Skin disease treatment electrode positioning system based on potential distribution

Abstract

The invention relates to the technical field of intelligent medical assistance, in particular to a skin disease treatment electrode positioning system based on potential distribution. The system comprises a potential gradient severe region identification module, a lesion minimum surrounding region identification module, a similar potential region grouping module, an optimal electrode position prediction module, an electric field distribution calculation module, a sweat gland dense region identification module, an optimal electrode placement point screening module, an enhanced feature generation module and a final electrode position determination module. By fusing multi-source information such as skin surface potential distribution, lesion geometric characteristics, tissue conductivity, historical treatment data and the like, systematic optimization is performed on electrode positioning in skin disease treatment, the effectiveness of electrode stimulation is maximized on the premise of ensuring the treatment safety, and the treatment efficiency is improved. An accurate electrode placement scheme is provided for treatment of different types of skin diseases, and the method is suitable for accurate determination of the electrode placement position in electrical stimulation treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent medical auxiliary technology, and particularly relates to a skin disease treatment electrode positioning system based on potential distribution. BACKGROUND

[0002] In the electric stimulation treatment of skin diseases, the rationality of the electrode placement position directly determines the treatment effect and safety. In current clinical practice, electrode positioning is mostly dependent on the experience of doctors, which has strong subjectivity and poor individual adaptability, and is prone to cause the electrode to be placed in an invalid area or a sensitive area, thereby causing low treatment efficiency, skin burns and other risks.

[0003] In the prior art, the traditional Canny algorithm mostly uses fixed parameters to process skin potential images, without considering the skin texture noise and individual differences in potential distribution, resulting in low accuracy of lesion area potential gradient edge recognition. The convex hull algorithm only pursues geometric minimum enclosure, which is prone to lesion coverage problems. The clustering analysis is mostly based on a single potential feature, and the clustering number depends on experience setting, which cannot adapt to the differences in potential distribution of different lesion types. The regression analysis and decision model do not fully integrate patient individual characteristics and clinical treatment rules, resulting in poor generalization of the prediction results. These technical defects make it difficult for the existing electrode positioning scheme to balance treatment effectiveness, safety and clinical practicability, and there is an urgent need for a multi-dimensional optimization of precise positioning method. SUMMARY

[0004] In view of the deficiencies of the existing methods and the needs of practical application, in order to solve the above problems. On the one hand, the present application provides a skin disease treatment electrode positioning system based on potential distribution, comprising: The system includes several modules: a module for identifying regions with dramatic potential gradients, used to accurately locate potential abrupt changes within the lesion area; a module for identifying the minimum bounding region of the lesion, used to extract the minimum bounding region of the lesion through convex hull expansion and lesion coverage verification; a module for grouping similar potential regions, used to divide the skin within the minimum bounding region into several similar potential groups based on features such as potential mean and variance; a module for predicting the optimal electrode position, used to predict the treatment effectiveness of each position within the minimum bounding region as an electrode placement point through data layering and feature engineering optimization, and to screen initial candidate positions; and a module for calculating the electric field distribution, used to accurately simulate the electric field generated by the initial candidate electrode positions within the skin tissue. The system comprises the following modules: an electric field distribution module; a sweat gland dense area identification module, used to extract the directional and coarse features of skin surface texture to identify densely populated sweat gland areas; an optimal electrode placement point selection module, used to find a set of placement points that balance optimal treatment effect and reasonable electrode distribution under the premise of satisfying all safety constraints through objective function quantification and constraint adaptation; an enhanced feature generation module, used to integrate similar potential region features and optimal electrode position features through feature selection and consistency verification to generate a more representative enhanced feature set; and a final electrode position determination module, used to transform enhanced features, electric field distribution, and optimal placement point features into decision rules through medical rule embedding and pruning optimization, ultimately outputting the electrode position.

[0005] This invention, through multi-module collaboration, precisely integrates multi-source information such as potential distribution and lesion characteristics, addressing the shortcomings of traditional electrode positioning methods that rely on experience and lack accuracy and safety. The potential gradient and lesion encirclement module enables precise lesion region localization, while similar potential grouping provides a basis for targeted therapy. The optimal location prediction and electric field simulation module, combining historical data and physical laws, enhances the scientific rigor of localization and reduces the risk of ineffective treatment. The sweat gland identification module avoids sensitive areas, ensuring treatment safety. The enhanced feature generation and final decision module, through feature integration and embedding of medical rules, outputs reliable electrode positions. This transforms localization from experience-driven to data-driven, improving treatment effectiveness and safety, adapting to different lesion types, reducing the workload of doctors, and providing precise and safe technical support for electrostimulation therapy of skin diseases.

[0006] Optionally, the precise location of potential abrupt change regions within the lesion area, and thus the identification of potential abrupt change regions within the lesion area, includes the following steps: The Sobel operator is used for gradient calculation; initial dual thresholds are calculated, and the edge pixels filtered by the gradient thresholds are connected using the 8-neighborhood connection rule to obtain the edge contour of the region with drastic potential gradient changes. This invention accurately solves the gradient using the Sobel operator to capture potential change features. Combined with initial dual threshold filtering and the 8-neighborhood connection rule, a complete edge contour is constructed, achieving precise locking of potential abrupt change regions within the lesion area. This provides core target points for subsequent lesion-surrounded region extraction and electrode position prediction, solving the problems of missed detections and contour breaks in traditional edge detection. It ensures that the localization focuses on the active core of the lesion, laying the foundation for improved treatment targeting and reduced ineffective stimulation. It is a key prerequisite for achieving precise electrode localization.

[0007] Optionally, the step of extracting the minimum bounding region of the lesion through convex hull expansion and lesion coverage verification includes the following steps: This invention introduces a distance-based deduplication mechanism to screen feature points; calculates the convex hull of the feature point set; and verifies the convex hull through full coverage, adaptively expanding the verification results to obtain the minimum enclosing region of the lesion. This invention improves the efficiency and accuracy of convex hull calculation by screening core feature points through distance deduplication; after constructing the convex hull based on the feature points, it ensures complete coverage of the lesion through full coverage verification and adaptive expansion. This effectively avoids the problem of missed lesion coverage caused by solely pursuing geometric minimization, providing a precise and comprehensive target range for grouping similar potential regions and predicting electrode positions, preventing missed lesion areas during treatment. Simultaneously, by standardizing region parameters, it lays a reliable geometric foundation for subsequent electric field simulation and other steps, serving as a key link between lesion identification and localization optimization.

[0008] Optionally, dividing the skin within the smallest enclosing region into several similar potential groups based on characteristics such as the mean and variance of the potential includes the following steps: The invention extracts a first feature vector within the smallest enclosing region; weights the first feature vector using entropy weighting; and groups similar potential regions using adaptive K-means clustering based on the number of clusters. This invention solves the problems of traditional clustering relying on single features and empirically based cluster numbers by extracting a multi-dimensional first feature vector, combining it with entropy weighting to highlight key feature weights, and then using adaptive K-means clustering to complete the grouping. This facilitates accurate classification of potential types in lesion areas, providing a targeted basis for subsequent electrode location prediction and ensuring that electrodes focus on highly active potential regions. Simultaneously, standardized region features provide a unified analysis unit for electric field simulation and multi-objective optimization, avoiding uneven treatment due to potential heterogeneity. This is a core supporting link for achieving personalized and precise electrode positioning.

[0009] Optionally, the step of predicting the treatment effectiveness of each location within the minimum enclosing region as an electrode placement point through data hierarchical and feature engineering optimization, and screening initial candidate locations, includes the following steps: By employing data hierarchical analysis and feature engineering optimization, a fusion feature set is constructed based on cluster group features. A patient-specific penalty term is introduced, with treatment effectiveness as the target variable. A random forest regression model is trained using this fusion feature set, and the trained model is used to select initial candidate locations. This invention, through data hierarchical analysis and feature engineering, fuses cluster group features to construct a comprehensive feature set, overcoming the limitation of single features in traditional models. The introduction of a patient-specific penalty term addresses the problem of poor generalization, making the random forest regression model more suitable for individual patient conditions. The model trained with treatment effectiveness as the target can accurately predict the efficacy of treatment at each location and select high-efficiency initial candidate locations. This provides high-quality basic samples for subsequent electrode positioning optimization, avoids interference from invalid locations, and improves the targeting of positioning. It is crucial for bridging regional grouping and precise positioning, laying a core data foundation for ensuring treatment effectiveness.

[0010] Optionally, the precise simulation of the electric field distribution generated by the initial candidate electrode locations within the skin tissue includes the following steps: Based on the layered structure of skin tissue, a finite element geometric model is constructed; a mesh is adaptively generated using tetrahedral elements; boundary conditions are set, and the electrostatic field control equations are solved to obtain the electric field distribution. This invention restores the physiological structure through layered skin modeling, and the tetrahedral adaptive mesh balances computational accuracy and efficiency. Combined with boundary conditions, the electric field distribution is solved, providing a core physical basis for electrode positioning. This facilitates the accurate quantification of the electric field characteristics of candidate locations, overcoming the deficiency of traditional positioning methods that lack empirical electric field measurements. Simultaneously, it provides electric field data support for multi-objective optimization, avoiding treatment ineffectiveness or skin burns caused by unreasonable electric field distribution. It is a crucial safety guarantee linking candidate location selection and final decision-making.

