be the error term. Image texture heterogeneity features are obtained by combining gray-level co-occurrence matrix with fractal dimension analysis to calculate the contrast, entropy and uniformity of local textures in the image, and to quantify the degree of heterogeneity using the spatial variation coefficient. This results in the comprehensive output of multi-dimensional ablation dynamic feature data, providing a data foundation for subsequent ablation effect evaluation.

[0023] Step S2: Based on the multi-dimensional ablation dynamic feature data and the real-time power parameters, the initial ablation state is estimated through a pre-trained deep convolutional network to obtain the initial ablation state estimation data. The initial ablation state estimation data and the microwave energy absorption feedback data are input into the dynamic feature relationship correction module to perform online correction on the mapping relationship between the ultrasound features and microwave absorption state implied in the deep convolutional network, and generate the corrected network weight parameters. For example, step S2 includes the following steps: Step S21: Based on the multi-dimensional ablation dynamic feature data and the real-time power parameters, perform initial ablation state estimation through a pre-trained deep convolutional network to obtain initial ablation state estimation data. The initial ablation state estimation data includes at least the microwave energy deposition distribution state and the tissue degeneration degree state. Step S22: Obtain microwave energy absorption feedback data measured by the microwave energy sensor during the current ablation cycle; Step S23: In the dynamic feature relationship correction module, calculate the state residual between the microwave energy deposition distribution state in the initial ablation state estimation data and the microwave energy absorption feedback data; Step S24: Based on the state residual, derive and adjust the weight parameters of the associated ultrasonic features and microwave absorption state in the deep convolutional network to generate the corrected network weight parameters.

[0024] Specifically, firstly, the multi-dimensional ablation dynamic feature data and real-time power parameters are spatially aligned and numerically normalized to form a high-dimensional feature cube, which is then input into a pre-trained deep convolutional network. This deep convolutional network adopts an encoder-decoder structure. The encoder part extracts abstract representations of input features layer by layer through multi-layer convolution and pooling operations, capturing cross-modal correlation patterns from local texture to global structure. The decoder part reconstructs these abstract features into a two-dimensional estimation map consistent with the spatial resolution of the original image through upsampling and deconvolution operations. The network output includes two channels: the first channel is a microwave energy deposition distribution state map, where each pixel value represents the estimated microwave energy absorbed per unit volume of tissue at that location. This value is calculated by combining the spatiotemporal convolution results of local ultrasound echo evolution rate, elastic change gradient, and current injected power; the second channel is a tissue degeneration degree state map, where each pixel value represents the probability of irreversible coagulative necrosis of the tissue. This probability is generated by the Sigmoid activation function of the last layer of the network. This network operates offline. The system uses a large amount of historical surgical data for supervised training. It iterative optimization minimizes the mean square error between the network's predicted energy distribution map and the theoretical simulation distribution, while maximizing the spatial overlap between the predicted necrotic area and the gold standard area to obtain initial network weights. Next, microwave energy absorption feedback data for the current ablation cycle is acquired from sensor arrays integrated into the microwave antenna or independently placed within the tissue. This data is typically provided as temperature field rise values ​​or direct energy absorption rate distribution maps, and coordinate transformation and interpolation ensure complete spatial registration with the ultrasound image and network estimation map. Subsequently, in the dynamic feature relationship correction module, the state residual between the microwave energy deposition distribution state map in the initial estimation data and the aforementioned measured microwave energy absorption feedback data is calculated. This calculation is performed in pixels or predefined tissue blocks, taking the absolute value of the difference between the estimated and measured values, and typically dividing by the measured value or a normalization factor to obtain the relative residual, thus generating a residual distribution map. This map quantifies the degree of deviation between network prediction and physical measurement at different anatomical locations.Finally, a lightweight online parameter correction is performed on the deep convolutional network based on this state residual distribution map. The principle is to construct an additional adaptive loss function from the sum of squares of the entire residual map. The gradient of this loss with respect to the weights of the last few convolutional kernels in the network is calculated using the backpropagation algorithm. Then, a very small, conservative learning rate is used to make minor adjustments to the weights along the gradient's inverse direction. This process only updates the network parameters locally and to a limited extent, aiming to quickly reduce the systematic bias in the network's state estimation for the current specific patient, tissue environment, and instrument state without compromising its general knowledge. This generates a set of corrected network weight parameters more suitable for the current real-time surgical scenario. This corrected network will be used for state estimation in the next ablation cycle, thus forming a closed loop of "estimation-actual comparison-model fine-tuning," enabling the system to dynamically adapt to individual differences and changing conditions during surgery.

