An AI-based radio frequency detection treatment system and method

Abstract

The application relates to the technical field of radio frequency treatment and medical monitoring, and discloses an AI-based radio frequency detection and treatment system and method. The method comprises the following steps: establishing a time-sharing direct current measurement channel during a radio frequency output intermittent period, applying a step direct current excitation, and synchronously collecting tissue voltage response and deformation images. According to image analysis of structural displacement, dynamic impedance mapping relationship is established in combination with voltage signals. The relationship is input into a pre-trained deep learning network, the network generates an energy deposition distribution map and a tissue degeneration state distribution map in the treatment area through the interactive calculation of the tissue state evolution reasoning branch and the heat diffusion simulation branch. Based on the former, the initial parameters of the next cycle radio frequency energy are calculated, based on the latter, the sub-regions of under-treatment are identified, and the composite radio frequency control instruction containing the time-space energy modulation scheme is comprehensively generated. The application realizes accurate monitoring and adaptive intelligent regulation and control of the treatment process.

Description

An AI-based radiofrequency detection and treatment system and method Technical Field

[0001] This invention relates to the field of radiofrequency therapy and medical monitoring technology, specifically to an AI-based radiofrequency detection and treatment system and method. Background Technology

[0002] Radiofrequency ablation is an important minimally invasive treatment for diseases such as tumor ablation. Its efficacy depends on precise control of energy deposition and tissue denaturation processes within the treatment area. Current technologies mainly assess the treatment progress indirectly by monitoring changes in radiofrequency impedance between treatment electrodes or by using equipment such as ultrasound and magnetic resonance imaging to observe tissue morphology and temperature, and adjust the output energy accordingly.

[0003] These conventional methods have significant limitations. On the one hand, during the main treatment period with continuous radiofrequency energy output, the strong treatment signal itself can mask subtle early changes in the electrical properties of the tissue, resulting in low signal-to-noise ratio and insufficient specificity in the monitoring data. Simple impedance monitoring can only reflect the overall trend and cannot actively detect the evolution of the tissue state before significant thermal damage occurs during treatment intervals. On the other hand, existing regulation models usually treat physical energy deposition and biological tissue denaturation processes separately, or use simplified empirical formulas to correlate them. This approach is difficult to accurately describe the nonlinear relationship of dynamic coupling and mutual influence between thermal diffusion and complex biochemical reactions such as protein denaturation and cell necrosis, resulting in limited accuracy of the generated regulation maps and difficulty in reliably identifying undertreated areas caused by tissue heterogeneity or blood flow heat dissipation.

[0004] Current technology struggles to achieve synchronized, high-precision visualization and prediction of the energy distribution and actual denaturation state within the tissue under visual field during treatment, hindering the development of radiofrequency ablation towards individualized, adaptive, and precise control. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based radiofrequency detection and treatment system and method to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an AI-based radiofrequency detection and treatment method, the method comprising:

[0007] During the radio frequency output interval, a time-division DC measurement channel is established to apply a stepped DC excitation signal to the target biological tissue, and the tissue voltage response signal and tissue deformation image sequence corresponding to the stepped DC excitation signal are acquired simultaneously.

[0008] Based on the tissue deformation image sequence, the structural displacement trajectory inside the tissue is analyzed, and combined with the tissue voltage response signal, a dynamic impedance mapping relationship with tissue structural displacement as the variable is established.

[0009] The dynamic impedance mapping relationship is input into a pre-trained deep learning network, which includes a tissue state evolution reasoning branch and a thermal diffusion simulation branch. Through the interactive calculation of the two branches, an energy deposition distribution map and a tissue degeneration state distribution map of the current treatment area inside the tissue are generated.

[0010] Based on the energy deposition distribution map, the initial configuration parameters of radiofrequency energy for the next treatment cycle are calculated. At the same time, based on the tissue denaturation state distribution map, the undertreated sub-regions that require energy replenishment are identified.

[0011] By combining the initial radio frequency energy configuration parameters and the location information of the undertreated sub-region, a composite radio frequency control command containing spatial energy modulation scheme and temporal energy modulation scheme is generated.

[0012] Preferably, establishing the dynamic impedance mapping relationship with organizational structure displacement as the variable includes the following steps:

[0013] The tissue deformation image sequence is analyzed frame by frame. The boundary contour features of different spatial locations of the tissue and the movement vector of the boundary contour of each spatial location are extracted from the single frame image. The starting point and the ending point of the movement vector are connected to form the microstructure motion trajectory line corresponding to different spatial locations inside the tissue.

[0014] The compressive displacement component along the RF electrode axis and the shear displacement component perpendicular to the RF electrode axis are separated from the motion trajectory line of the microstructure.

[0015] Synchronize the timestamps to align the timing of each step change in the stepped DC excitation signal with the acquisition timing of the compression displacement component and the shear displacement component.

[0016] The ratio of the change in the compressive displacement component to the change in the tissue voltage response before and after each DC excitation step change is calculated to obtain the compressive dynamic impedance coefficient of the tissue.

[0017] The ratio of the change in shear displacement component to the change in tissue voltage response before and after each DC excitation step change is calculated to obtain the shear dynamic impedance coefficient of the tissue.

[0018] The compression dynamic impedance coefficient and the shear dynamic impedance coefficient, which change with time, are constructed into a dynamic impedance field in three-dimensional space according to their corresponding spatial coordinates. The dynamic impedance field serves as the dynamic impedance mapping relationship.

[0019] Preferably, inputting the dynamic impedance mapping relationship into the pre-trained deep learning network includes the following parallel processing steps:

[0020] The current snapshot of the dynamic impedance field is input into the tissue state evolution inference branch. This branch extracts the spatial texture features of the impedance field through a multi-layer convolutional network and predicts the local impedance change trend of the tissue at different spatial locations in the next moment based on historical time-series impedance field data. The output is a tissue degeneration probability cloud map.

[0021] The current snapshot of the dynamic impedance field and the current radio frequency energy parameters are input into the thermal diffusion simulation branch. This branch simulates the conduction and absorption process of radio frequency energy in the electrical properties of the tissue described by the dynamic impedance field, and outputs the temperature field and energy density field at different spatial locations inside the tissue in theory.

[0022] In the middle layer of the deep learning network, an interactive channel is established connecting the organization state evolution inference branch and the heat diffusion simulation branch;

[0023] After multiple iterations and interactions, the final output of the tissue state evolution inference branch is defined as the tissue denaturation state distribution map, and the final output of the thermal diffusion simulation branch is defined as the energy deposition distribution map.

[0024] Preferably, the step of calculating the initial configuration parameters of the radiofrequency energy for the next treatment cycle based on the energy deposition distribution map includes:

[0025] On the energy deposition distribution map, with the preset energy deposition uniformity as the target, the variance of energy deposition at different spatial locations is calculated;

[0026] The spatial locations in the energy deposition distribution map where the energy deposition value is lower than a first preset threshold are identified and marked as energy-insufficient areas; the spatial locations where the energy deposition value is higher than a second preset threshold are identified and marked as energy-overload areas.

[0027] Based on the total area and spatial distribution of the energy-deficient region, the total radio frequency energy value that needs to be supplemented is calculated, and the focusing parameters of the radio frequency electrode are recommended according to its distribution shape.