[0011] Optionally, the extraction of the directionality and coarseness features of skin surface texture to identify areas with dense sweat glands includes the following steps: This invention extracts gray-level co-occurrence matrix features; calculates the similarity of each feature by comparing it with a normal skin sweat gland texture template library to obtain sweat gland texture sensitive features; and identifies densely populated sweat gland areas based on these sensitive features. The invention accurately extracts skin texture features using a gray-level co-occurrence matrix and combines this with a normal skin sweat gland texture template library for comparison and screening of sensitive features, achieving precise identification of densely populated sweat gland areas. It can clearly identify sweat gland features and other skin textures, clearly define treatment contraindications, and provide a safe boundary for electrode positioning; it avoids stimulation reactions caused by placing electrodes in densely populated sweat gland areas, improving treatment safety; and it provides key constraints for subsequent multi-objective optimization and selection of electrode positions. It is a core link connecting texture analysis and safe positioning, laying the foundation for dual protection of treatment safety and efficacy.

[0012] Optionally, the step of finding a set of placement points that balances optimal treatment effect and reasonable electrode distribution by quantifying the objective function and adapting to constraints, while satisfying all safety constraints, includes the following steps: This invention constructs a core objective function, sets up a hierarchical constraint mechanism, and obtains the optimal set of electrode placement points using the NSGA-III algorithm. It quantifies the core requirements of efficacy, electric field, and distribution through the objective function, clarifies safety baselines and compatibility requirements using a hierarchical constraint mechanism, and efficiently solves for multi-objective optimal solutions using the NSGA-III algorithm. This effectively balances efficacy with safety and distribution rationality, avoiding treatment ineffectiveness or increased risk caused by a single-objective optimization. Simultaneously, it connects the candidate location and enhanced feature generation stages, providing a high-quality set of placement points that considers multi-dimensional needs for subsequent decision-making. This is key to achieving a unified approach to electrode positioning effectiveness, safety, and practicality, further improving the clinical adaptability and reliability of the solution.

[0013] Optionally, the step of integrating similar potential region features and optimal electrode location features through feature filtering and consistency verification to generate a more representative enhanced feature set includes the following steps: This invention constructs an association mapping between the optimal electrode placement point and similar potential groups based on spatial attribution; it performs feature set consistency verification using a historical database of similar cases as a reference; and it structurally splices standardized core features, local-global association features, and historical correction features to generate a final enhanced feature set with hierarchical feature expression capabilities. This invention establishes an intrinsic connection between electrode locations and similar potential regions through spatial association mapping, avoiding feature isolation; it improves feature reliability and corrects deviations caused by individual differences by combining consistency verification with historical cases; and it structurally splices multi-dimensional features to form a hierarchical expression, strengthening feature representativeness. Connecting the similar potential grouping and electrode location screening stages provides high-quality feature input for subsequent decision trees, which is the core support for improving electrode positioning accuracy and result stability, ensuring that the final positioning takes into account both data rationality and clinical suitability.

[0014] Optionally, the step of transforming enhancement features, electric field distribution, and optimal placement point features into decision rules through medical rule embedding and pruning optimization, and finally outputting the electrode position, includes the following steps: This invention constructs a dual-core training system driven by data and guided by rules; employs pessimistic pruning to eliminate overfitting risks; and obtains the final output electrode positions based on a three-level decision-making rule system of point-by-point judgment, quantity completion, and quality verification. Through this dual-core training system, the model considers both data patterns and medical safety rules, addressing the problem of insufficient clinical adaptability in purely data-driven models. Pessimistic pruning effectively eliminates overfitting, improving decision reliability. The three-level decision-making rule system, progressively advancing, ensures accurate electrode positions, compliant quantity, and quality. It transforms multi-source features into decision-making criteria that conform to clinical logic, avoiding ineffective or dangerous localization, providing a rigid guarantee for the final electrode position output, and improving treatment safety and effectiveness. Attached Figure Description

[0015] Figure 1 This is a framework diagram of a skin disease treatment electrode positioning system based on potential distribution, provided in an embodiment of the present invention. Detailed Implementation

[0016] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0017] Throughout this specification, references to an embodiment, example, or illustration mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, phrases appearing in various places throughout the specification, such as "in one embodiment," "in an embodiment," "an example," or "an illustration," do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in any suitable combination and / or sub-combination in one or more embodiments or examples. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0018] Please see Figure 1 To address the aforementioned problems, this invention provides a skin disease treatment electrode positioning system based on potential distribution, such as... Figure 1 As shown, it includes: The module for identifying regions with severe potential gradients is used to accurately locate regions of potential abrupt change within the lesion area, thereby identifying such regions.

[0019] Based on the traditional Canny edge detection algorithm, and considering the characteristics of large noise interference (such as skin texture and measurement error) and blurred gradient boundaries in skin potential distribution images, this algorithm is adaptively optimized through dynamic threshold adjustment and texture suppression mechanism. It retains the significant features of potential gradient changes while filtering out invalid edge information that is not related to the lesion, thereby achieving accurate localization of potential abrupt change regions within the lesion area.

[0020] Specifically, the method of accurately locating potential change regions within the lesion area and thereby identifying potential change regions within the lesion area includes the following steps: S11. Use the Sobel operator to solve for the gradient.

[0021] Considering that the natural skin texture (such as skin grooves and ridges) and random noise introduced by the measurement equipment in the skin potential distribution image can seriously interfere with the accurate identification of the potential gradient, Gaussian filtering is preferred for targeted noise reduction. This filter has linear smoothing characteristics, which can suppress noise while preserving edge details to the greatest extent and avoid the edge blurring problem caused by traditional mean filtering.

[0022] The filter kernel size is dynamically adapted based on the texture density of the lesion area: First, using the initial manually marked bounding box of the lesion area as the range, a 10×10 pixel sliding window is used to calculate the texture grayscale variance σ within each window (the larger the variance, the coarser the texture and the stronger the noise interference). When σ>30, a 5×5 kernel is selected to enhance noise reduction capability; when σ≤30, a 3×3 kernel is selected to reduce excessive smoothing of subtle gradient information. Through this dynamic adjustment strategy, the interference of skin texture and measurement errors on subsequent gradient detection can be effectively reduced, laying the foundation for accurate extraction of potential change information.

[0023] The gradient is a core indicator reflecting the drasticness of potential changes. The Sobel operator is used for gradient calculation, and through pixel-weighted operations within a 3×3 neighborhood, the response to local grayscale changes is effectively enhanced, and the computational efficiency is suitable for high-resolution skin images. Specifically, the x-direction (horizontal) operator Gx=[[-1,0,1],[-2,0,2],[-1,0,1]] and the y-direction (vertical) operator Gy=[[-1,-2,-1],[0,0,0],[1,2,1]] are convolved with the potential image to obtain the gradient components Gx and Gy in the two directions.

[0024] Compared to the limitations of traditional algorithms that rely solely on gradient magnitude, this invention adds a gradient direction consistency verification mechanism: First, curve fitting is performed on the edge pixels of the manually labeled bounding box using the least squares method to obtain the overall edge trend curve of the lesion area (e.g., eczema lesions often present as irregular curves, while vitiligo often presents as near-circular curves). The tangent direction of this curve at each pixel is calculated, and then the gradient direction of each pixel is solved. Only gradient values ​​with an angle of less than 30° between the gradient direction and the corresponding edge tangent direction are retained. Then, the gradient magnitude is calculated to complete the screening and enhancement of effective gradient information.

[0025] S12. Calculate the initial double thresholds, and use the 8-neighborhood connection rule to connect the edge pixels after the gradient threshold filtering to obtain the edge contour of the region with drastic change in potential gradient.

[0026] The Otsu method is used to automatically calculate the initial dual thresholds (high threshold TH, low threshold TL). By traversing all possible thresholds, the threshold combination that maximizes the variance between the foreground (edge) and background classes is found to achieve adaptive segmentation. However, there are significant individual differences and lesion type differences in skin potential distribution (e.g., infectious skin diseases often show high potential clustering, while allergic skin diseases have a more gradual potential distribution). Relying solely on the Otsu threshold is prone to edge missed detection or false detection. Therefore, the mean potential μ is introduced for secondary adjustment: First, the mean potential μ of all pixels in the entire lesion area is calculated (0-255 gray values ​​correspond to actual potential 0-100mV, μ>128 corresponds to actual potential >50mV, defined as a high potential region; μ≤128 corresponds to actual potential ≤50mV, defined as a low potential region). When μ > 128, gradient boundaries in high-potential regions tend to be more blurred, requiring a 15% increase in TH and a 10% increase in TL to enhance the sensitivity of edge extraction. When μ ≤ 128, low-potential regions are susceptible to noise interference, resulting in false edges, requiring a 10% decrease in TH and a 5% decrease in TL to suppress false edges. This dynamic adjustment can adapt to lesions with different potential distribution characteristics, significantly improving the accuracy of gradient boundary identification.