[0025] For example, step S24 includes the following steps: Step S241: Calculate the contribution gradient of each feature dimension in the multi-dimensional ablation dynamic feature data to the state residual based on the state residual, and obtain the feature contribution gradient data; Step S242: Adaptively update the feature weight matrix of the fully connected layer in the deep convolutional network based on the feature contribution gradient data to obtain the updated feature weight matrix; Step S243: Use the updated feature weight matrix as the corrected network weight parameters.

[0026] Specifically, step S24 implements a network parameter refinement correction mechanism based on feature contribution analysis. Its operating principle is to identify and strengthen or weaken the decision weights of feature channels that bear primary responsibility for the current estimation error. First, based on the state residual, the contribution gradient of each feature dimension in the multi-dimensional ablation dynamic feature data to the total residual is calculated using the error backpropagation algorithm. This process involves using the sum of squares of the entire residual map as the objective function, and applying it to each feature channel in the multi-dimensional ablation dynamic feature data input to the network (e.g., the data layer representing the echo intensity evolution rate, the data layer representing the elastic modulus, etc.). The partial derivatives of the data layers representing the gradient changes and the data layers representing texture heterogeneity indices are calculated. The resulting gradient vector is the feature contribution gradient data. The magnitude and sign of each component of this data indicate the direction and strength of the corresponding feature dimension's contribution in this estimation. Positive and large components mean that enhancing the feature will increase the estimation error, while negative components mean that the feature's contribution may be insufficient. Next, based on the calculated feature contribution gradient data, the feature weight matrix of the fully connected layer in the deep convolutional network, which is responsible for fusing high-level features and mapping them to state estimation, is adaptively updated using a momentum-driven update strategy. The gradient descent method, specifically, uses the feature contribution gradient data as an adjustment signal, multiplies it by a preset, extremely small adaptive learning rate, and adds the momentum term from the previous update step. Together, these factors are used to calculate the adjustment amount for each element in the current step's weight matrix. This adjustment amount is then subtracted from the original weights to obtain the updated feature weight matrix. Essentially, this process dynamically redistributes the "voice" of different ultrasound features in the final state estimation decision based on real-time feedback, reducing the weights of features that cause the current bias and increasing the weights of features that help correct the bias. However, the entire update is limited to a very small amplitude. To prevent the network from forgetting the basic knowledge it gained during pre-training, the updated feature weight matrix from this iteration, along with other convolutional layer weight parameters that were not adjusted in this step, are packaged together as a set of corrected network weight parameters temporarily optimized for the current surgical environment. This set of parameters will be immediately loaded into the network to process the data at the next time step or the next ablation site, thus completing a closed-loop adaptive learning cycle from error assessment to targeted parameter correction. This allows the system to continuously and dynamically adjust its internal decision-making logic to better match the tissue characteristics of individual patients and changes in real-time surgical conditions.

[0027] Step S3: Update the deep convolutional network using the corrected network weight parameters, and re-estimate the multi-dimensional ablation dynamic feature data based on the updated deep convolutional network to obtain the corrected ablation state data for the current ablation cycle. Obtain the real-time collected temperature field data of the ablation region, and fuse the corrected ablation state data with the temperature field data of the ablation region to perform a comprehensive state assessment of microwave energy absorption and thermal diffusion, generating comprehensive ablation state assessment data. For example, step S3 includes the following steps: Step S31: Update the deep convolutional network based on the corrected network weight parameters; Step S32: Input the multi-dimensional ablation dynamic feature data into the updated deep convolutional network to obtain the corrected ablation state data for the current ablation cycle; Step S33: Collect temperature field data of the ablation area in real time using a temperature sensor; Step S34: Perform spatiotemporal alignment and feature-level fusion of the corrected ablation state data and the temperature field data of the ablation region to obtain fused state feature data; Step S35: Based on the fusion state characteristic data, perform a comprehensive quantitative evaluation of microwave energy absorption efficiency and thermal diffusion range to generate ablation comprehensive state evaluation data.