[0028] Based on the spatial location of the energy overload region, calculate its relative distance to the energy underload region, and adjust the distribution weight of the radio frequency output among the multiple electrodes accordingly.

[0029] Based on the total radiofrequency energy value that needs to be supplemented, the recommended radiofrequency electrode focusing parameters, and the adjusted radiofrequency output allocation weight, the initial values ​​of voltage, current, frequency, and duty cycle that the radiofrequency generator should load when the next treatment cycle starts are generated as the initial configuration parameters of the radiofrequency energy.

[0030] Preferably, identifying undertreated sub-regions requiring energy replenishment based on the tissue degeneration state distribution map includes:

[0031] On the tissue denaturation state distribution map, a probability threshold representing the state in which the tissue has undergone sufficient denaturation is set;

[0032] Traverse all spatial cells in the tissue degeneration state distribution map, filter out cells whose tissue degeneration state probability is lower than the state probability threshold, and mark these cells as potential undertreated cells;

[0033] Spatial clustering analysis is performed on the potential undertreated units to group spatially continuous or adjacent potential undertreated units into the same treatment sub-region;

[0034] Calculate the geometric center location, volume, and average and gradient of the probability of tissue degeneration state within each treatment sub-region;

[0035] Based on the average of the region volume and the probability of the tissue degeneration state, the additional energy dose required to bring the region to the state probability threshold is estimated, and this dose, the geometric center location, and the gradient information of the probability of the tissue degeneration state are used together as a complete description of the undertreated sub-region.

[0036] Preferably, the generation of composite radio frequency control commands including spatial energy modulation schemes and temporal energy modulation schemes includes:

[0037] An optimization function is established with the objectives of minimizing total treatment energy and maximizing the uniformity of treatment effect, where the decision variables are the time history curves of the output power of each electrode in the radiofrequency electrode array;

[0038] The initial configuration parameters of the radio frequency energy are used as the initial solution of the optimization function;

[0039] The geometric center location of the undertreated sub-region, the required additional energy dose information, and the location of the energy overload region in the energy deposition distribution map are used as constraints input into the optimization function.

[0040] By solving the optimization function, a set of optimal RF electrode output power time history curves are obtained;

[0041] From the optimal RF electrode output power time history curve, the power distribution ratio between each electrode at the same time is analyzed and encoded as the spatial energy modulation scheme.

[0042] From the optimal RF electrode output power time history curve, the sequence of power change of a single electrode over time is extracted and encoded as the time energy modulation scheme.

[0043] The spatial energy modulation scheme and the temporal energy modulation scheme are packaged together to generate the composite radio frequency control command that can be directly parsed and executed by the radio frequency power amplifier.

[0044] Preferably, the synchronous acquisition of tissue voltage response signals and tissue deformation image sequences corresponding to the stepped-changing DC excitation signals includes:

[0045] A DC signal source is controlled to generate a voltage step sequence with a fixed time interval, and the voltage amplitude of each step is adaptively adjusted according to the tissue voltage response measured in the previous step.

[0046] After each voltage step is applied, a preset electrical stabilization delay time is waited, and then the DC current value flowing through the tissue is acquired by a high-precision analog-to-digital converter, and the corresponding excitation voltage value is recorded, which together constitute a set of tissue voltage response signals.

[0047] Throughout the application of the voltage step sequence, a high-speed optical coherence tomography imaging system is synchronously triggered to continuously acquire cross-sectional tomographic image sequences of the target biological tissue at a frame rate no less than the frequency of the voltage step change.

[0048] Motion artifact correction and image registration are performed on the cross-sectional tomographic image sequence, and displacement data of specific marker points at the tissue-electrode contact interface and inside the tissue are extracted to form the tissue deformation image sequence that is strictly synchronized in time with the voltage step sequence.

[0049] Preferably, the step of performing frame-by-frame analysis on the tissue deformation image sequence to extract the boundary contour features of the tissue at different spatial locations in a single frame image and the movement vector of the boundary contour at each spatial location includes:

[0050] An edge detection algorithm is used to process the images in the single-frame tissue deformation image sequence to identify clear boundary lines at different spatial locations between the tissue and the external medium, and integrate them to form a complete tissue boundary contour.

[0051] On the clearly defined boundary line, a series of feature points are selected at fixed intervals according to different spatial locations;

[0052] In two adjacent frames, for the same feature point, template matching is performed within a preset neighborhood range in the next frame to find the best matching position for the feature point.

[0053] Calculate the spatial coordinate difference of the feature point in two adjacent frames of the image. This difference is the movement vector of the feature point during this time interval.

[0054] Statistical analysis is performed on the movement vectors calculated from all feature points on the clear boundary line to filter out abnormal movement vectors caused by noise, thereby obtaining a set of effective movement vectors that characterize the overall deformation trend of the organization.

[0055] Preferably, the pre-training process of the deep learning network includes:

[0056] A training dataset containing a large number of samples is constructed. Each sample includes: time-series data of the dynamic impedance field generated by simulation or acquired in history, the corresponding real radio frequency energy parameters, the real tissue energy deposition distribution results obtained by infrared thermography or pathological sections, and the real tissue degeneration status results obtained by postoperative medical imaging evaluation.

[0057] The time-series data of the dynamic impedance field in the sample and the corresponding real radio frequency energy parameters are input into the deep learning network to be trained to obtain the predicted energy deposition distribution map and the predicted tissue degeneration state distribution map.

[0058] Calculate the first loss function between the predicted energy deposition distribution map and the actual tissue energy deposition distribution result;

[0059] Calculate a second loss function between the predicted tissue degeneration state distribution map and the actual tissue degeneration state result;

[0060] The first loss function and the second loss function are weighted and summed to obtain the total loss function;

[0061] The backpropagation algorithm is used to adjust the network weight parameters of the tissue state evolution inference branch and the heat diffusion simulation branch in the deep learning network in order to minimize the total loss function;

[0062] The above process is repeated iteratively until the total loss function converges, thus completing the training of the deep learning network and obtaining the pre-trained deep learning network.

[0063] Preferably, when the processor executes the computer program, it implements the steps of the AI-based radio frequency detection and treatment method as described in any of the above-mentioned embodiments.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] During the energy output intervals of radiofrequency therapy, a stepped DC excitation signal is actively applied, and the resulting tissue voltage response and tissue deformation images are simultaneously acquired. Utilizing the interference-free window during treatment breaks, the high sensitivity of low-frequency DC signals to changes in the tissue's microscopic electrical properties is actively detected. The simultaneously acquired electrical response and mechanical deformation images constitute multi-dimensional data reflecting the tissue's dielectric and structural mechanical properties. This multimodal synchronous acquisition method can capture subtle features that are impossible to obtain with traditional single radiofrequency impedance monitoring or static imaging, indicating earlier changes in tissue state.