[0027] Furthermore, edge pixels after gradient thresholding are often discrete and need to be connected to form a complete edge contour. An 8-neighborhood connection rule (where pixels satisfying a threshold in the vertical, horizontal, and four diagonal directions are considered connected) is used, which more completely restores the irregular shape of lesion edges compared to 4-neighborhood connections, especially suitable for skin lesions with tortuous boundaries. After connection, some isolated edge segments formed by residual noise remain, requiring selection based on clinical practice: according to dermatological clinical data, the minimum diameter of most pathological skin lesions is not less than 1mm, while the image pixel resolution is 500dpi (i.e., 1mm corresponds to approximately 20 pixels). Therefore, isolated edge segments shorter than 20 pixels are directly deleted to exclude pseudo-edges unrelated to the lesion. The final output edge contour of the region with drastic potential gradient changes is presented as a coordinate set in a skin surface coordinate system (such as a two-dimensional rectangular coordinate system with the tip of the nose as the origin), with a coordinate accuracy to 0.01mm, providing high-precision input for feature point extraction in subsequent convex hull algorithms.

[0028] Regions with sharp potential gradients typically correspond to the boundary between diseased and normal tissues or the core area of ​​disease activity, serving as key targets for electrode stimulation. Accurate identification of these regions provides the core target range for subsequent electrode placement, preventing electrodes from being placed in ineffective areas with gentle gradients. This invention combines texture features with the mean potential value to dynamically adjust algorithm parameters and introduces gradient direction consistency verification, resolving the issue of potential gradient edges easily being confused with texture edges against complex skin backgrounds.

[0029] The lesion minimum bounding region identification module is used to extract the lesion minimum bounding region through convex hull expansion and lesion coverage verification.

[0030] Based on the Graham scanning convex hull algorithm, and considering that skin lesions are more likely to present irregular non-convex shapes, an adaptive optimization of convex hull expansion and lesion coverage verification is performed to construct a minimized candidate region for electrode placement while ensuring complete coverage of the lesion area (including areas with severe potential gradients), thereby reducing the scope of ineffective treatment.

[0031] Specifically, the extraction of the minimum bounding region of the lesion through convex hull expansion and lesion coverage verification includes the following steps: S21. Introduce a distance-based deduplication mechanism to filter feature points.

[0032] The quality of feature points directly determines the accuracy of convex hull calculation, so it is necessary to balance the integrity of the lesion region with the targeting of the core target. First, all boundary pixels are extracted from the pixel-level labeled data of the complete lesion region (these points constitute the physical outline of the lesion, ensuring that the enclosed area does not miss the main body of the lesion), and then merged with the edge coordinate set of the region with a strong potential gradient (these points correspond to the active core area of ​​the lesion and are the key target points for electrode stimulation).

[0033] To avoid redundant calculations due to duplicate points, a distance deduplication mechanism is introduced: the merged point set is traversed, and the Euclidean distance between any two points is calculated. If the distance is less than 5 pixels (considering the image resolution of 1 pixel = 0.05 mm, corresponding to a physical distance of 0.25 mm, this precision can filter out duplicate annotations caused by measurement errors), the gradient edge coordinate points are retained (prioritizing core target point information), and finally the initial feature point set P is obtained. This point set contains both the overall contour information of the lesion and highlights the core area of ​​potential change, providing high-quality input for subsequent convex hull calculation.

[0034] S22. Calculate the convex hull of the feature point set.

[0035] The Graham scan method is used to calculate the convex hull H0 of the feature point set P. The convex hull is efficiently constructed through polar angle sorting and stack operations, which can adapt to the more irregular point set distribution of skin lesions.

[0036] The specific steps are broken down as follows: ① Starting point selection: Traverse the point set P and select the point with the smallest y-coordinate as the starting point p0 (if there are multiple points with the same y-coordinate, select the point with the smallest x-coordinate). This point can be guaranteed to be a vertex of the convex hull, avoiding ambiguity in subsequent sorting. ② Polar Angle Sort: With p0 as the pole, calculate the polar angle (range 0°-360°) of the line connecting all other points to p0. Sort these points according to the polar angle from smallest to largest. If the polar angles are the same, sort them according to their distance from p0 from closest to farthest. This sorting can ensure that the scanning process traverses the point set in an orderly manner in either clockwise or counterclockwise direction. ③ Constructing the convex hull using stack operations: Initialize the stack and push p0 and the first two sorted points onto the stack. Then, add the remaining points to the stack sequentially. For each new point added, determine the direction of the broken line formed by the top three points of the stack by calculating the cross product. If the cross product result is ≤0 (indicating that the three points are clockwise or collinear, collinear points will lead to redundancy in the convex hull), pop the middle point from the top of the stack. Repeat this process until the top three points are counterclockwise, then push the new point onto the stack. After traversal, the remaining points in the stack constitute the vertex set of the initial convex hull H0, which is the smallest convex polygon containing all feature points.

[0037] S23. By performing a full-coverage verification of the convex hull, the verification results are adaptively expanded to obtain the minimum enclosing area of ​​the lesion.

[0038] The initial convex hull H0 is constructed based on feature points only. Incomplete feature point sampling may lead to missed coverage of lesion areas. Therefore, full coverage verification is required to ensure clinical treatment needs.

[0039] In the specific operation, firstly, all pixels in the complete lesion area (including all labeled pixels inside and at the boundary of the lesion) are traversed. The ray method is used to determine whether each pixel is inside the convex hull H0: a ray is emitted from the pixel in the positive x-axis direction, and the number of intersections between the ray and each side of the convex hull H0 is counted. If the number of intersections is odd, the point is inside; if it is even, the point is outside.

[0040] After the traversal is complete, count the number of uncovered lesion pixels and calculate their proportion of the total number of lesion pixels. , Directly reflects the coverage integrity of the convex hull, if If the electrode is too large, it may cause some lesion areas to fall outside the electrode stimulation range, affecting the treatment effect; if... If it is too small, it may over-expand the surrounding area, increasing the risk of ineffective stimulation.

[0041] Furthermore, based on clinical treatment experience in dermatology, Set an acceptable undercoverage threshold (this threshold ensures that more than 98% of the lesion area is covered, while avoiding excessive expansion of the surrounding area).

[0042] like Then the convex hull expansion process is initiated: ① Locating the uncovered areas: By using coordinate clustering, all uncovered lesion pixels are divided into several continuous regions, and the geometric center of each uncovered area is determined; ② Select the target convex hull edge: Calculate the vertical distance from the center of each missed coverage area to each edge of the convex hull H0, and select the convex hull edge with the smallest distance as the expansion object (this edge is the boundary closest to the missed coverage area, and the expansion efficiency is the highest). ③ Calculate the extension distance d: First, calculate the maximum vertical distance d1 from all pixels in the missed coverage area to the target convex hull edge (to ensure that all pixels in the area can be covered after the extension), and then add a 2mm safety margin (to consider the deformation of the skin during movement and the error of electrode placement, to avoid missed coverage in dynamic state), that is, d=d1+2mm. ④ Generate a new convex hull H1: Move the midpoint of the target convex hull edge a distance d along the normal vector direction of the edge (pointing to the side of the missing coverage area) to obtain a new vertex, replace the two endpoints of the original convex hull edge, and reconstruct the convex hull vertex set.

[0043] Repeat the coverage check-convex hull expansion process until... Finally, we obtain the convex hull H that balances minimization and full coverage, which is the minimum enclosing area of ​​the lesion.

[0044] Furthermore, the convex hull H in the image pixel coordinate system is converted to the physical coordinate system of the skin surface. Based on the acquisition parameters of the input image (high-resolution skin images typically use 1 pixel = 0.05 mm, this conversion is determined by the pixel size of the image sensor and the optical magnification), the coordinates of all vertices of the convex hull H are converted to millimeter-level physical coordinates using a coordinate conversion formula. Simultaneously, the core parameters for calculating the minimum bounding region are: the area bounded by the convex hull H is calculated using the shoelace formula, and the perimeter and roundness (roundness = 4π × area / perimeter) are calculated in conjunction with the convex hull vertex coordinates. 2 (This reflects the regularity of the shape of the lesion).

[0045] The final output includes: the physical coordinates of the vertices of the minimum bounding region (accurate to 0.01 mm) and the area (unit: cm²). 2 The parameters, such as perimeter (in cm) and roundness, provide spatial constraints for subsequent cluster analysis and help clinicians determine the number of electrodes (e.g., per 5 cm). 2 (Placing one electrode) provides direct evidence.

[0046] Determining the minimum enclosing area of ​​the lesion limits the core range for electrode placement, ensuring that electrode stimulation is concentrated within the lesion area while avoiding electrodes exceeding the skin's safe tolerance range (such as near sensitive areas like the eyes and mucous membranes), thus providing spatial constraints for electrode positioning. This invention introduces a coverage verification-dynamic expansion mechanism, transforming the complete coverage requirements of clinical treatment into quantified algorithmic constraints, and iteratively optimizing the balance between minimizing the enclosing area and ensuring complete lesion coverage.

[0047] The similar potential region grouping module is used to divide the skin within the smallest enclosing area into several similar potential groups based on characteristics such as the mean and variance of the potential.

[0048] Based on the K-means clustering algorithm, and considering the characteristics of local similarity and global heterogeneity in skin potential distribution, optimization is performed by weighting potential features and adaptively determining the number of clusters. According to features such as the mean and variance of the potential, the skin within the smallest enclosing region is divided into several similar potential groups, providing a basis for the grouping and placement of multiple electrodes.

[0049] Specifically, the step of dividing the skin within the minimum enclosing region into several similar potential groups based on characteristics such as the mean and variance of the potential includes the following steps: S31. Extract the first feature vector within the smallest enclosing region.