[0028] Specifically, firstly, the corrected network weight parameters are loaded into the deep convolutional network, replacing the original weight matrices of the corresponding layers, thus completing the online update of the network. This update process ensures that the feature response mechanism within the network has been adaptively adjusted to the specific patient tissue characteristics and surgical environment. Subsequently, the multi-dimensional ablation dynamic feature data (i.e., the spatiotemporal evolution of tissue echo intensity, dynamic changes in elastic modulus, and image texture heterogeneity features) acquired in real time during the current ablation cycle are input into the updated deep convolutional network. The network performs forward propagation calculations, reweighting and nonlinearly transforming the input features using the new weights, and finally outputs the corrected ablation state data for the current ablation cycle. This data also contains high... The system generates a high-resolution microwave energy deposition distribution map and a probability map of tissue degeneration, but the estimation results are more closely aligned with current physical reality due to weight correction. Simultaneously, a multi-point temperature sensor array deployed around the ablation needle tip or within the tissue collects real-time temperature field data of the ablation area at millisecond-level sampling frequencies. This data is a spatially discrete set of temperature points, which is reconstructed into a continuous two-dimensional temperature distribution map using thin-plate spline interpolation or Kriging spatial interpolation methods. This data is synchronized with the ultrasound image sequence in the temporal dimension, achieving spatiotemporal alignment between the temperature field data and the ultrasound image data. Next, the corrected ablation state data and the spatiotemporally aligned ablation area temperature field data are fused at the feature level, and the fusion is performed within a shared context. The process is conducted in a spatial coordinate system. For each spatial location, the network-estimated energy deposition value, tissue denaturation probability value, measured temperature value, and temperature change rate over time are combined to form a multimodal feature vector. A lightweight feature cross-network is used to calculate the interaction weights between different feature channels, and these are then weighted, concatenated, and dimensionality-reduced to generate a fused state feature data that simultaneously reflects energy absorption, biological tissue response, and physical thermal diffusion. Finally, a comprehensive quantitative evaluation of microwave energy absorption efficiency and thermal diffusion range is performed based on the fused state feature data. The microwave energy absorption efficiency is evaluated by calculating the ratio of the effective thermal energy reflected by the sensor's measured temperature rise to the total deposition energy estimated by the network. Within the region of interest, the estimated energy corresponding to pixels with significant temperature increases is summed and then divided by the total energy deposition estimated by the network in that region to obtain an efficiency coefficient between 0 and 1. The thermal diffusion range is quantified by analyzing the volume or area enclosed by isothermal surfaces exceeding a preset biological tissue denaturation threshold in the temperature field data. Simultaneously, the high-probability regions of the tissue denaturation probability map in the fusion features are combined to calculate the spatial overlap and difference between the two, thereby comprehensively judging the adequacy and safety of ablation. Finally, a comprehensive ablation status assessment data is generated, which includes multiple quantitative indicators such as absorption efficiency coefficient, effective thermal diffusion volume, overheating risk area indication, and the matching degree between predicted necrotic area and hot zone, for real-time guidance of clinical decision-making.

[0029] For example, step S35 includes the following steps: Step S351: Perform joint feature reconstruction on the fusion state feature data, extract the features characterizing the core region of energy deposition and the features characterizing the edge region of thermal diffusion, and obtain joint energy-thermal diffusion feature data; Step S352: Based on the energy-thermal diffusion joint characteristic data, evaluate the effective ablation volume and potential overheating risk index under the current ablation state; Step S353: Combining the effective ablation volume and the potential overheating risk index, calculate the comprehensive state score using the state scoring function, and use the comprehensive state score as the comprehensive ablation state assessment data.