[0066] The dynamic impedance relationship obtained from the analysis is input into a pre-trained deep learning network, which includes a tissue state evolution inference branch and a thermal diffusion simulation branch, and they interact to perform computations. The inference branch learns a complex mapping from multimodal features to the degree of tissue denaturation, while the simulation branch calculates the spatial distribution of energy based on the laws of biological heat transfer physics. The two branches continuously interact during the computation process, enabling the real-time inferred tissue state to dynamically correct thermophysical parameters, while the updated temperature field also feeds back to influence the prediction of the denaturation rate. This coupled computation mechanism realizes the joint dynamic simulation and inversion of the physical processes of energy deposition and the biochemical processes of tissue denaturation.

[0067] Parameter calculation and region identification are performed based on high-precision energy deposition distribution maps and tissue denaturation state distribution maps. The energy deposition distribution is used to derive the initial energy configuration for the next treatment cycle, while the tissue denaturation state distribution is used to accurately locate sub-regions that have not been sufficiently denatured after energy application. The generated composite radiofrequency control commands, with spatial energy modulation for targeted compensation of undertreated areas and temporal energy modulation for optimizing the output rhythm based on predicted efficiency, ultimately achieve coordinated closed-loop control of radiofrequency energy in the spatiotemporal dimensions. Attached Figure Description

[0068] Figure 1 is a schematic diagram illustrating the working principle of the AI-based radiofrequency detection and treatment method described in this invention.

[0069] Figure 2 is a flowchart for establishing dynamic impedance mapping relationships;

[0070] Figure 3 is a flowchart for identifying undertreated sub-regions;

[0071] Figure 4 shows the 3D dynamic impedance field distribution;

[0072] Figure 5 shows the spatial distribution cloud map of the dynamic impedance coefficient of compression. Detailed Implementation

[0073] 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.

[0074] Referring to Figure 1, this invention provides an AI-based radiofrequency detection and treatment method. The method includes: during the intermittent period of radiofrequency energy output, the system switches to a DC measurement mode and establishes a time-division DC measurement channel. A series of stepped-variable DC excitation signals are applied to the target biological tissue through this channel, and tissue voltage response signals and tissue deformation image sequences corresponding to the stepped-variable DC excitation signals are acquired simultaneously. Based on the tissue deformation image sequence, the structural displacement trajectory inside the tissue can be analyzed. Combined with the tissue voltage response signal, a dynamic impedance mapping relationship with tissue structural displacement as the variable can be established. The dynamic impedance mapping relationship is input into a pre-trained deep learning network, which includes a tissue state evolution inference branch and a thermal diffusion simulation branch. Through the interactive calculation of the two branches, an energy deposition distribution map and a tissue denaturation state distribution map of the current treatment area inside the tissue are generated. Based on the energy deposition distribution map, the initial configuration parameters of the radiofrequency energy for the next treatment cycle are calculated. Simultaneously, based on the tissue denaturation state distribution map, undertreated sub-regions requiring energy replenishment are identified. Finally, by combining the initial radio frequency energy configuration parameters with the location information of the undertreated sub-region, a composite radio frequency control command containing spatial energy modulation and temporal energy modulation schemes is generated to guide the precise treatment in the next cycle.

[0075] Example 1: Referring to Figure 2, a time-division DC measurement channel is established during the radio frequency output interval. A stepped DC excitation signal is applied to the target biological tissue, and the tissue voltage response signal and tissue deformation image sequence corresponding to the stepped DC excitation signal are acquired simultaneously. A dynamic impedance mapping relationship with tissue structure displacement as the variable is established, specifically including the following steps: the tissue deformation image sequence is analyzed frame by frame, and the movement vector of the tissue boundary contour is extracted using an edge detection algorithm. The starting point and ending point of the movement vector are connected to form the microstructure motion trajectory line inside the tissue. The compression displacement component along the RF electrode axis and the shear displacement component perpendicular to the RF electrode axis are separated from the microstructure motion trajectory line. The timestamp is synchronized, and the time of each step change of the stepped DC excitation signal is aligned with the acquisition time of the compression displacement component and the shear displacement component. The ratio of the change in the compression displacement component to the change in the tissue voltage response before and after each DC excitation step change is calculated to obtain the tissue's compression dynamic impedance coefficient. The ratio of the change in shear displacement component to the change in tissue voltage response before and after each DC excitation step change is calculated to obtain the shear dynamic impedance coefficient of the tissue.

[0076] The compression dynamic impedance coefficient and the shear dynamic impedance coefficient, which vary with time, are constructed into a three-dimensional dynamic impedance field according to their corresponding spatial coordinates. This dynamic impedance field serves as the dynamic impedance mapping relationship. The dynamic impedance mapping relationship is input into a pre-trained deep learning network. This process includes the following parallel processing steps: A snapshot of the current dynamic impedance field is input into the tissue state evolution inference branch. This branch extracts the spatial texture features of the impedance field through a multi-layer convolutional network and predicts the local impedance change trend of the tissue at the next moment based on historical temporal impedance field data, outputting a tissue denaturation probability cloud map. The current snapshot of the dynamic impedance field and the current radio frequency energy parameters are input into the thermal diffusion simulation branch. This branch simulates the conduction and absorption process of radio frequency energy in the tissue with the electrical properties described by the dynamic impedance field, outputting the theoretical internal temperature field and energy density field of the tissue. An interactive channel connecting the tissue state evolution inference branch and the thermal diffusion simulation branch is established in the middle layer of the deep learning network. After multiple iterations and interactions, the final output of the tissue state evolution inference branch is defined as the tissue denaturation state distribution map, and the final output of the thermal diffusion simulation branch is defined as the energy deposition distribution map.

[0077] In practical implementation, a time-division DC measurement channel is established during the radio frequency output interval. A stepped DC excitation signal is applied to the target liver tissue, and the tissue voltage response signal and tissue deformation image sequence corresponding to the stepped DC excitation signal are acquired simultaneously. Based on the tissue deformation image sequence, the structural displacement trajectory within the tissue is analyzed, and a dynamic impedance mapping relationship is established. This includes frame-by-frame analysis of the tissue deformation image sequence, using an edge detection algorithm to process single-frame images, identifying clear boundary lines at different spatial locations between the tissue and the external medium, and integrating them to form a complete tissue boundary contour. On the clear boundary lines, a series of feature points are selected at fixed intervals according to different spatial locations. In two adjacent frames, for the same feature point, template matching is performed within a preset neighborhood range in the next frame to find the optimal matching position for that feature point. The spatial coordinate difference of the feature point in two adjacent frames is calculated to obtain the movement vector of the feature point within this time interval. Statistical analysis is performed on the movement vectors calculated for all feature points on the clear boundary lines, filtering out abnormal movement vectors caused by noise, to obtain a set of effective movement vectors characterizing the overall tissue deformation trend. Connecting the starting and ending points of the movement vectors forms the microstructural motion trajectory lines corresponding to different spatial locations within the organization.

[0078] In some embodiments, the compressive displacement component along the RF electrode axis and the shear displacement component perpendicular to the RF electrode axis are separated from the microstructure motion trajectory. A synchronization timestamp aligns the timing of each step change in the stepped DC excitation signal with the acquisition timing of the compressive and shear displacement components. The ratio of the change in the compressive displacement component to the change in the tissue voltage response before and after each DC excitation step change is calculated to obtain the tissue's compressive dynamic impedance coefficient. Similarly, the ratio of the change in the shear displacement component to the change in the tissue voltage response before and after each DC excitation step change is calculated to obtain the tissue's shear dynamic impedance coefficient. The time-varying compressive and shear dynamic impedance coefficients are then constructed into a three-dimensional dynamic impedance field according to their corresponding spatial coordinates, serving as a dynamic impedance mapping relationship. The following formula is used to calculate the compressive dynamic impedance coefficient:

[0079] Where: Z c (t) represents the compressive dynamic impedance coefficient at time t, Δd c ΔV(t) represents the change in the compressive displacement component at time t, and ΔV(t) represents the change in the tissue voltage response at time t.