[0050] Within the minimum bounding area, a dual-mode sampling method with a fixed number of samples and a dynamic interval is used: when the area is ≥5cm². 2 At that time, 1000 sampling points were selected evenly; when the area of ​​the region is <5cm² 2At that time, sampling is performed at 5×5 pixel intervals (corresponding to a physical interval of 0.25mm, which avoids computational redundancy caused by overly dense sampling and ensures that no local potential changes are missed). Sampling points need to exclude areas with concentrated skin texture noise (by filtering through texture variance data, removing pixels with texture variance σ>30) to ensure that the sampling points are focused on the lesion-related areas.

[0051] Furthermore, the first feature vector is a three-dimensional vector, including the actual potential value V of the sampling point and the standard deviation of the potential within the 3×3 neighborhood of the sampling point. The tissue conductivity corresponding to this point (Based on pixel-level labeling of lesions—regions labeled as lesions use the average conductivity value for this type of skin disease, while regions labeled as normal use the baseline conductivity value for healthy skin. Conductivity data is sourced from a clinical measurement database, covering statistical results of epidermal / dermal conductivity from over 2000 samples.) After feature extraction, outlier processing is required between the actual potential value and the standard deviation of the potential at the sampling points. The 3σ criterion is used to remove extreme values ​​outside the range to avoid measurement errors affecting subsequent clustering.

[0052] S32. The first feature vector is weighted using the entropy weighting method.

[0053] The actual potential value V at the sampling point and the standard deviation of the potential in the 3×3 neighborhood of the sampling point are used. The tissue conductivity corresponding to this point The min-max method is used to map to the [0,1] interval to eliminate dimensional differences; For the potential value V, its standardized data is divided into 8 intervals, and the probability of each interval is calculated to obtain the information entropy. (The higher the entropy value, the higher the feature uncertainty, and the lower the weight should be.) The potential value V is the core basis for electrode stimulation, therefore the weighting , The maximum entropy value represents the feature, ensuring that V, which has a low entropy value and high determinism, receives a higher weight; Reflects local potential fluctuations, is related to lesion activity but has a lower priority than V, and a lower weight. ; It directly affects the electric field distribution and is crucial to treatment safety, therefore it is given a higher secondary weight. .

[0054] By using weighted summation to obtain the weighted feature vector of each sampling point, we can highlight the core features while also taking into account the impact of tissue characteristics on treatment, thus providing a foundation for accurate clustering.

[0055] S33. Based on the number of clusters, group similar potential regions using adaptive K-means clustering.

[0056] The rationality of the cluster number K directly determines the clinical value of the grouping. The K value is determined by a dual mechanism of elbow rule + potential characteristic correction.

[0057] First, apply the elbow rule: ① Calculate the sum of squared errors (SSE) within clusters from K=2 to K=10; ② Plot the K-SSE curve, and find the elbow corresponding to the sudden change in the rate of decrease of SSE η between adjacent K values ​​by calculating the rate of decrease η (e.g., when K increases from 3 to 4, η drops sharply from 0.3 to 0.1, then K0=3). Then, adjust K0 based on the potential distribution characteristics: Calculate the potential range ΔV = Vmax - Vmin within the smallest enclosed region (Vmax is the maximum potential within the region, and Vmin is the minimum potential). When ΔV > 50mV (significant potential difference, requiring more refined grouping to match different stimulation parameters), K = K0 + 1; when ΔV < 20mV (gradual potential distribution, excessive grouping has no clinical significance), K = K0 - 1; when ΔV is between 20-50mV, keep K = K0. For example, psoriasis lesions often show ΔV>50mV, requiring an increase in the number of clusters to distinguish between high-potential active areas and low-potential peripheral areas; while eczema lesions mostly have ΔV in the range of 20-50mV, and K0 can meet the grouping requirements, ensuring that the number of clusters is highly consistent with clinical treatment needs.

[0058] To address the problem of traditional K-means random initialization of cluster centers easily getting trapped in local optima, this invention implements two core optimizations. First, the initialization of cluster centers is optimized by prioritizing the selection of potential extreme points from the sampling points as initial centers, including the points with the maximum, minimum, and median values ​​of V within the region. If the K value is greater than 3, additional centers are selected. The largest possible points (ensuring the center covers the key potential area) are selected, and these points often correspond to the core or boundary of the lesion, making the cluster centers more clinically representative. Secondly, the iteration termination condition is optimized: traditional K-means uses the number of iterations or the change in center value as a condition. This step combines the potential measurement accuracy to set a termination threshold. After each iteration, the potential deviation between the new cluster center and the previous round's center is calculated. Iteration stops when the Δc of all cluster centers is less than 0.5mV (0.5mV is the minimum effective potential difference for clinical electrode stimulation; below this value, the center is stable, and further iteration is meaningless). During iteration, weighted Euclidean distance is used to calculate the similarity between samples and centers, ensuring that the similarity calculation matches the feature weights. After clustering is completed, complete information for each cluster group is output: including the physical coordinates of the region, the average potential within the group, the average conductivity, and the number of samples.

[0059] This invention incorporates tissue conductivity and local potential stability into the feature system, dynamically allocates weights using the entropy weight method, and optimizes the number of clusters by combining potential range, thus solving the problem of unstable clustering results caused by large differences in potential distribution among different lesion types.

[0060] The optimal electrode location prediction module is used to predict the treatment effectiveness of each location within the smallest enclosing area as an electrode placement point through data hierarchical and feature engineering optimization, and to screen initial candidate locations.

[0061] Based on the random forest regression algorithm, and considering the sample heterogeneity (different skin disease types and individual patient differences) in historical treatment data, a multi-source feature-treatment effect mapping model is constructed through data stratification and feature engineering optimization. This model predicts the treatment effectiveness of each location within the smallest enclosing region as an electrode placement point and screens initial candidate locations.

[0062] Specifically, the process of predicting the treatment effectiveness of each location within the minimum enclosing region as an electrode placement point through data hierarchical and feature engineering optimization, and screening initial candidate locations, includes the following steps: S41. Through data layering and feature engineering optimization, construct a fusion feature set based on cluster group features.

[0063] Based on the characteristics of each cluster group (mean potential, region area, etc.), the historical treatment database (including patient basic information, lesion type, potential distribution characteristics, electrode location coordinates, treatment effectiveness rate, and adverse reaction rate), and the current patient's lesion type and tissue conductivity data, the historical data are stratified according to lesion type (such as eczema, psoriasis, vitiligo), and a sample set S consistent with the current patient's lesion type is selected; the missing potential characteristics in S are filled with the median of the same type of samples, and abnormal samples with an adverse reaction rate >5% are deleted.

[0064] Furthermore, a fusion feature set is constructed based on the weighted first feature vector, lesion geometric features (perimeter / area ratio of the smallest enclosing region, roundness), patient individual features (age standardized value, skin thickness), and electrode position statistical features of historical similar cases (mean of the optimal position in the same cluster group).

[0065] S42. Introduce a patient individual difference penalty term, with the treatment effectiveness rate as the target variable, train a random forest regression model through the fused feature set, and use the trained regression model to screen initial candidate positions.

[0066] The sample set S was divided into a training set and a validation set in a 7:3 ratio. A random forest regression model was trained with the treatment effectiveness rate as the target variable. The parameters were optimized as follows: the number of decision trees was determined according to the sample size (10 trees for every 100 samples), and the maximum depth was limited to 10 (to avoid overfitting). A patient individual difference penalty term was introduced to reduce the weight of samples whose skin thickness difference from the current patient was >0.2 mm (weight = 1 - |current thickness - sample thickness| / 0.5).

[0067] The current patient's fusion features are input into the trained model to predict the treatment effectiveness of all sampling points within the minimum bounding area. The top 20% of effective points are selected as the initial electrode candidate location set, and duplicate locations with a distance of less than 5 mm are excluded.

[0068] This invention combines macroscopic lesion types with microscopic individual characteristics through an optimized mechanism of hierarchical data screening, individual difference penalty, and similar case feature fusion, effectively improving the generalization of historical data.

[0069] The electric field distribution calculation module is used to accurately simulate the electric field distribution generated in the skin tissue at the initial candidate electrode locations.

[0070] Based on the finite element method (FEM), and considering the layered structure (epidermis, dermis, and subcutaneous tissue) and conductivity differences of skin tissue, this study accurately simulates the electric field distribution generated by the initial candidate electrode locations within the skin tissue through adaptive mesh generation and boundary condition optimization, thus verifying the safety and effectiveness of the electrode locations.

[0071] Specifically, the precise simulation of the electric field distribution generated by the initial candidate electrode locations within the skin tissue includes the following steps: S51. Based on the layered structure of skin tissue, construct a finite element geometric model.

[0072] Based on the initial candidate electrode location set, skin layer structure data (thickness and conductivity of each layer, obtained through ultrasound measurement), electrode parameters (circular shape, 10mm diameter, conductive silicone material, stimulation voltage range 0-10V), and minimum enclosing region coordinates, a finite element geometric model is constructed.