[0030] Specifically, joint feature reconstruction is performed on the fused state feature data. This process utilizes an autoencoder network to reduce the dimensionality and remove redundancy from the high-dimensional fused features. Simultaneously, two types of key region features are separated from the dimensionality-reduced feature map using a spatial gradient operator and a region growing algorithm: one type represents the core region of energy deposition, defined by selecting spatially continuous regions in the fused features where the estimated microwave energy deposition value is extremely high, accompanied by a tissue denaturation probability value higher than a preset deterministic threshold and a measured temperature exceeding the tissue coagulation threshold. The mean and spatial distribution density of features within this region are extracted. The other type represents the features of the thermal diffusion edge region, identified by using isotherms in the measured temperature field between the effective ablation temperature and the safe temperature threshold. The gradient difference between the estimated energy value and the measured temperature value of the network within the strip region, as well as the coefficient of variation of the texture heterogeneity features, are extracted to define the energy-thermal diffusion joint feature data that simultaneously includes the core region's strength and stability, and the edge region's transition and inhomogeneity. Subsequently, quantitative evaluations are performed based on this joint feature data. The effective ablation volume is obtained by reconstructing the spatial point set of the energy deposition core region in three dimensions and calculating its containment volume. The calculation considers not only the number of voxels but also the average probability of tissue degeneration in the fusion features within the region to weight the volume, distinguishing between complete necrosis and partial damage, thus obtaining a weighted effective ablation volume that better reflects the pathological reality. The potential overheating risk index is calculated by comprehensively considering three sub-indicators. The calculation yields three sub-indicators: the first is the ratio of the volume of the heat diffusion edge region to the effective ablation volume, reflecting the relative degree of heat diffusion; the second is the maximum measured temperature gradient within the edge region, reflecting the presence of abrupt temperature changes and potential hotspots; and the third is the degree of disorder in the changes of ultrasonic elastic characteristics within the edge region, quantified by calculating the spatial standard deviation of the rate of change of elastic modulus. These three sub-indicators are then normalized, multiplied by different risk weights, and summed to obtain a potential overheating risk index ranging from 0 to 1. A higher index indicates a greater risk of thermal damage to surrounding normal tissues. Finally, the effective ablation volume and the potential overheating risk index are combined, and a comprehensive assessment is performed using a pre-defined state scoring function. The status score, designed to encourage maximizing effective ablation volume while penalizing excessive overheating risk, is typically constructed by dividing the effective ablation volume by a preset target volume to obtain a completion ratio. Simultaneously, a potential overheating risk index is mapped to a risk penalty coefficient between 0 and 1. The overall status score is the product of the completion ratio and the risk penalty coefficient, multiplied by a standard full score coefficient. This score directly quantifies the degree to which the current ablation status approaches the ideal target; a higher score indicates that the ablation is both sufficient and safe. This overall status score, along with its derived weighted effective ablation volume and potential overheating risk index, constitutes the final comprehensive ablation status assessment data, providing clinicians with immediate and quantitative decision support.

[0031] For example, step S352 includes the following sub-steps: The energy-thermal diffusion joint feature data is segmented into feature regions to extract feature subsets representing the energy deposition core region and feature subsets representing the thermal diffusion edge region, thereby obtaining energy deposition core feature data and thermal diffusion edge feature data. Based on the core feature data of energy deposition, the continuity analysis and volume quantization of the high absorption stable region are performed, and the total volume of the pixel region that meets the preset energy deposition threshold is calculated to obtain preliminary effective ablation volume data. Based on the thermal diffusion edge feature data, analyze the non-uniformity and abrupt change regions of the thermal diffusion gradient distribution, identify abnormal diffusion regions where the temperature exceeds the safety threshold, and calculate their volume to obtain abnormal thermal diffusion volume data. The preliminary effective ablation volume data is corrected based on the abnormal thermal diffusion volume data, and then the effective ablation volume is obtained according to the correction result. Specifically, the volume of the part that overlaps with the abnormal thermal diffusion region in space is subtracted from the preliminary effective ablation volume data to obtain the final effective ablation volume. Based on the abnormal thermal diffusion volume data and the temperature gradient change rate in the thermal diffusion edge feature data, the severity and expansion trend of thermal diffusion runaway are assessed, and a potential overheating risk index is obtained through weighted fusion calculation.

[0032] Specifically, in step S352, the comprehensive evaluation of the joint energy-thermal diffusion feature data begins with refined region segmentation of the feature data. This segmentation is achieved through an unsupervised clustering algorithm combined with spatial continuity constraints. Feature vectors with high energy deposition values, high tissue denaturation probability, and stable temperature are classified into the core energy deposition feature subset, while feature vectors with moderate energy deposition values, significant temperature gradients, and drastic elasticity changes are classified into the thermal diffusion edge feature subset, thus separating the core energy deposition feature data and the thermal diffusion edge feature data. When calculating the preliminary effective ablation volume based on the core feature data, firstly, all spatially continuous regions that meet the preset energy deposition threshold are extracted through connected component analysis. Then, the number of pixels in the region is converted into physical volume according to the spatial resolution of the medical image, and weighted by multiplying by a confidence coefficient determined by the average tissue denaturation probability of the region, thereby obtaining the preliminary effective ablation volume data. For the thermal diffusion edge feature data, the second-order gradient of its spatial temperature distribution is calculated to identify steeply changing regions where the gradient change rate exceeds the safety threshold, and combined with the measured temperature exceeding the preset safety upper limit. The process involves locating the abnormal heat diffusion area and calculating its volume using the same pixel-to-volume conversion method. Volume correction is then performed by comparing spatial locations to identify the overlap between the abnormal diffusion area and the initial effective ablation area. Since this overlap area, although reaching the ablation temperature, may pose a risk of overheating leading to carbonization or vapor explosion, its therapeutic value is considered negative. Therefore, the volume of this overlapping portion is precisely subtracted from the initial effective ablation volume to obtain a corrected, safer final effective ablation volume. Finally, the calculation of the potential overheating risk index integrates the relative size of the abnormal heat diffusion volume and the runaway trend of heat diffusion. Specifically, the ratio of the abnormal heat diffusion volume to the total heat diffusion edge area volume is used as a basic risk indicator. Simultaneously, the spatial maximum value of the temperature gradient change rate is extracted from the heat diffusion edge feature data as a risk aggravation indicator. These two indicators are normalized, and the basic risk indicator is assigned a higher weight, followed by the risk aggravation indicator, before being weighted and summed to generate a quantified potential overheating risk index. This index comprehensively reflects the severity and dynamic trend of abnormal heat diffusion to non-target areas during the current ablation process.