[0080] It can be understood that inputting the dynamic impedance mapping relationship into the pre-trained deep learning network involves the following parallel processing: A snapshot of the current dynamic impedance field is input into the tissue state evolution inference branch. This branch extracts the spatial texture features of the impedance field through a multi-layer convolutional network and predicts the local impedance change trend of the tissue at different spatial locations in the next moment based on historical temporal impedance field data, outputting a tissue denaturation probability cloud map. The current snapshot of the dynamic impedance field and the current radio frequency energy parameters are input into the thermal diffusion simulation branch. This branch simulates the conduction and absorption process of radio frequency energy in the tissue with the electrical properties described by the dynamic impedance field, outputting the theoretical temperature field and energy density field at different spatial locations within the tissue. In the middle layer of the deep learning network, an interactive channel connecting the tissue state evolution inference branch and the thermal diffusion simulation branch is established. After multiple iterations, the final output of the tissue state evolution inference branch is defined as the tissue denaturation state distribution map, and the final output of the thermal diffusion simulation branch is defined as the energy deposition distribution map.

[0081] In practice, the interaction channel between the tissue state evolution inference branch and the thermal diffusion simulation branch allows the probability information in the tissue denaturation probability cloud map to be fed back to the thermal diffusion simulation branch to adjust the energy conduction simulation parameters. At the same time, the distribution information in the energy density field is fed back to the tissue state evolution inference branch to correct the impedance change trend prediction.

[0082] It is understandable that the construction of the dynamic impedance field and the input processing of the dual-branch deep learning network present data comparisons in the example scenario. For instance, the compressive and shear displacement components extracted from tissue deformation image sequences numerically reflect the differences in tissue response in different mechanical directions, while the calculated compressive and shear dynamic impedance coefficients quantify the changes in electro-mechanical coupling characteristics in different spatial directions. After these data are input into the deep learning network as components of the dynamic impedance field, the tissue degeneration probability cloud map output by the tissue state evolution inference branch and the energy density field output by the thermal diffusion simulation branch may show correlation in spatial distribution to support interactive computation. Optionally, during parallel processing, the current snapshot of the dynamic impedance field is fed into both branches simultaneously, but the tissue state evolution inference branch additionally relies on historical temporal impedance field data sequences to capture temporal evolution patterns, while the thermal diffusion simulation branch combines real-time radio frequency energy parameters such as power and frequency for physical simulation. In specific implementations, the interactive channel is implemented based on an attention mechanism, where the feature maps generated by the tissue state evolution inference branch and the feature maps generated by the thermal diffusion simulation branch undergo cross-branch information fusion at a specific network layer to update their respective computational paths.

[0083] Example 2: Referring to Figure 3, based on the energy deposition distribution map, the initial configuration parameters of radiofrequency energy for the next treatment cycle are calculated. This process includes: on the energy deposition distribution map, with a preset energy deposition uniformity as the target, calculating the variance of energy deposition at different spatial locations. Identifying spatial locations in the energy deposition distribution map where the energy deposition value is lower than a first preset threshold and marking them as energy-deficient areas; identifying spatial locations where the energy deposition value is higher than a second preset threshold and marking them as energy-overload areas. Calculating the total radiofrequency energy value that needs to be supplemented based on the total area and spatial distribution of the energy-deficient areas, and recommending the focusing parameters of the radiofrequency electrodes based on their distribution shape. Calculating the relative distance between the energy-overload areas and the energy-deficient areas based on their spatial location, and adjusting the distribution weight of the radiofrequency output among the multiple electrodes accordingly. Combining the total radiofrequency energy value that needs to be supplemented, the recommended radiofrequency electrode focusing parameters, and the adjusted radiofrequency output distribution weight, generating the initial values ​​of voltage, current, frequency, and duty cycle that the radiofrequency generator should load when the next treatment cycle starts, as the initial configuration parameters of the radiofrequency energy.

[0084] Based on the tissue degeneration state distribution map, undertreated sub-regions requiring energy replenishment are identified. This process includes: setting a state probability threshold representing sufficient tissue degeneration on the tissue degeneration state distribution map; traversing all spatial cells in the tissue degeneration state distribution map and filtering out cells with a tissue degeneration state probability lower than the state probability threshold, marking these cells as potential undertreated cells; performing spatial clustering analysis on the potential undertreated cells, grouping spatially continuous or adjacent potential undertreated cells into the same treatment sub-region; calculating the geometric center location, region volume, and the average and gradient of the tissue degeneration state probability within each treatment sub-region; estimating the additional energy dose required to bring the region to the state probability threshold based on the region volume and the average tissue degeneration state probability, and using this dose, the geometric center location, and the gradient information of the tissue degeneration state probability together as a complete description of the undertreated sub-region.

[0085] In practice, the initial configuration parameters of radiofrequency energy for the next treatment cycle are calculated based on the energy deposition distribution map. This process includes calculating the variance of the current energy deposition on the energy deposition distribution map with a preset energy deposition uniformity as the target. The variance is calculated using the following formula:

[0086]

[0087] σ 2 E represents the variance of the energy deposition values, N represents the total number of spatial units within the energy deposition distribution map, and E represents the variance of the energy deposition values. iLet represent the energy deposition value of the i-th spatial unit, and μ represent the arithmetic mean of the energy deposition values ​​of all spatial units. Regions in the energy deposition distribution map with energy deposition values ​​below a first preset threshold are identified and marked as energy-deficient regions, while regions with energy deposition values ​​above a second preset threshold are identified and marked as energy-overloaded regions.

[0088] In some embodiments, the total radiofrequency energy required to be supplemented is calculated based on the total area and spatial distribution of the energy-deficient region, and focusing parameters of the radiofrequency electrodes, such as focusing depth and focal spot width, are recommended based on the distribution shape of the energy-deficient region. The relative distance between the energy-overloaded region and the energy-deficient region is calculated based on the spatial location of the energy-overloaded region, and the distribution weight of the radiofrequency output among the multiple electrodes is adjusted accordingly. When the relative distance is close, the distribution weight tends to reduce the power output of the corresponding electrode to avoid heat accumulation. Combining the total radiofrequency energy required to be supplemented, the recommended radiofrequency electrode focusing parameters, and the adjusted radiofrequency output distribution weight, initial values ​​for the voltage, current, frequency, and duty cycle that the radiofrequency generator should load at the start of the next treatment cycle are generated as the initial radiofrequency energy configuration parameters.