[0073] First, the skin layer structure data is dynamically adjusted based on the current patient's ultrasound measurement results. The lesion area is scanned using 15MHz high-frequency ultrasound to obtain the actual thicknesses of the epidermis (0.12-0.18mm), dermis (1.5-2.5mm), and subcutaneous tissue (6-8mm). Conductivity parameters are taken from a clinical database of skin samples. The conductivity of the lesion area is multiplied by a correction factor of 1.3-1.5 based on the corresponding normal tissue (as edema or inflammation in the lesion tissue can increase conductivity). The finite element model is constructed using ANSYS 2023R1, which supports multiphysics coupling and has excellent mesh processing capabilities. The analysis domain is defined with the minimum bounding region as the core, extending outwards by 5mm to form a buffer zone (to avoid distortion in electric field calculations caused by boundary effects). The electrode model is constructed according to actual clinical parameters: a circular structure with a diameter of 10mm, a relative permittivity of 3.5 for the conductive silicone material, and a surface-to-surface contact constraint between the electrode and the skin to ensure realistic simulation of the fit. Finally, the initial candidate electrode positions were accurately implanted into the model, and the error of the electrode center coordinates was controlled within ±0.01mm, providing a high-fidelity geometric basis for subsequent electric field calculations.

[0074] S52. Adaptively mesh based on tetrahedral elements.

[0075] Mesh quality is key to the accuracy of electric field calculations. Tetrahedral elements are used for meshing, and a differentiated strategy of denser core regions and sparser outer regions is implemented.

[0076] Specifically, the electrode contact area (within a diameter of 10 mm) is the starting point of the electric field, with a large field strength gradient, and the unit size is set to 0.1 mm; Regions with dramatic potential gradients are the core of lesion activity, and a 0.1mm dense grid is also used to ensure the capture of details of electric field abrupt changes; The skin layering interface (epidermis-dermis, dermis-subcutaneous tissue) is prone to electric field refraction due to abrupt changes in conductivity. A transition grid (unit size 0.2mm) is set to avoid numerical oscillation. The normal tissue area (the area away from the lesion and the electrode) uses a sparse grid of 0.5mm. While ensuring the calculation accuracy, the total number of grids is controlled between 500,000 and 800,000, balancing the calculation efficiency and resource consumption.

[0077] After mesh generation, quality verification is required: Use the mesh quality assessment tool provided by ANSYS to ensure that the element distortion rate is ≤15% and the aspect ratio is ≤3. Unqualified meshes need to be regenerated to avoid distortion of calculation results due to mesh defects.

[0078] S53. Set boundary conditions and solve the electrostatic field control equations to obtain the electric field distribution.

[0079] Optimize boundary constraints by combining skin physiological characteristics with clinical treatment scenarios.

[0080] ① Skin surface boundary: The skin surface exchanges charges with the air, so it is set as a convective boundary. The electric field intensity attenuation coefficient is taken as 0.02 (this value is determined by the skin surface charge dissipation experiment and can accurately simulate the attenuation law of the electric field at the air-skin interface). ②Bottom boundary of subcutaneous tissue: set as ground boundary (potential V=0). Since the muscles and bones below the subcutaneous tissue have high electrical conductivity, they can be regarded as the infinitely far boundary where charge dissipates, which conforms to the physical law of electric field propagation. ③ Electrode-skin contact boundary: Introduce contact resistance (set to 1kΩ). This parameter is derived from clinical testing - through contact resistance testing between conductive silicone electrodes and different skin types, 1kΩ is the average value, to avoid the problem of overestimation of field strength caused by ignoring contact resistance; ④ Electrode Excitation Boundaries: A constant voltage of 5V is applied to the electrode surface. This value is a commonly used safe threshold in clinical practice for electrical stimulation therapy of skin diseases (below 3V is insufficient for therapeutic effect, and above 8V is prone to causing skin irritation). The voltage loading method is set to gradual loading (increasing from 0 to 5V within 1ms) to simulate the electrode activation process in actual treatment and avoid numerical shocks caused by instantaneous high voltage. After all boundary conditions are set, a consistency check must be performed to ensure that the electrode excitation direction, boundary constraint type, and electric field propagation physical laws match.

[0081] Furthermore, the finite element method is used to solve the electrostatic field control equations. Since the electrode stimulation frequency is 50-100Hz (belonging to a low-frequency electric field), the influence of displacement current can be neglected, and Maxwell's equations are simplified to... ( For electrical conductivity, (where the potential is denoted as ), this simplification reduces computational complexity while maintaining computational accuracy.

[0082] The solver used is ANSYS's PCG iterative solver. After solving, an electric field distribution cloud map is generated (colored according to the magnitude of the electric field, ranging from 0-200V / m). Based on this, three core features are extracted: ① Average electric field strength in the lesion area (reflecting the overall stimulation intensity, with a clinically optimal range of 80-120V / m); ② Standard deviation of electric field strength (reflecting the uniformity of the electric field distribution, with a standard deviation / average electric field strength ≤ 0.3 being ideal, avoiding insufficient or excessive local stimulation); ③ Maximum electric field strength (a core safety constraint indicator, strictly controlled within 200V / m—this threshold originates from the skin electrical stimulation safety standard; exceeding this value can lead to stratum corneum damage and the risk of burns). Simultaneously, a cross-sectional image of the electric field strength at 500μm below the electrode (superficial dermis, the main distribution area of ​​the lesion) needs to be output to visually demonstrate the penetration depth and distribution pattern of the electric field within the lesion tissue.

[0083] Electric field distribution is a direct physical manifestation of the therapeutic effect of electrodes. Finite element simulation can help identify the risk of skin burns caused by excessively high electric field strength in advance, while ensuring a uniform electric field distribution in the lesion area, providing a physical basis for selecting candidate electrode locations. This invention combines ultrasound-measured skin structure data to construct a layered model, uses adaptive meshing to divide the core area, and introduces contact resistance and convection boundary conditions to make the simulation results closer to clinical reality.

[0084] The sweat gland dense area recognition module is used to extract the directional and coarse features of skin surface texture to identify dense areas of sweat glands.

[0085] Based on gray-level co-occurrence matrix (GLCM) texture analysis, and considering the characteristics of fine skin sweat gland texture and susceptibility to lesions, this study optimizes the process by using texture feature screening and adaptive threshold adjustment. This process extracts the directional and coarse features of skin surface texture, identifies areas with dense sweat glands, and avoids interference from sweat conductivity caused by placing electrodes in these areas.

[0086] Specifically, the extraction of the directional and coarse features of skin surface texture to identify areas with dense sweat glands includes the following steps: S61. Extract the features of the gray-level co-occurrence matrix.

[0087] First, an RGB-to-grayscale conversion is performed using a weighted average method (grayscale value = 0.299×R + 0.587×G + 0.114×B). This aligns with the human eye's highest sensitivity to green and lowest sensitivity to blue, thus more accurately preserving the grayscale levels of skin texture. Next, adaptive histogram equalization (CLAHE) is applied, dividing the image into 8×8 pixel sub-blocks. Histograms are calculated individually for each sub-block, and contrast gain is limited. This enhances the brightness differences of sweat gland textures while preventing skin blemishes (such as pigmentation) from being overemphasized.

[0088] For linear noise such as hair, morphological opening operation (erosion followed by dilation) is used: a 3×3 rectangular structuring element is selected, which matches the diameter of clinically common hair (about 50-80μm, corresponding to 2-3 pixels in a 1000dpi image). The erosion operation can cut off the connectivity of hair, while the dilation operation restores the complete morphology of sweat gland texture.

[0089] After processing, quality verification is required. The signal-to-noise ratio (SNR) of the texture region is calculated, and only images with an SNR ≥ 15dB are retained for subsequent analysis to ensure that the preprocessing effect meets the requirements for feature extraction. Furthermore, gray-level co-occurrence matrix (GLCM) features are extracted. The GLCM is a core tool for capturing spatial correlations in texture, and its parameters need to be optimized to account for the subtle and weakly directional characteristics of sweat gland textures. The sampling method uses 8×8 pixel blocks sliding, which corresponds to the actual physical size of 0.2032mm × 0.2032mm (1000dpi = 0.0254mm / pixel). This size can fully encompass the texture features of a single sweat gland opening (the diameter of a sweat gland opening is approximately 0.1-0.3mm) while avoiding excessively large blocks that would cause feature blurring.

[0090] The setting of distance d=2 is based on experimental verification: when d=1, it is easily affected by pixel-level noise interference, and when d≥3, the adjacent correlation information of sweat gland texture will be lost. d=2 can achieve a balance between noise suppression and detail preservation; the angle covers 0° / 45° / 90° / 135° to ensure that the distribution characteristics of sweat gland texture without obvious directionality are fully captured.

[0091] Furthermore, the four extracted core features need to be quantified in conjunction with their clinical significance: ① Energy: Due to the high degree of texture repetition, the energy value in areas with dense sweat glands is usually ≥0.6; ②Entropy: The texture of sparse sweat gland areas is simple, with an entropy value ≤3.5, while the entropy value of dense areas is ≥4.2; ③ Contrast: The difference in grayscale between the sweat gland opening and the surrounding skin makes the contrast of dense areas ≥500; ④ Correlation: The spatial continuity of sweat gland texture makes the correlation ≥ 0.3.

[0092] Ultimately, each pixel block generates a 4×4=16-dimensional feature vector (4 angles × 4 types of features), providing multi-dimensional basis for subsequent classification.

[0093] S62. Compare the similarity of each feature with the normal skin sweat gland texture template library to obtain the sweat gland texture sensitive features.

[0094] By comparing the template library of normal skin sweat gland textures, the similarity of each feature is calculated (using cosine similarity), and the top 30% of features with the highest similarity are selected as sensitive features of sweat gland texture.

[0095] S63. Identify areas with dense sweat glands based on the sweat gland texture sensitivity features.