[0033] Step S4: Make microwave power adjustment decisions based on the ablation comprehensive status assessment data, generate microwave power adaptive adjustment instructions, and adjust the transmission power of the microwave ablation system in real time based on the microwave power adaptive adjustment instructions to perform adaptive control of microwave ablation power.

[0034] For example, step S4 includes the following steps: Step S41: Based on the ablation comprehensive state assessment data, a decision is made through a power control model driven by a strategy gradient algorithm. The power control model aims to maximize the expected comprehensive state score for several future cycles, outputs the optimal power adjustment action, and generates an adaptive microwave power adjustment command. Step S42: Send the microwave power adaptive adjustment command to the power controller of the microwave ablation system; Step S43: The power controller adjusts the output power of the microwave transmitter in real time according to the microwave power adaptive adjustment command, thus completing the power adaptive control closed loop.

[0035] Specifically, the ablation comprehensive state assessment data (including comprehensive state score, effective ablation volume, and potential overheating risk index) is input as the current environmental state into a pre-trained power control model driven by a policy gradient algorithm. The core of this model is a policy function composed of a deep neural network. Its internal calculation process is as follows: After receiving the current state, the model performs feature abstraction and value evaluation through a multi-layer fully connected network. The network output layer represents the probability distribution of all selectable power adjustment actions. Simultaneously, an auxiliary value network evaluates the cumulative discount value of the expected comprehensive state score obtainable over several ablation cycles after performing various actions in the current state. During decision-making, a specific action is sampled based on the probability distribution output by the policy network. However, for optimization, the policy gradient algorithm updates the network parameters through interaction with the environment. The principle is: during the training phase, the model tries different power adjustment actions and observes the subsequent actual comprehensive state scores. The action is judged by calculating the difference in the advantage function between the actual score after the action is executed and the score predicted by the value network. The model assesses the strengths and weaknesses of different power adjustment actions and then adjusts the weight parameters of the policy network through backpropagation along the direction that increases the advantage function value. This makes the model more inclined to select power adjustment actions that bring higher future expected scores. In this real-time control, the model uses the pre-trained policy network to directly output the power adjustment action with the highest probability based on the latest input state. This action is encoded as a specific microwave power adaptive adjustment command. Immediately afterward, this command is sent to the embedded power controller of the microwave ablation system via a standard communication protocol. Finally, the power controller parses the received command and generates corresponding pulse width modulation signals or direct digital control signals through its internal digital signal processor or field-programmable gate array. It then adjusts the driving voltage or current of the microwave transmitter in real time, thereby accurately and quickly adjusting its output power to the target value set by the command. This completes a power adaptive control closed loop from "state assessment" to "intelligent decision-making" to "execution feedback," enabling the system to dynamically optimize the energy delivery strategy and proactively manage thermal risks while ensuring ablation effectiveness.

[0036] For example, the training method of the power control model driven by the policy gradient algorithm in step S41 includes the following steps: Step S411: In the offline stage, a state-action-reward sample pool is constructed using historical ablation surgery data. The state is the comprehensive ablation state evaluation data, the action is the power adjustment amount, and the reward is calculated based on the degree of consistency between the ablation effect and the expected target. Step S412: Initialize the policy network parameters, wherein the policy network is used to output the probability distribution of actions based on the current state; Step S413: Using the policy gradient algorithm, perform multiple rounds of iterative training with the data in the state-action-reward sample pool to update the policy network parameters, so that the policy network outputs a power adjustment policy that maximizes the cumulative expected reward.