[0089] It is understandable that identifying undertreated sub-regions requiring energy replenishment based on the tissue degeneration state distribution map involves setting a state probability threshold on the tissue degeneration state distribution map to represent the state of sufficient tissue degeneration, traversing all spatial units in the tissue degeneration state distribution map, and filtering out units whose tissue degeneration state probability is lower than the state probability threshold, marking these units as potential undertreated units. Optionally, spatial clustering analysis can be performed on potential undertreated units to group spatially continuous or adjacent potential undertreated units into the same treatment sub-region, for example, using a distance-based clustering algorithm to merge sets of units with a centroid distance less than a set value.

[0090] In some embodiments, the geometric center location, volume, and average and gradient of the probability of tissue degeneration states within each treatment sub-region are calculated. The additional energy dose required to bring the region to a state probability threshold is estimated based on the region volume and the average probability of tissue degeneration states. It is understood that the estimation of the additional energy dose may be based on a pre-defined energy-state probability conversion model. This additional energy dose, along with the geometric center location and the gradient information of the probability of tissue degeneration states, together constitute a complete description of the undertreated sub-region. In specific implementations, the data comparison is reflected in the fact that the spatial locations of the energy-deficient region and the undertreated sub-region may not completely overlap. The energy-deficient region mainly reflects the non-uniformity of energy deposition in the current cycle, while the undertreated sub-region reflects the lag in the tissue degeneration process. Both indicate, from different dimensions, the focal areas where the treatment strategy needs to be specifically adjusted in the next cycle.

[0091] Example 3: Combining the initial radio frequency energy configuration parameters and the location information of the undertreated sub-region, a composite radio frequency control command containing spatial energy modulation and temporal energy modulation schemes is generated. This process includes: establishing an optimization function with the objectives of minimizing total treatment energy and maximizing the uniformity of treatment effect, where the decision variables are the output power time history curves of each electrode in the radio frequency electrode array. The initial radio frequency energy configuration parameters are used as the initial solution of the optimization function. The geometric center position of the undertreated sub-region, the required additional energy dose information, and the energy overload region position in the energy deposition distribution map are used as constraints input into the optimization function. By solving the optimization function, a set of optimal radio frequency electrode output power time history curves are obtained. From the optimal radio frequency electrode output power time history curves, the power distribution ratio between each electrode at the same time is parsed out and encoded as the spatial energy modulation scheme. From the optimal radio frequency electrode output power time history curves, the power change sequence of a single electrode over time is parsed out and encoded as the temporal energy modulation scheme. The spatial energy modulation scheme and the temporal energy modulation scheme are packaged to generate the composite radio frequency control command that can be directly parsed and executed by the radio frequency power amplifier.

[0092] In practical implementation, a composite radio frequency control command, incorporating both spatial and temporal energy modulation schemes, is generated by integrating the initial radio frequency energy configuration parameters with the location information of the undertreated sub-region. This process involves establishing an optimization function aimed at minimizing total treatment energy and maximizing the uniformity of treatment effect, where the decision variable is the time history curve of the output power of each electrode in the radio frequency electrode array. The specific form of the optimization function can be expressed as:

[0093]

[0094] in: T represents the dimensionless objective function value to be minimized, where λ and η are preset dimensionless weighting coefficients. f This indicates the total duration of the next treatment cycle, where M represents the total number of electrodes in the radiofrequency electrode array. This indicates that the m-th electrode is at time [time missing]. The output power, E ref This indicates a preset reference energy value used for normalization, and V represents the set of all voxels within the treatment area. This represents the predicted probability value of tissue degeneration state for voxel v. This represents the average of the predicted tissue denaturation probabilities for all voxels. The initial configuration parameters of the radio frequency energy are used as the initial solution to the optimization function; for example, the initial solution is set so that all electrodes operate at a constant power specified by the initial configuration parameters of the radio frequency energy during the cycle.

[0095] In some embodiments, the geometric center location of the undertreated sub-region, the required additional energy dose information, and the location of the energy overload region in the energy deposition distribution map are used as constraints input into the optimization function. Specifically, the constraints include that the accumulated energy deposition of each undertreated sub-region at the end of treatment must not be less than the sum of its required additional energy dose and baseline energy, while the energy deposition rate of the energy overload region must not exceed a set safety limit at any time during treatment. A set of optimal radiofrequency electrode output power time-history curves are obtained by solving the optimization function. The solution process can employ numerical optimization algorithms, such as sequential quadratic programming, to iteratively calculate the optimization function until the convergence condition is met. In its implementation, the sequential quadratic programming algorithm, as a numerical optimization method for solving optimization functions, involves transforming the nonlinear optimization problem into a series of quadratic programming subproblems for iterative solution. The algorithm initializes with initial RF energy configuration parameters as the starting point. In each iteration, a quadratic approximation model of the current point is constructed, containing gradient information of the objective function, Hessian matrix estimation, and linearization of constraints. During the solution process, the algorithm updates candidate solutions to the electrode output power time history curves by calculating the search direction and determining the step size, while rigorously verifying the satisfaction of the additional energy dose constraints required for the undertreated sub-region and the safety upper limit constraints for the energy overload region. Iterative calculations continue until the change in the objective function value between two adjacent iterations is less than a preset threshold or the adjustment magnitude of the decision variables approaches zero. At this point, the optimization process is considered to have met the convergence condition, and the final set of optimal RF electrode output power time history curves is output. The power distribution ratio between electrodes at the same time is extracted from the optimal RF electrode output power time history curves, for example, at time t. k Calculate the power P of each electrode. m (t k ) accounts for the total power of all electrodes The percentage is encoded as a space energy modulation scheme.

[0096] It is understandable that the sequence of power changes of a single electrode over time can be extracted from the optimal RF electrode output power time history curve, for example, extracting the power of the m-th electrode at discrete time points. Power value sequence on The spatial energy modulation scheme is encoded into a time-energy modulation scheme. Optionally, the spatial energy modulation scheme and the time-energy modulation scheme are packaged to generate a composite RF control instruction that can be directly parsed and executed by the RF power amplifier. The composite RF control instruction encapsulates the power allocation ratio matrix in the spatial energy modulation scheme and the power time series array in the time-energy modulation scheme using a data structure. In specific implementation, the data comparison is reflected in the fact that based on the same set of initial conditions and constraints, different weighting coefficients λ and η are used to obtain different optimal RF electrode output power time history curves, which leads to differences in the power allocation ratio in the spatial energy modulation scheme and differences in the power change pattern in the time-energy modulation scheme. For example, increasing the weight of λ will make the optimization result more inclined to reduce the total energy consumption, which may reduce the overall amplitude of the power time series. On the other hand, increasing the weight of η will make the optimization result more inclined to improve the treatment uniformity, which may make the power allocation in the spatial energy modulation scheme focus more on the edge region. The final form of the composite RF control instruction is uniquely determined by the optimal RF electrode output power time history curve obtained by actual solution.