[0096] A sweat gland texture classifier is trained using a support vector machine (SVM) to classify sampled pixel blocks into two categories: dense sweat glands and sparse sweat glands.

[0097] Pixel blocks classified as having dense sweat glands were subjected to connectivity analysis, retaining those with an area greater than 1 mm. 2 The region is determined by combining the findings with human physiological patterns (e.g., high density of sweat glands in the armpit and groin areas). The coordinate range of the densely sweated area is then output and marked as the electrode placement taboo zone.

[0098] Sweat is conductive, and placing electrodes in areas with dense sweat glands can lead to disordered electric field distribution, affecting treatment efficacy and increasing the risk of skin irritation. Identifying this area can provide contraindications for electrode placement and improve treatment safety.

[0099] By constructing a dedicated texture template library, screening sweat gland sensitive features, and verifying them in conjunction with human physiological laws, the accuracy of micro-texture recognition under disease backgrounds can be effectively improved.

[0100] The optimal electrode placement point selection module is used to find a set of placement points that balance optimal treatment effect and reasonable electrode distribution by quantifying the objective function and adapting the constraints, while satisfying all safety constraints.

[0101] Based on the non-dominated sorting genetic algorithm (NSGA-Ⅲ), and considering the multi-objective conflict between treatment effect, safety constraints, and spatial distribution in electrode positioning, optimization is performed through objective function quantification and constraint condition adaptation. Under the premise of satisfying all safety constraints, a set of placement points that balances optimal treatment effect and reasonable electrode distribution is found.

[0102] Specifically, the process of finding a set of placement points that balance optimal treatment efficacy and reasonable electrode distribution through objective function quantification and constraint adaptation, while satisfying all safety constraints, includes the following steps: S71. Construct the core objective function.

[0103] Based on the core requirement of electrode positioning, three core objective functions are constructed, and the requirements are optimized using a maximization approach. The specific definitions are as follows: ①The treatment effect target F1 = the predicted value of the treatment effectiveness of the candidate position × the coverage of the lesion area, where the coverage of the lesion area is defined based on the effective range of electrode stimulation (5mm), that is, the proportion of the lesion area within 5mm of the center of the electrode to the lesion area of ​​the cluster group to which the electrode belongs. ② Rational goals in electric fields ,in, The average electric field strength in the lesion area. The clinically optimal electric field strength threshold for this lesion type (determined based on the 95% confidence interval of historical effective treatment cases). ③ The spatial distribution target F3 = 1 - (minimum distance between electrodes / average distance between electrodes) is used to quantify the uniformity of the spatial distribution of electrodes. The average distance between electrodes is calculated from the total area of ​​the lesion and the preset number of electrodes. The minimum distance between electrodes is the minimum distance between all electrode pairs.

[0104] S72. Set up a hierarchical constraint mechanism.

[0105] A hierarchical constraint mechanism is adopted, which uses hard constraints to ensure safety and soft constraints to improve efficacy. All constraints are transformed into mathematical inequalities that the algorithm can recognize, ensuring that the optimization results conform to clinical operating procedures and physiological characteristics. ① Hard constraints (unbreakable safety and boundary criteria): a. Spatial boundary constraints: The coordinates of the electrode center must satisfy the set of inequalities formed by the coordinates of the vertices of the minimum enclosing region, and the distance from the edge of the region must be ≥2mm; b. Constraints related to sweat glands: The distance between the electrode edge and the edge of the densely populated area of ​​sweat glands must be ≥3mm; c. Electric field safety constraints: The maximum electric field strength corresponding to the electrode is ≤200V / m; d. Quantity matching constraint: The number of electrodes must be strictly limited to every 5cm. 2 ±0.5cm 2One treatment is performed, calculated based on the lesion area (e.g., 4.8cm). 2 Place one, 7.2cm 2 Place two (to avoid insufficient treatment due to insufficient quantity or excessive burden on the patient due to excessive quantity).

[0106] ② Soft constraints (prioritizing and guiding principles to improve treatment efficacy): a. Target priority constraint: If the candidate location falls within a region of intense potential gradient, its therapeutic effect target F1 value will be multiplied by a weighting coefficient of 1.2; b. Individual adaptation constraint: For thin skin areas with a skin thickness of <1.5mm (such as the face and neck), the electrode spacing needs to be increased by 20% from the basic average spacing (lesion area / number of electrodes) to avoid the electric field penetrating too deeply and stimulating subcutaneous nerves. This constraint is achieved by dynamically adjusting the average spacing threshold. It is not forcibly excluded but will reduce the calculated value of the spatial distribution target F3.

[0107] Hard constraints are handled using a penalty function method (if violated, the individual's fitness is directly set to 0, and they are eliminated from the population). Soft constraints are incorporated into the objective function through weight adjustment to ensure that the optimization direction is consistent with the decision-making logic of prioritizing clinical safety and efficacy.

[0108] S73. Obtain the optimal set of electrode placement points according to the NSGA-Ⅲ algorithm.

[0109] ① Population Initialization: Based on the initial candidate site set, an elite retention + random supplementation strategy is used to construct the initial population. High-efficacy candidate sites with F1 ≥ 0.7 are prioritized for screening, and the remaining sites are randomly selected from those with F1 ≥ 0.5. The population size is set to 5 times the number of electrodes (if 3 electrodes are to be placed, the population size = 15). Individuals in the population are coded with electrode position coordinates (x, y), with the coding accuracy retained to 0.01 mm to match the clinical placement accuracy.

[0110] ② Genetic operation design: The selection operation adopts the tournament selection method, randomly selecting 3 individuals to compete, and selecting individuals with better non-dominance level and greater crowding to enter the next generation to ensure the inheritance of superior genes; the crossover operation adopts single-point crossover, randomly selecting crossover points based on the electrode coordinate encoding, and exchanging part of the genes of two parent individuals, with the crossover probability set to 0.8; the mutation operation adopts Gaussian mutation, adding Gaussian perturbations with a mean of 0 and a standard deviation of 0.5 mm to the electrode coordinates, with the mutation probability set to 0.05 to avoid excessive mutation destroying the already formed superior solutions.

[0111] ③ Iteration Termination and Solution Screening: Iteration termination adopts dual threshold control. The maximum number of iterations is set to 100 generations (after exceeding this number, the improvement of the solution is <0.5%, and continuing the iteration is meaningless), or the compromise solution in the Pareto optimal solution set remains unchanged for 10 consecutive generations (fitness fluctuation <0.01). After the iteration is completed, the final solution is screened from the Pareto optimal solution set based on clinical priority: First, ensure that F1≥0.75 (the bottom line of efficacy, corresponding to a clinical effectiveness rate ≥85%), then screen F2≥0.8 (the bottom line of electric field rationality, electric field deviation ≤20%), and finally select the solution with F3 closest to 0.2 (relatively uniform spatial distribution) as the optimal electrode placement point set. This screening logic is completely consistent with the decision order of clinical efficacy first, safety second, and operation third.

[0112] Furthermore, a dual verification mechanism of weighted comprehensive scoring and tiered clinical validation is constructed to ensure the reliability and practicality of the output scheme and avoid the disconnect between algorithm optimization and clinical practice. ① Comprehensive score calculation: Based on the weight allocation of clinical needs, the treatment effect target F1 accounts for 50% (core efficacy indicator), the electric field rationality target F2 accounts for 30% (safety key indicator), and the spatial distribution target F3 accounts for 20% (operational auxiliary indicator). The comprehensive score S = F1×0.5 + F2×0.3 + F3×0.2.

[0113] ②Graded verification process: a. Compliance verification: If S≥0.8, it is necessary to further verify the hard constraint satisfaction (whether the electrodes are all within the enclosed area, whether the maximum field strength is ≤200V / m, and whether the distance from the sweat gland area is ≥3mm). If all are satisfied, the solution is directly output, and the F1 and F2 values ​​of each electrode and its cluster group are marked. b. Fine-tuning verification: If S≥0.8 but there is a slight violation of soft constraints (such as one electrode not falling in the region of severe potential gradient), output the solution and indicate that it is recommended to adjust it to the XX coordinate (within the region of severe gradient). After adjustment, the F1 is expected to improve to XX. c. Handling of Non-Compliance: If S < 0.8, the optimization is deemed unsatisfactory, and candidate location prediction is re-executed. In this case, the feature engineering dimensions need to be expanded (e.g., adding lesion stage, patient skin moisture characteristics), or the parameters of the random forest regression model need to be adjusted (e.g., increasing the number of decision trees to 15 trees per 100 samples) to improve the quality of candidate locations before re-entering the multi-objective optimization process. After validation, a validation report needs to be generated, clearly indicating the scoring composition, constraint satisfaction status, and adjustment suggestions, providing clinicians with complete decision-making support.

[0114] By comprehensively balancing therapeutic efficacy, safety, and operational feasibility, the optimal electrode placement scheme is selected from initial candidate locations, resolving the problem of overlooking certain aspects in single-objective optimization and providing clear guidance for electrode placement in clinical practice. This invention transforms the qualitative requirements of clinical treatment into a quantitative objective function, introduces lesion-specific optimal electric field parameters as constraints, and finds a multi-objective compromise solution through the NSGA-Ⅲ algorithm.

[0115] The enhanced feature generation module is used to integrate similar potential region features and optimal electrode position features through feature filtering and consistency verification to generate a more representative enhanced feature set.