[0037] Specifically, the training method for the power control model driven by the policy gradient algorithm in step S41 first constructs a state-action-reward sample pool using historical ablation surgery data in the offline stage. Specifically, this involves replaying the surgical records stored in the database, where each record contains multiple consecutive time slices. The comprehensive ablation status assessment data (including comprehensive status score, effective ablation volume, and potential overheating risk index) corresponding to each slice is used as the state representation. The actual power adjustment amount performed in that slice is used as the action label, and an immediate reward value is calculated for this state-action pair. This reward value is calculated based on a preset reward function, the core of which is to encourage… The ablation effect approaches the expected target, and its typical design is as follows: a positive reward is given when the effective ablation volume increases and the potential overheating risk index does not exceed the safety threshold, with the reward value positively correlated with the volume increase; zero or slight negative rewards are given when the overheating risk index exceeds the threshold or the effective ablation volume does not increase significantly; and a large negative reward is given when the effective ablation volume decreases or the risk increases sharply. This method labels each decision moment in the massive historical data with a quantitative "good" or "bad" evaluation, thereby constructing a sample pool for training. Next, the parameters of the policy network are initialized. This policy network is typically a multi-layer fully connected neural network, with its input layer dimension and state representation dimension being equal. The output layer dimension is consistent with the number of selectable actions. The output value is converted into a probability distribution for selecting each action using the Softmax function. The network parameters are initialized using Xavier or He initialization methods to ensure stable gradient flow in the early stages of training. Finally, the policy gradient algorithm is used to perform multiple rounds of iterative training with the sample pool to update the policy network parameters. The core principle is: in each round of training, a batch of state-action-reward sequences are sampled from the sample pool. First, the "advantage function" of each action is calculated, which is the cumulative reward actually brought by the action minus the average expected reward of all actions in that state, to measure the relative merit of the action. Then, the policy gradient algorithm is applied. The goal of the algorithm is to increase the output probability of actions with positive dominance function values ​​while decreasing the probability of actions with negative dominance function values. Specifically, this is achieved by calculating the gradient of the logarithm of the policy network's output probability with respect to the network parameters, and multiplying this gradient by the corresponding dominance function value to obtain the direction and magnitude of the parameter update. Through numerous iterations, the parameters within the policy network are gradually adjusted, eventually learning a mapping relationship: for a specific ablation state, it can output an action probability distribution, such that selecting power-adjusted actions according to this distribution yields the maximum cumulative expected reward in the long run, which means it can most effectively guide the ablation process to reach a target state that is both sufficient and safe.

[0038] Example 2 This application also discloses a microwave ablation power adaptive control system based on deep learning.

[0039] Reference Figure 2 A deep learning-based microwave ablation power adaptive control system includes: The data acquisition module is used to acquire the intraoperative multimodal ultrasound image sequence corresponding to the target tissue and the real-time power parameters of the microwave ablation system. At the same time, it acquires the microwave energy absorption feedback data measured by the microwave energy sensor. Based on the intraoperative multimodal ultrasound image sequence, it performs multidimensional ablation dynamic feature extraction to obtain multidimensional ablation dynamic feature data. The calibration module is used to estimate the initial ablation state through a pre-trained deep convolutional network based on the multi-dimensional ablation dynamic feature data and the real-time power parameters, to obtain the initial ablation state estimation data. The initial ablation state estimation data and the microwave energy absorption feedback data are then input into the dynamic feature relationship correction module to perform online correction of the mapping relationship between the ultrasound features and microwave absorption state implicit in the deep convolutional network, and to generate the corrected network weight parameters. The evaluation module is used to update the deep convolutional network using the corrected network weight parameters, and to re-estimate the multi-dimensional ablation dynamic feature data based on the updated deep convolutional network to obtain the corrected ablation state data for the current ablation cycle, acquire the real-time collected temperature field data of the ablation region, and fuse the corrected ablation state data with the temperature field data of the ablation region to perform a comprehensive state evaluation of microwave energy absorption and thermal diffusion, thereby generating comprehensive ablation state evaluation data. The control module is used to make microwave power adjustment decisions based on the comprehensive ablation status assessment data, generate microwave power adaptive adjustment commands, and adjust the transmission power of the microwave ablation system in real time based on the microwave power adaptive adjustment commands to execute adaptive control of microwave ablation power.

[0040] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention, they should all fall within the protection scope of the present invention.