[0097] Example 4: Synchronously acquiring tissue voltage response signals and tissue deformation image sequences corresponding to the stepped DC excitation signal. Specifically, this involves controlling a DC signal source to generate a voltage step sequence with fixed time intervals. The voltage amplitude of each step is adaptively adjusted based on the previously measured tissue voltage response. After each voltage step is applied, a preset electrical stabilization delay is waited for, and then the DC current value flowing through the tissue is acquired via a high-precision analog-to-digital converter, and the corresponding excitation voltage value is recorded, together forming a set of tissue voltage response signals. Throughout the application of the voltage step sequence, a high-speed optical coherence tomography (OCT) imaging system is synchronously triggered to continuously acquire cross-sectional tomographic image sequences of the target biological tissue at a frame rate no less than the voltage step change frequency. Motion artifact correction and image registration are performed on the cross-sectional tomographic image sequences, and displacement data of the tissue-electrode contact interface and specific marker points within the tissue are extracted to form the tissue deformation image sequence that is strictly synchronized in time with the voltage step sequence. The process involves frame-by-frame analysis of the tissue deformation image sequence to extract the movement vectors of the tissue boundary contours. This includes: processing images in a single frame of the tissue deformation image sequence using an edge detection algorithm to identify clear boundary lines at different spatial locations between the tissue and the external medium, and integrating these lines to form a complete tissue boundary contour. On these clear boundary lines, a series of feature points are selected at fixed intervals according to different spatial locations. In two adjacent frames, for the same feature point, template matching is performed within a preset neighborhood range in the next frame to find the optimal matching position for that feature point. The spatial coordinate difference between the feature point and the adjacent frames is calculated; this difference is the movement vector of the feature point within that time interval. Statistical analysis is performed on the movement vectors calculated for all feature points on the clear boundary lines to filter out abnormal movement vectors caused by noise, resulting in a set of effective movement vectors characterizing the overall tissue deformation trend.

[0098] In practice, the synchronous acquisition of tissue voltage response signals and tissue deformation image sequences corresponding to the stepped DC excitation signals includes controlling a DC signal source to generate a voltage step sequence with fixed time intervals. The voltage amplitude of each step is adaptively adjusted based on the tissue voltage response measured in the previous measurement. For example, if the tissue voltage response in the previous measurement showed a high impedance, the voltage amplitude of the next step can be appropriately increased to ensure the significance of the measured signal. After each voltage step is applied, a preset electrical stabilization delay time is waited, and then the DC current value flowing through the tissue is acquired through a high-precision analog-to-digital converter, and the corresponding excitation voltage value is recorded to form a set of tissue voltage response signals. The electrical stabilization delay time is used to eliminate transient responses and ensure that the acquired signal is in a steady state. Throughout the entire process of applying the voltage step sequence, a high-speed optical coherence tomography imaging system is synchronously triggered to continuously acquire cross-sectional tomographic image sequences of the target biological tissue at a frame rate no less than the voltage step change frequency, ensuring that electrical measurements and image acquisition are strictly synchronized in time. Motion artifact correction and image registration were performed on the cross-sectional tomographic image sequence. Displacement data of specific marker points at the tissue-electrode contact interface and inside the tissue were extracted to form a tissue deformation image sequence that was strictly synchronized in time with the voltage step sequence, as shown in Table 1.

[0099] Table 1: Exemplary Voltage Step Sequence and Synchronous Acquisition Data Table

[0100]

[0101] It is understood that frame-by-frame analysis of tissue deformation image sequences and extraction of movement vectors of tissue boundary contours includes processing images in a single frame of the tissue deformation image sequence using an edge detection algorithm to identify clear boundary lines between the tissue and the external medium, and selecting a series of feature points at fixed intervals on the clear boundary lines. In some embodiments, template matching is performed on the same feature point in two adjacent frames within a preset neighborhood of the next frame to find the optimal matching position of the feature point. Template matching can be calculated using a normalized cross-correlation algorithm, and the spatial coordinate difference of the feature point in two adjacent frames is calculated as the movement vector of the feature point within this time interval. Optionally, statistical analysis is performed on the movement vectors calculated for all feature points on the clear boundary lines to filter out abnormal movement vectors caused by noise. The statistical analysis includes calculating the average direction and amplitude of all movement vectors and removing vectors that deviate from the average by more than three standard deviations, resulting in a set of effective movement vectors characterizing the overall deformation trend of the tissue.

[0102] In practical implementation, data comparison reveals a correlation in amplitude between the tissue voltage response signal acquired under different excitation voltage steps and the movement vector resolved from the tissue deformation image sequence. As the voltage step increases, the acquired current value increases accordingly, and the amplitude of the feature point's movement vector also shows an increasing trend. This verifies the synchronicity between electro-mechanical excitation and response. The amplitude of the feature point's movement vector is calculated using the following formula:

[0103]

[0104] in: Let represent the magnitude of the movement vector of the i-th feature point between two adjacent image frames (time t to t+1). and These represent the two-dimensional spatial coordinates of the feature point in the images at times t and t+1, respectively. In some embodiments, the accuracy of synchronous acquisition is guaranteed by a timestamp alignment mechanism. Each step change of the DC excitation signal, the sampling time of the high-precision analog-to-digital converter, and the imaging exposure time of the high-speed optical coherence tomography imaging system are all triggered by a unified precision clock source. This ensures that there is a precise temporal correspondence between each data point in the tissue voltage response signal and a specific frame in the tissue deformation image sequence. This correspondence is the basis for the subsequent construction of dynamic impedance mapping.

[0105] Referring to Figure 4, the distribution characteristics of the 3D dynamic impedance field in the dimensions of "step time - feature point number - comprehensive dynamic impedance coefficient" are presented. Specifically, this graph constructs a multi-dimensional dynamic impedance field visualization model with the step time (ms) as the time axis, the feature point number as the spatial feature dimension, and the comprehensive dynamic impedance coefficient (Ω) as the electromechanical response quantification index. In the data mapping logic, the color gradient in the figure corresponds to the numerical range of the comprehensive dynamic impedance coefficient (3000Ω to 4200Ω): when the step time is small (e.g., 0-200ms), the impedance coefficient corresponding to the feature point is in the high value range above 4000Ω (shown as yellow-green); as the step time increases (e.g., 300-600ms), the impedance coefficient transitions to the low value range of 3000-3600Ω (shown as blue-purple). This distribution characteristic reflects the electromechanical coupling response law of the tissue under DC excitation step action over time: during the continuous excitation process, the cumulative displacement of the tissue microstructure leads to a gradient decrease in the comprehensive dynamic impedance coefficient, and the impedance change of different feature points has spatial heterogeneity.

[0106] Example 5: The pre-training process of the deep learning network includes: constructing a training dataset containing a large number of samples, each sample including: time-series data of the dynamic impedance field generated by simulation or historically collected data, corresponding real radio frequency energy parameters, real tissue energy deposition distribution results obtained through infrared thermal imaging or pathological sections, and real tissue degeneration state results obtained through postoperative medical image evaluation. The time-series data of the dynamic impedance field and the corresponding real radio frequency energy parameters in the samples are input into the deep learning network to be trained to obtain a predicted energy deposition distribution map and a predicted tissue degeneration state distribution map. A first loss function is calculated between the predicted energy deposition distribution map and the real tissue energy deposition distribution results. A second loss function is calculated between the predicted tissue degeneration state distribution map and the real tissue degeneration state results. The first loss function and the second loss function are weighted and summed to obtain a total loss function. The backpropagation algorithm is used to adjust the network weight parameters of the tissue state evolution inference branch and the thermal diffusion simulation branch in the deep learning network to minimize the total loss function. The above process is repeated iteratively until the total loss function converges, thus completing the training of the deep learning network and obtaining the pre-trained deep learning network.