[0116] With feature-level data fusion as the core, and considering the redundancy and complementarity of multi-source features, optimization is achieved through feature selection and consistency verification. By integrating features of similar potential regions and the optimal electrode position features, a more representative enhanced feature set is generated, thereby improving the accuracy of subsequent decisions.

[0117] Specifically, the process of integrating similar potential region features and optimal electrode location features through feature filtering and consistency verification to generate a more representative enhanced feature set includes the following steps: S81. Construct an association mapping between the optimal electrode placement point and similar potential groups based on spatial attribution.

[0118] A two-stage standardization strategy, involving classification and redundancy removal, is employed to lay the data foundation for the effective fusion of multi-source features. First, differential standardization is performed based on feature attributes: ① Numerical features (such as average electric potential, electric field strength, etc.) are normalized using the min-max method to map the feature values ​​to the [0,1] interval.

[0119] ② For proportional features (such as treatment effectiveness rate, lesion coverage, etc.), since the original value range is already in the [0,1] interval, only precision calibration (retaining 3 decimal places) is required to ensure data format uniformity.

[0120] After standardization, feature redundancy screening was performed: using the Pearson correlation coefficient (r) as a quantitative evaluation index, the correlation of all pairwise feature combinations in the feature set was calculated, and features were judged to be highly redundant when |r|>0.8. Redundant feature processing followed clinical priority principles: for example, when there was redundancy in electrode spacing and distribution uniformity, electrode spacing that directly guided clinical operation was retained first; when there was redundancy in mean potential and potential deviation, potential deviation, which better reflected the characteristics of local lesions, was retained. The final redundant feature removal ratio was controlled within 15%, achieving a balance between feature dimensional simplification and core information retention.

[0121] Furthermore, a mapping between the optimal electrode placement point and a similar potential group is constructed based on spatial attribution. The similar potential group to which each electrode's position coordinates belong is determined using the ray casting method, clarifying the correspondence between a single electrode and the group potential region. For electrodes with confirmed attribution, their core features (treatment effectiveness, local electric field strength, tissue conductivity) are extracted. The deviation parameters (including absolute and relative deviations) between the core features of this location and the average features of the corresponding potential group are calculated, forming a correlation feature dimension between local features and the global benchmark, to quantify the difference between local electrode characteristics and the group potential pattern.

[0122] S82. Using a database of similar historical cases as a reference, conduct a consistency check of the feature set.

[0123] First, cases with a ≥85% match in lesion type and skin physiological parameters with the current patient are selected from the database to construct a similar case feature set. The cosine similarity between the current enhanced feature set and the similar case feature set is calculated. Simultaneously, the absolute deviation rate between the current predicted treatment effect and the actual treatment effect of similar cases is quantified. A consistency index β is calculated using the formula: cosine similarity of feature vectors × (1 - absolute deviation rate of treatment effect) to quantify the matching degree between the current feature set and historical effective cases. When β < 0.6, it is considered insufficiently consistent, and feature correction coefficients from historical similar cases (such as potential weight adjustment factors and electric field parameter correction values) need to be introduced. A weighted fusion strategy is then used to optimize the current feature set to ensure that the correlation between features and treatment effects conforms to clinical practice consensus.

[0124] S83. The standardized core features, local-global correlation features, and historical correction features are structurally spliced ​​together to generate the final enhanced feature set with hierarchical feature expression capabilities.

[0125] Standardized core features, local-global correlation features, and historical correction features are structurally concatenated to generate a final enhanced feature set with hierarchical feature representation capabilities. This feature set contains two dimensions: ① Electrode-level characteristics, covering the predicted treatment efficacy, local electric field parameters, tissue conductivity and characteristic deviation parameters of a single electrode, focusing on the effectiveness and specificity of a single electrode; ② Global-level features, including spatial distribution parameters of multiple electrodes (average spacing, minimum spacing, distribution uniformity), overall potential characteristics of the lesion area (potential range, average potential), and comprehensive efficacy prediction indicators, reflecting the overall performance and treatment coverage of the electrode system.

[0126] By introducing a local-global correlation and historical consistency verification mechanism, feature fusion is linked to clinical effects to ensure the effectiveness and reliability of enhanced features. Through the fusion and optimization of multi-source features, the limitations of single features are eliminated, and enhanced features that more comprehensively reflect the rationality of electrode placement are generated, providing more reliable data support for the final decision.

[0127] The final electrode location determination module is used to transform enhancement features, electric field distribution, and optimal placement point features into decision rules through medical rule embedding and pruning optimization, and finally outputs the electrode location.

[0128] Based on the decision tree algorithm, and considering the need for clear rule interpretation in clinical decision-making, this paper transforms enhanced features, electric field distribution, and optimal placement point features into interpretable decision rules through medical rule embedding and pruning optimization. Finally, it outputs electrode positions and operation suggestions that meet clinical needs.

[0129] Specifically, the process of transforming enhancement features, electric field distribution, and optimal placement point features into decision rules through medical rule embedding and pruning optimization, and finally outputting the electrode position, includes the following steps: S91. Construct a dual-core training system that is data-driven and rule-guided.

[0130] We construct a dual-core training system that combines data-driven and rule-guided approaches to ensure that the decision tree conforms to both historical patterns and clinical norms.

[0131] ① Refined Sample Library Construction: The historical treatment decision case library undergoes a triple screening process. Case lesion types must be consistent with the applicable scope (covering 8 common skin diseases such as eczema and psoriasis), containing complete feature-effect-decision records (features include 14-dimensional data such as electric potential and electric field; decisions include the doctor's final placement plan and follow-up feedback), excluding cases with extreme individual differences, and finally retaining valid cases as training samples. Sample labels employ a three-level clinical judgment labeling system: jointly evaluated by 2 attending physicians and 1 algorithm engineer. Recommended placement requires a treatment effectiveness rate ≥85% and no adverse reactions; cautious placement corresponds to an effectiveness rate of 70%-85% or a slight risk of irritation; prohibited placement is defined as an effectiveness rate <70% or a violation of safety constraints. Only cases with a Kappa value ≥0.85 are included in training.

[0132] ② Feature Input and Algorithm Configuration: The input features are defined as the final enhanced feature set (including 8-dimensional electrode level + 6-dimensional global level) and core electric field parameters, totaling 17-dimensional features; when using the C4.5 algorithm, the splitting attribute is selected based on the information gain ratio.

[0133] ③ Clinical rule embedding: Clinical contraindication rules are transformed into algorithmic prior conditions. A pre-judgment node is added before the decision tree is constructed. If a feature meets any of the following conditions: distance from the eye periphery < 5mm, location within a dense sweat gland area, and maximum field strength > 200V / m, it is directly judged as prohibited from placement and the branch splitting is terminated. For special scenarios such as thin-skinned areas of the face and joint mobility areas, feature weighting (such as increasing the weight of skin thickness features by 20%) guides the decision tree to prioritize individual suitability, ensuring that rule embedding does not destroy data regularity while enhancing clinical safety.

[0134] S92. Use pessimistic pruning to eliminate the risk of overfitting.

[0135] Pessimistic pruning (PEP) is used to eliminate the risk of overfitting, and rules with strong generalization ability are retained through iterative logic of error estimation and branch selection.

[0136] ① Refinement of the pruning process: First, calculate the empirical error rate of each branch of the original decision tree (error rate = number of incorrect samples / total number of samples in the branch). Then, assume that a certain branch is pruned to a leaf node, and calculate the pessimistic error rate after pruning based on the binomial distribution. If the pessimistic error rate after pruning is ≤ the original branch error rate, and the difference is ≥ 0.02, then pruning is performed.

[0137] ② Basis for setting the error threshold: The 5% error rate threshold is derived from clinical decision-making requirements. After 10-fold cross-validation, when the branch error rate is >5%, the misjudgment rate of the decision rule in new cases will rise sharply to more than 18%, while when it is ≤5%, the misjudgment rate can be controlled within 6%, which meets the clinical low-risk decision-making needs.

[0138] ③ Rule Set Transformation and Optimization: After pruning, the decision tree is transformed into an IF-THEN rule set along the path from the root node to the leaf node. Each rule must include three elements: condition set, conclusion, and confidence level. For example, a complete rule is: IF Treatment efficacy rate > 0.8 AND maximum field strength < 180V / m AND distance from dense sweat gland area ≥ 3mm AND skin thickness ≥ 1.5mm THEN Recommended placement (confidence level 0.92). Rule priority is ordered as follows: safety constraints > efficacy indicators > distribution characteristics. When multiple rules conflict, the rule containing the prohibition of placement safety constraint is executed first to ensure clinical safety.

[0139] S93. Based on the three-level decision-making rule of point-by-point judgment, quantity completion, and quality verification, the final output electrode position is obtained.

[0140] A three-level decision-making process of point-by-point judgment, quantity completion, and quality verification is constructed to ensure that the output position conforms to both algorithm optimization and clinical needs.

[0141] ① Point-by-point judgment logic: Input the 17-dimensional features (enhanced features + electric field parameters) of the current patient into the rule set one by one, and perform batch judgment on the optimal placement point. The judgment order follows the principle of safety first and efficacy second. First, check whether the prohibition rule is violated, then assess the confidence of the recommended placement, and finally mark the cautious placement category. Each judgment result needs to be associated with the corresponding rule number for easy traceability.