[0041] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0042] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A microwave ablation power adaptive control method based on deep learning, characterized in that, Includes the following steps: Step S1: Obtain the intraoperative multimodal ultrasound image sequence corresponding to the target tissue and the real-time power parameters of the microwave ablation system. At the same time, obtain the microwave energy absorption feedback data measured by the microwave energy sensor. Based on the intraoperative multimodal ultrasound image sequence, perform multidimensional ablation dynamic feature extraction to obtain multidimensional ablation dynamic feature data. Step S2: Based on the multi-dimensional ablation dynamic feature data and the real-time power parameters, the initial ablation state is estimated through a pre-trained deep convolutional network to obtain the initial ablation state estimation data. The initial ablation state estimation data and the microwave energy absorption feedback data are input into the dynamic feature relationship correction module to perform online correction on the mapping relationship between the ultrasound features and microwave absorption state implied in the deep convolutional network, and generate the corrected network weight parameters. Step S3: Update the deep convolutional network using the corrected network weight parameters, and re-estimate the multi-dimensional ablation dynamic feature data based on the updated deep convolutional network to obtain the corrected ablation state data for the current ablation cycle. Obtain the real-time collected temperature field data of the ablation region, and fuse the corrected ablation state data with the temperature field data of the ablation region to perform a comprehensive state assessment of microwave energy absorption and thermal diffusion, generating comprehensive ablation state assessment data. Step S4: Make microwave power adjustment decisions based on the ablation comprehensive status assessment data, generate microwave power adaptive adjustment instructions, and adjust the transmission power of the microwave ablation system in real time based on the microwave power adaptive adjustment instructions to perform adaptive control of microwave ablation power.

2. The microwave ablation power adaptive control method based on deep learning according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Obtain the intraoperative multimodal ultrasound image sequence corresponding to the target tissue, wherein the multimodal ultrasound image sequence includes at least a B-mode image sequence and an ultrasound elastography image sequence; Step S12: Perform image registration and time synchronization processing on the intraoperative multimodal ultrasound image sequence to obtain spatiotemporally registered multimodal image data; Step S13: Based on the spatiotemporally registered multimodal image data, perform multidimensional ablation dynamic feature extraction. The multidimensional ablation dynamic features include the spatiotemporal evolution features of tissue echo intensity, the dynamic change features of tissue elastic modulus, and the heterogeneity features of image texture, to obtain multidimensional ablation dynamic feature data.

3. The microwave ablation power adaptive control method based on deep learning according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Based on the multi-dimensional ablation dynamic feature data and the real-time power parameters, perform initial ablation state estimation through a pre-trained deep convolutional network to obtain initial ablation state estimation data. The initial ablation state estimation data includes at least the microwave energy deposition distribution state and the tissue degeneration degree state. Step S22: Obtain microwave energy absorption feedback data measured by the microwave energy sensor during the current ablation cycle; Step S23: In the dynamic feature relationship correction module, calculate the state residual between the microwave energy deposition distribution state in the initial ablation state estimation data and the microwave energy absorption feedback data; Step S24: Based on the state residual, derive and adjust the weight parameters of the associated ultrasonic features and microwave absorption state in the deep convolutional network to generate the corrected network weight parameters.

4. The microwave ablation power adaptive control method based on deep learning according to claim 3, characterized in that, Step S24 includes the following steps: Step S241: Calculate the contribution gradient of each feature dimension in the multi-dimensional ablation dynamic feature data to the state residual based on the state residual, and obtain the feature contribution gradient data; Step S242: Adaptively update the feature weight matrix of the fully connected layer in the deep convolutional network based on the feature contribution gradient data to obtain the updated feature weight matrix; Step S243: Use the updated feature weight matrix as the corrected network weight parameters.

5. The microwave ablation power adaptive control method based on deep learning according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Update the deep convolutional network based on the corrected network weight parameters; Step S32: Input the multi-dimensional ablation dynamic feature data into the updated deep convolutional network to obtain the corrected ablation state data for the current ablation cycle; Step S33: Collect temperature field data of the ablation area in real time using a temperature sensor; Step S34: Perform spatiotemporal alignment and feature-level fusion of the corrected ablation state data and the temperature field data of the ablation region to obtain fused state feature data; Step S35: Based on the fusion state characteristic data, perform a comprehensive quantitative evaluation of microwave energy absorption efficiency and thermal diffusion range to generate ablation comprehensive state evaluation data.

6. The microwave ablation power adaptive control method based on deep learning according to claim 5, characterized in that, Step S35 includes the following steps: Step S351: Perform joint feature reconstruction on the fusion state feature data, extract the features characterizing the core region of energy deposition and the features characterizing the edge region of thermal diffusion, and obtain joint energy-thermal diffusion feature data; Step S352: Based on the energy-thermal diffusion joint characteristic data, evaluate the effective ablation volume and potential overheating risk index under the current ablation state; Step S353: Combining the effective ablation volume and the potential overheating risk index, calculate the comprehensive state score using the state scoring function, and use the comprehensive state score as the comprehensive ablation state assessment data.