[0107] In practice, the pre-training process of the deep learning network includes constructing a training dataset containing a large number of samples. Each sample includes time-series data of the simulated or historically acquired dynamic impedance field, corresponding real radiofrequency energy parameters, real tissue energy deposition distribution results obtained through infrared thermography or pathological sections, and real tissue degeneration status results obtained through postoperative medical imaging assessment. Specifically, the time-series data of the simulated dynamic impedance field is calculated based on a finite element model of the conductivity and mechanical properties of biological tissue. The time-series data of the historically acquired dynamic impedance field comes from records actually measured and stored during radiofrequency intervals in past clinical treatments. The real radiofrequency energy parameters record the actual output power, voltage, frequency, and time series of the treatment device. The real tissue energy deposition distribution results are obtained by infrared thermography in model experiments or intraoperative monitoring. The real tissue degeneration status results are determined by comparing preoperative and postoperative CT or MRI images to assess the tissue necrosis area.

[0108] The time-series data of the dynamic impedance field in the sample and the corresponding real radio frequency energy parameters are input into the deep learning network to be trained to obtain the predicted energy deposition distribution map and the predicted tissue degeneration state distribution map. A first loss function is calculated between the predicted energy deposition distribution map and the real tissue energy deposition distribution results. In some embodiments, the first loss function is calculated using a pixel-based mean square error form, as shown in the formula:

[0109]

[0110] Where: L E This represents the value of the first loss function, where W and H represent the width and height of the energy deposition distribution map, respectively. Let E(x,y) represent the predicted energy deposition value at coordinates (x,y), and E(x,y) represent the actual tissue energy deposition value at coordinates (x,y). A second loss function is calculated between the predicted tissue degeneration state distribution map and the actual tissue degeneration state results. This second loss function can combine cross-entropy loss and Dice loss to evaluate the accuracy of pixel classification and region overlap.

[0111] It is understandable that the total loss function is obtained by weighted summation of the first and second loss functions, with the weighting coefficients set according to the relative importance of energy deposition prediction and state prediction in clinical practice. The backpropagation algorithm is used to adjust the network weight parameters of the tissue state evolution inference branch and the thermal diffusion simulation branch in the deep learning network to minimize the total loss function. This process is iteratively executed until the total loss function converges, completing the training of the deep learning network and obtaining a pre-trained deep learning network. In specific implementations, data comparisons show that the total loss function value decreases with increasing iterations during training. Simultaneously, the spatial similarity between the predicted energy deposition distribution map and the actual tissue energy deposition distribution, as well as the spatial overlap between the predicted tissue degeneration state distribution map and the actual tissue degeneration state, both show an increasing trend in the later stages of training. Optionally, the network weights are optimized using an adaptive moment estimation algorithm, with the learning rate dynamically decaying according to the training progress. In some embodiments, the training dataset samples need to undergo preprocessing, including normalization of the time-series data of the dynamic impedance field, spatial interpolation of the actual tissue energy deposition distribution to match the network output resolution, and binarization annotation of the actual tissue degeneration state results.

[0112] Referring to Figure 5, the distribution of compressible dynamic impedance coefficients during the construction phase of the dynamic impedance field is presented in the AI-based radiofrequency detection and treatment method. Specifically, the figure uses axial and tangential coordinates (both in mm) to form a two-dimensional space. Color mapping (from blue to red corresponding to 60~140 Ω·mm) shows the distribution characteristics of the compressible dynamic impedance coefficients of the target biological tissue at different spatial locations during the radiofrequency output interval: the blue areas correspond to lower compressible dynamic impedance coefficients, and the red areas correspond to higher compressible dynamic impedance coefficients, exhibiting a multi-regional heterogeneous distribution. This distribution is obtained based on the compressive displacement component resolved from the tissue deformation image sequence and the synchronously acquired tissue voltage response signal. The ratio of the changes before and after the DC excitation step change is calculated, and this ratio is used to construct the components of the dynamic impedance field according to the corresponding spatial coordinates.

[0113] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based radiofrequency detection and treatment system, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements an AI-based radiofrequency detection and treatment method. The method includes: establishing a time-division DC measurement channel during radiofrequency output intervals; applying a stepped-variable DC excitation signal to the target biological tissue; simultaneously acquiring tissue voltage response signals and tissue deformation image sequences corresponding to the stepped-variable DC excitation signals; analyzing the structural displacement trajectory within the tissue based on the tissue deformation image sequence; and establishing a dynamic impedance mapping relationship with tissue structural displacement as the variable, in conjunction with the tissue voltage response signal. The system inputs the dynamic impedance mapping relationship into a pre-trained deep learning network, which includes a tissue state evolution inference branch and a thermal diffusion simulation branch. Through the interactive calculation of the two branches, an energy deposition distribution map and a tissue denaturation state distribution map of the current treatment area within the tissue are generated. Based on the energy deposition distribution map, the initial configuration parameters of radiofrequency energy for the next treatment cycle are calculated. At the same time, based on the tissue denaturation state distribution map, undertreated sub-regions that require energy replenishment are identified. Combining the initial configuration parameters of radiofrequency energy and the location information of the undertreated sub-regions, a composite radiofrequency control command including a spatial energy modulation scheme and a temporal energy modulation scheme is generated.

2. The AI-based radiofrequency detection and treatment system according to claim 1, characterized in that, The establishment of a dynamic impedance mapping relationship with tissue structure displacement as the variable includes the following steps: performing frame-by-frame analysis on the tissue deformation image sequence, extracting the boundary contour features of different spatial locations of the tissue in a single frame image and the movement vector of the boundary contour of each spatial location, connecting the starting point and the ending point of the movement vector to form the microstructure motion trajectory line corresponding to different spatial locations within the tissue; separating the compression displacement component along the axial direction of the radio frequency electrode and the shear displacement component perpendicular to the axial direction of the radio frequency electrode from the microstructure motion trajectory line; synchronizing the timestamp, aligning the time of each step change of the stepped DC excitation signal with the acquisition time of the compression displacement component and the shear displacement component; calculating the ratio of the change in the compression displacement component to the change in the tissue voltage response before and after each DC excitation step change to obtain the compression dynamic impedance coefficient of the tissue; calculating the ratio of the change in the shear displacement component to the change in the tissue voltage response before and after each DC excitation step change to obtain the shear dynamic impedance coefficient of the tissue; constructing a dynamic impedance field in three-dimensional space based on the time-varying compression dynamic impedance coefficient and the shear dynamic impedance coefficient according to their corresponding spatial coordinates, and using the dynamic impedance field as the dynamic impedance mapping relationship.