[0142] ②Comprehensive scoring supplementation mechanism: If the number of recommended placement points does not meet clinical requirements (e.g., 7cm) 2If a lesion requires two electrodes but only one is selected, then the electrode should be chosen from the cautious placement points based on a comprehensive score. The scoring formula is S = 0.4 × efficacy confidence + 0.3 × electric field compliance + 0.2 × safety distance + 0.1 × ease of operation. Efficacy confidence is taken from the predicted effectiveness value, and electric field compliance is... The safe distance is the minimum distance from the forbidden area, and the ease of operation is assigned a value based on whether it is located in the joint's active zone (0.5 for the active zone and 1 for the inactive zone).

[0143] ③ Supplementing quality verification: The supplemented cautious placement points must meet two additional conditions: the distance from the selected recommended placement points is ≥15mm (to avoid electric field superposition) and the comprehensive score is ≥0.7 (to ensure basic efficacy and safety). If the number requirement is still not met, the number of iterations of the NSGA-Ⅲ algorithm will be readjusted (increased to 120 generations) to re-optimize the candidate points and avoid affecting treatment coverage due to insufficient number.

[0144] Furthermore, the system employs a dual output format of structured data and visual guidance to ensure that clinicians can intuitively understand and accurately operate the system.

[0145] ① Core data output: The final electrode position is represented by coordinate pairs in the physical coordinate system of the skin surface (e.g., with the center of the lesion area as the origin), with an accuracy of 0.01mm. At the same time, the cluster group to which each position belongs is marked. The stimulation parameters are recommended based on the statistical results of the top 3 historical similar cases. The voltage range should match the calculated electric field value (e.g., when the average field strength = 100V / m, the recommended voltage is 4-6V). The frequency is set to 50-80Hz (this range ensures both therapeutic effect and avoids muscle tremors). The parameter deviation is controlled within ±10%.

[0146] ② Supplementary Clinical Guidelines: Precautions should be combined with the preceding steps to clarify the causal relationship. For example, avoid areas with dense sweat glands (coordinate range XX-XX): sweat in this area will cause the electric field distortion rate to be >15%, resulting in fluctuations in stimulation intensity. Keep facial electrodes more than 5mm away from the eye area: the skin thickness around the eyes is <0.8mm, and getting too close can easily cause corneal irritation. At the same time, a visual annotation diagram of the lesion area is attached, using different colors to mark the electrode position (green recommended, yellow with caution), contraindication areas (red), and potential gradient areas (blue gradient). The diagram annotation clearly explains the meaning of each mark and the operation priority.

[0147] ③ Additional traceability information: Output the rule matching records and feature consistency verification results of the decision tree judgment, providing data support for subsequent efficacy follow-up and treatment plan adjustment.

[0148] By embedding explicit clinical treatment rules into the decision tree construction process, and balancing the scientific validity and practicality of the rules through pruning optimization, complex multi-source data is transformed into clinically understandable decision rules. This ensures that the final electrode placement not only meets the optimization results but also satisfies clinical operating procedures, achieving a seamless connection between technical optimization and clinical practice.

[0149] In summary, this invention, through multi-module collaboration, precisely integrates multi-source information such as potential distribution and lesion characteristics, addressing the shortcomings of traditional electrode positioning, which relies on experience and suffers from insufficient accuracy and safety. The potential gradient and lesion encirclement module enables precise lesion region localization, while similar potential grouping provides a basis for targeted therapy. The optimal location prediction and electric field simulation module, combining historical data and physical laws, enhances the scientific accuracy of positioning and reduces the risk of ineffective treatment. The sweat gland identification module avoids sensitive areas, ensuring treatment safety. The enhanced feature generation and final decision module, through feature integration and embedding of medical rules, outputs reliable electrode positions. This transforms positioning from experience-driven to data-driven, improving treatment effectiveness and safety, adapting to different lesion types, reducing the workload of doctors, and providing precise and safe technical support for electrostimulation therapy of skin diseases.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A skin disease treatment electrode positioning system based on potential distribution, characterized in that, include: The module for identifying regions with severe potential gradients is used to accurately locate regions of potential abrupt change within the lesion area, thereby identifying such regions. The lesion minimum bounding region identification module is used to extract the lesion minimum bounding region through convex hull expansion and lesion coverage verification; The similar potential region grouping module is used to divide the skin within the smallest bounding area into several similar potential groups based on features such as the mean and variance of the potential. The optimal electrode location prediction module is used to predict the treatment effectiveness of each location within the minimum bounding area as an electrode placement point through data hierarchical and feature engineering optimization, and to screen initial candidate locations. The electric field distribution calculation module is used to accurately simulate the electric field distribution generated by the initial candidate electrode positions within the skin tissue; The sweat gland dense area recognition module is used to extract the directional and coarse features of skin surface texture to identify dense sweat gland areas; The optimal electrode placement point screening module is used to find a set of placement points that balance optimal treatment effect and reasonable electrode distribution by quantifying the objective function and adapting the constraint conditions, while satisfying all safety constraints. The enhanced feature generation module is used to integrate similar potential region features and optimal electrode position features through feature filtering and consistency verification to generate a more representative enhanced feature set. The final electrode location determination module is used to transform enhancement features, electric field distribution, and optimal placement point features into decision rules through medical rule embedding and pruning optimization, and finally outputs the electrode location.

2. The skin disease treatment electrode positioning system based on potential distribution according to claim 1, characterized in that, The method for accurately locating potential abrupt change regions within a lesion area, and thereby identifying such regions, includes the following steps: Gradient calculation is performed using the Sobel operator; Calculate the initial double thresholds, and use the 8-neighborhood connection rule to connect the edge pixels after gradient threshold filtering to obtain the edge contour of the region with drastic potential gradient change.

3. The skin disease treatment electrode positioning system based on potential distribution according to claim 1, characterized in that, The extraction of the minimum bounding region of the lesion through convex hull expansion and lesion coverage verification includes the following steps: A distance-based deduplication mechanism is introduced to filter feature points; Calculate the convex hull of the feature point set; By performing a full-coverage verification of the convex hull, the verification results are adaptively expanded to obtain the minimum enclosing area of ​​the lesion.

4. The skin disease treatment electrode positioning system based on potential distribution according to claim 1, characterized in that, The process of dividing the skin within the minimum bounding region into several similar potential groups based on characteristics such as the mean and variance of the potential includes the following steps: Extract the first feature vector within the smallest bounding region; The first feature vector is weighted using the entropy weighting method; Based on the number of clusters, adaptive K-means clustering is used to group regions with similar potentials.

5. The skin disease treatment electrode positioning system based on potential distribution according to claim 1, characterized in that, The process of predicting the treatment effectiveness of each location within the minimum bounding region as an electrode placement point through data hierarchical analysis and feature engineering optimization, and screening initial candidate locations, includes the following steps: By optimizing data hierarchical structure and feature engineering, a fusion feature set is constructed based on cluster group features. A patient individual difference penalty term is introduced, with the treatment effectiveness rate as the target variable. A random forest regression model is trained using the fused feature set, and the trained regression model is used to screen initial candidate positions.

6. The skin disease treatment electrode positioning system based on potential distribution according to claim 1, characterized in that, The precise simulation of the electric field distribution generated by the initial candidate electrode locations within the skin tissue includes the following steps: A finite element geometric model was constructed based on the layered structure of skin tissue. Mesh generation is adaptively performed based on tetrahedral elements; Set boundary conditions and solve the electrostatic field control equations to obtain the electric field distribution.

7. The skin disease treatment electrode positioning system based on potential distribution according to claim 1, characterized in that, The extraction of the directional and coarse features of skin surface texture to identify areas with dense sweat glands includes the following steps: Extract features from the gray-level co-occurrence matrix; The similarity of each feature is calculated by comparing it with a normal skin sweat gland texture template library to obtain sweat gland texture sensitive features; Based on the sweat gland texture sensitivity features, areas with dense sweat glands are identified.

8. The skin disease treatment electrode positioning system based on potential distribution according to claim 1, characterized in that, The process of finding a set of placement points that balance optimal treatment efficacy and reasonable electrode distribution by quantifying the objective function and adapting to constraints, while satisfying all safety constraints, includes the following steps: Construct the core objective function; Set up a hierarchical constraint mechanism; The optimal set of electrode placement points is obtained based on the NSGA-Ⅲ algorithm.

9. The skin disease treatment electrode positioning system based on potential distribution according to claim 1, characterized in that, The process of integrating similar potential region features and optimal electrode location features through feature filtering and consistency verification to generate a more representative enhanced feature set includes the following steps: Construct an association mapping between the optimal electrode placement point and similar potential groups based on spatial attribution; Using a database of similar historical cases as a reference, conduct feature set consistency checks; The standardized core features, local-global correlation features, and historical correction features are structurally spliced ​​together to generate the final enhanced feature set with hierarchical feature expression capabilities.

10. The skin disease treatment electrode positioning system based on potential distribution according to claim 1, characterized in that, The process of embedding and pruning medical rules to optimize the electrode placement involves transforming enhanced features, electric field distribution, and optimal placement point features into decision rules, ultimately outputting the electrode position. This includes the following steps: Construct a dual-core training system that combines data-driven and rule-guided approaches; A pessimistic pruning method is used to eliminate the risk of overfitting. The final output electrode position is obtained by using a three-level decision-making rule of point-by-point judgment, quantity completion, and quality verification.

Images