7. The microwave ablation power adaptive control method based on deep learning according to claim 6, characterized in that, Step S352 includes the following sub-steps: The energy-thermal diffusion joint feature data is segmented into feature regions to extract feature subsets representing the energy deposition core region and feature subsets representing the thermal diffusion edge region, thereby obtaining energy deposition core feature data and thermal diffusion edge feature data. Based on the core feature data of energy deposition, the continuity analysis and volume quantization of the high absorption stable region are performed, and the total volume of the pixel region that meets the preset energy deposition threshold is calculated to obtain preliminary effective ablation volume data. Based on the thermal diffusion edge feature data, analyze the non-uniformity and abrupt change regions of the thermal diffusion gradient distribution, identify abnormal diffusion regions where the temperature exceeds the safety threshold, and calculate their volume to obtain abnormal thermal diffusion volume data. The preliminary effective ablation volume data is corrected based on the abnormal thermal diffusion volume data, and then the effective ablation volume is obtained based on the correction result. Based on the abnormal thermal diffusion volume data and the temperature gradient change rate in the thermal diffusion edge feature data, the severity and expansion trend of thermal diffusion runaway are assessed, and a potential overheating risk index is obtained through weighted fusion calculation.

8. The microwave ablation power adaptive control method based on deep learning according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Based on the ablation comprehensive state assessment data, a decision is made through a power control model driven by a strategy gradient algorithm. The power control model aims to maximize the expected comprehensive state score for several future cycles, outputs the optimal power adjustment action, and generates an adaptive microwave power adjustment command. Step S42: Send the microwave power adaptive adjustment command to the power controller of the microwave ablation system; Step S43: The power controller adjusts the output power of the microwave transmitter in real time according to the microwave power adaptive adjustment command, thus completing the power adaptive control closed loop.

9. The microwave ablation power adaptive control method based on deep learning according to claim 7, characterized in that, The training method for the power control model driven by the policy gradient algorithm in step S41 includes the following steps: Step S411: In the offline stage, a state-action-reward sample pool is constructed using historical ablation surgery data. The state is the comprehensive ablation state evaluation data, the action is the power adjustment amount, and the reward is calculated based on the degree of consistency between the ablation effect and the expected target. Step S412: Initialize the policy network parameters, wherein the policy network is used to output the probability distribution of actions based on the current state; Step S413: Using the policy gradient algorithm, perform multiple rounds of iterative training with the data in the state-action-reward sample pool to update the policy network parameters, so that the policy network outputs a power adjustment policy that maximizes the cumulative expected reward.

10. A deep learning-based adaptive control system for microwave ablation power, applied to the deep learning-based adaptive control method for microwave ablation power described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire the intraoperative multimodal ultrasound image sequence corresponding to the target tissue and the real-time power parameters of the microwave ablation system. At the same time, it acquires the microwave energy absorption feedback data measured by the microwave energy sensor. Based on the intraoperative multimodal ultrasound image sequence, it performs multidimensional ablation dynamic feature extraction to obtain multidimensional ablation dynamic feature data. The calibration module is used to estimate the initial ablation state through a pre-trained deep convolutional network based on the multi-dimensional ablation dynamic feature data and the real-time power parameters, to obtain the initial ablation state estimation data. The initial ablation state estimation data and the microwave energy absorption feedback data are then input into the dynamic feature relationship correction module to perform online correction of the mapping relationship between the ultrasound features and microwave absorption state implicit in the deep convolutional network, and to generate the corrected network weight parameters. The evaluation module is used to update the deep convolutional network using the corrected network weight parameters, and to re-estimate the multi-dimensional ablation dynamic feature data based on the updated deep convolutional network to obtain the corrected ablation state data for the current ablation cycle, acquire the real-time collected temperature field data of the ablation region, and fuse the corrected ablation state data with the temperature field data of the ablation region to perform a comprehensive state evaluation of microwave energy absorption and thermal diffusion, thereby generating comprehensive ablation state evaluation data. The control module is used to make microwave power adjustment decisions based on the comprehensive ablation status assessment data, generate microwave power adaptive adjustment commands, and adjust the transmission power of the microwave ablation system in real time based on the microwave power adaptive adjustment commands to execute adaptive control of microwave ablation power.

Images