3. The AI-based radiofrequency detection and treatment system according to claim 2, characterized in that, The process of inputting the dynamic impedance mapping relationship into the pre-trained deep learning network includes the following parallel processing steps: The current snapshot of the dynamic impedance field is input into the tissue state evolution inference branch. This branch extracts the spatial texture features of the impedance field through a multi-layer convolutional network and predicts the local impedance change trend of the tissue at different spatial locations in the next moment based on historical temporal impedance field data, outputting a tissue denaturation probability cloud map. The current snapshot of the dynamic impedance field and the current radio frequency energy parameters are input into the thermal diffusion simulation branch. This branch simulates the conduction and absorption process of radio frequency energy in the tissue with electrical properties described by the dynamic impedance field, outputting the theoretical temperature field and energy density field at different spatial locations within the tissue. An interactive channel connecting the tissue state evolution inference branch and the thermal diffusion simulation branch is established in the middle layer of the deep learning network. After multiple iterations and interactions, the final output of the tissue state evolution inference branch is defined as the tissue denaturation state distribution map, and the final output of the thermal diffusion simulation branch is defined as the energy deposition distribution map.

4. The AI-based radiofrequency detection and treatment system according to claim 3, characterized in that, The step of calculating the initial radiofrequency energy configuration parameters for the next treatment cycle based on the energy deposition distribution map includes: calculating the variance of energy deposition at different spatial locations on the energy deposition distribution map, with a preset energy deposition uniformity as the target; identifying spatial locations in the energy deposition distribution map where the energy deposition value is lower than a first preset threshold and marking them as energy-deficient areas, and identifying spatial locations where the energy deposition value is higher than a second preset threshold and marking them as energy-overload areas; calculating the total radiofrequency energy value that needs to be supplemented based on the total area and spatial distribution of the energy-deficient areas, and recommending the focusing parameters of the radiofrequency electrodes based on their distribution shape; calculating the relative distance between the energy-overload areas and the energy-deficient areas based on their spatial location, and adjusting the distribution weight of the radiofrequency output among the multiple electrodes accordingly; and generating the initial values ​​of voltage, current, frequency, and duty cycle that the radiofrequency generator should load when the next treatment cycle starts, based on the total radiofrequency energy value that needs to be supplemented, the recommended radiofrequency electrode focusing parameters, and the adjusted radiofrequency output distribution weight, as the initial radiofrequency energy configuration parameters.

5. The AI-based radiofrequency detection and treatment system according to claim 4, characterized in that, The step of identifying undertreated sub-regions requiring energy replenishment based on the tissue degeneration state distribution map includes: setting a state probability threshold on the tissue degeneration state distribution map to indicate that the tissue has undergone sufficient degeneration; traversing all spatial units in the tissue degeneration state distribution map, filtering out units whose tissue degeneration state probability is lower than the state probability threshold, and marking these units as potential undertreated units; performing spatial clustering analysis on the potential undertreated units, grouping spatially continuous or adjacent potential undertreated units into the same treatment sub-region; calculating the geometric center position, region volume, and average and gradient of the tissue degeneration state probability within each treatment sub-region; estimating the additional energy dose required to bring the region to the state probability threshold based on the region volume and the average tissue degeneration state probability, and using the dose, the geometric center position, and the gradient information of the tissue degeneration state probability together as a complete description of the undertreated sub-region.

6. The AI-based radiofrequency detection and treatment system according to claim 5, characterized in that, The generation of a composite radio frequency (RF) control command comprising a spatial energy modulation scheme and a temporal energy modulation scheme includes: establishing an optimization function with the objectives of minimizing total treatment energy and maximizing the uniformity of treatment effect, wherein the decision variables are the output power time history curves of each electrode in the RF electrode array; using the initial configuration parameters of the RF energy as the initial solution of the optimization function; inputting the geometric center position of the undertreated sub-region, the required additional energy dose information, and the energy overload region position in the energy deposition distribution map as constraints into the optimization function; obtaining a set of optimal RF electrode output power time history curves by solving the optimization function; parsing the power distribution ratio between each electrode at the same time from the optimal RF electrode output power time history curves and encoding it as the spatial energy modulation scheme; parsing the power change sequence of a single electrode over time from the optimal RF electrode output power time history curves and encoding it as the temporal energy modulation scheme; and packaging the spatial energy modulation scheme and the temporal energy modulation scheme to generate the composite RF control command that can be directly parsed and executed by the RF power amplifier.

7. The AI-based radiofrequency detection and treatment system according to claim 1, characterized in that, The synchronous acquisition of tissue voltage response signals and tissue deformation image sequences corresponding to the stepped DC excitation signals includes: controlling a DC signal source to generate a voltage step sequence with fixed time intervals, wherein the voltage amplitude of each step is adaptively adjusted according to the previously measured tissue voltage response; after each voltage step is applied, waiting for a preset electrical stabilization delay time, and then acquiring the DC current value flowing through the tissue through a high-precision analog-to-digital converter and recording the corresponding excitation voltage value, which together constitute a set of the tissue voltage response signals; throughout the entire process of applying the voltage step sequence, synchronously triggering a high-speed optical coherence tomography imaging system to continuously acquire a cross-sectional tomographic image sequence of the target biological tissue at a frame rate not lower than the voltage step change frequency; performing motion artifact correction and image registration on the cross-sectional tomographic image sequence, extracting displacement data of the tissue-electrode contact interface and specific marker points inside the tissue, and forming the tissue deformation image sequence that is strictly synchronized with the voltage step sequence in time.

8. The AI-based radiofrequency detection and treatment system according to claim 2, characterized in that, The step of performing frame-by-frame analysis of the tissue deformation image sequence to extract the boundary contour features of different spatial locations of the tissue and the movement vectors of the boundary contours of each spatial location in a single frame image includes: processing the images in a single frame of the tissue deformation image sequence using an edge detection algorithm to identify clear boundary lines at different spatial locations between the tissue and the external medium, and integrating them to form a complete tissue boundary contour; selecting a series of feature points at fixed intervals according to different spatial locations on the clear boundary lines; performing template matching within a preset neighborhood range in the next frame image for the same feature point in two adjacent frames to find the best matching position of the feature point; calculating the spatial coordinate difference of the feature point in two adjacent frames, which is the movement vector of the feature point in this time interval; and performing statistical analysis on the movement vectors calculated for all feature points on the clear boundary lines to filter out abnormal movement vectors caused by noise, thereby obtaining a set of effective movement vectors characterizing the overall deformation trend of the tissue.

9. The AI-based radiofrequency detection and treatment system according to claim 3, characterized in that, The pre-training process of the deep learning network includes: constructing a training dataset containing a large number of samples, each sample including: time-series data of the dynamic impedance field generated by simulation or historically collected, corresponding real radio frequency energy parameters, real tissue energy deposition distribution results obtained through infrared thermal imaging or pathological sections, and real tissue degeneration state results obtained through postoperative medical image evaluation; inputting the time-series data of the dynamic impedance field and the corresponding real radio frequency energy parameters in the samples into the deep learning network to be trained to obtain a predicted energy deposition distribution map and a predicted tissue degeneration state distribution map; calculating a first loss function between the predicted energy deposition distribution map and the real tissue energy deposition distribution results; calculating a second loss function between the predicted tissue degeneration state distribution map and the real tissue degeneration state results; weighted summing of the first loss function and the second loss function to obtain a total loss function; using the backpropagation algorithm to adjust the network weight parameters of the tissue state evolution inference branch and the thermal diffusion simulation branch in the deep learning network to minimize the total loss function; iteratively executing the above process until the total loss function converges, completing the training of the deep learning network, and obtaining the pre-trained deep learning network.

Images