Intelligent regulation and control method based on radio frequency treatment comfort parameters

Abstract

The invention discloses an intelligent regulation and control method based on radio frequency treatment comfort parameters, and the method comprises the steps: collecting a video of a treated patient, and carrying out the parallel processing through employing two linear sparse self-encoding networks with complementary functions: a first network focuses on the extraction of contour information representing the posture change of the patient from the video, and generating a first matrix; the second network has a deeper level and stronger low signal-to-noise ratio robustness, focuses on extraction of fine facial expression change information, generation of a second matrix, fusion of contour information and expression change information and formation of video data, comprehensively judges real-time comfort information of the patient on the basis of the video data, and when the judged comfort information meets a preset condition, gives an alarm to the patient. The system automatically adjusts the output power of the radio frequency treatment equipment, the moving speed of the treatment head and other comfortable parameters. According to the invention, objective, real-time and multi-dimensional evaluation of the physiological state of the patient is realized, automatic and intelligent regulation and control of treatment parameters are driven, and the safety, accuracy and patient experience of radio frequency treatment are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and discloses an intelligent control method based on radiofrequency therapy comfort parameters. Background Technology

[0002] In the field of radiofrequency ablation, ensuring patient comfort while providing effective treatment is crucial for improving clinical experience and safety. Currently, treatment comfort primarily relies on patients' proactive subjective feedback or medical staff's experiential observation for rough assessment and manual parameter adjustment. This approach suffers from significant lag and subjectivity, making it difficult to achieve real-time, precise, and personalized control of treatment parameters, thus impacting treatment effectiveness and patient satisfaction.

[0003] Some existing assistive technologies attempt to indirectly assess discomfort by monitoring patients' physiological or behavioral signals. For example, traditional image processing or simple pattern recognition methods are used to analyze videos of the treatment area, aiming to extract information from changes in posture or facial expressions. However, these methods often face significant challenges: factors such as changes in lighting in the medical environment, equipment occlusion, and individual patient differences lead to low signal-to-noise ratios in the videos; single-dimensional feature analysis, such as focusing only on macroscopic contours or rough expressions, cannot comprehensively and stably capture complex physiological responses; and conventional algorithms lack sufficient robustness and discriminative power against the aforementioned noise and complex patterns, making them prone to misjudgment and thus difficult to use directly as a reliable basis for automated control.

[0004] Therefore, the current field of radiofrequency ablation urgently needs an intelligent method that can overcome environmental interference, comprehensively utilize multi-dimensional complementary visual information, and possess highly robust analytical capabilities to objectively, in real time, and accurately assess the patient's comfort level during treatment. Based on this assessment, the method can achieve automated and intelligent control of treatment parameters, thereby filling the gap in existing technologies for precise and comfortable treatment. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] A smart control method based on comfort parameters of radiofrequency therapy, comprising:

[0007] S1, capturing videos of patient treatment;

[0008] S2, Perform first linear sparse autoencoder network processing on the video of the treated patient to collect contour information;

[0009] S3, the video of the treated patient is processed by a second linear sparse autoencoder network to collect facial expression change information. The second linear sparse autoencoder network is deeper in level than the first linear sparse autoencoder network, and the signal-to-noise ratio of the second linear sparse autoencoder network is lower than that of the first linear sparse autoencoder network. The second linear sparse autoencoder network is complementary to the first linear sparse autoencoder network. The contour information and the facial expression change information are complementary information.

[0010] S4, fuse the contour information and the facial expression change information to collect video data, wherein the contour information is a first matrix representing contour information, the facial expression change information is a second matrix representing facial expression change information, and the video data includes contour information and facial expression change information;

[0011] S5. Based on the video data, determine the patient's comfort information;

[0012] S6, when the comfort information meets the preset conditions, adjust the radiofrequency treatment comfort parameters.

[0013] Furthermore, S1 specifically includes:

[0014] During radiofrequency treatment, a camera device is simultaneously activated, and the camera device is configured to capture a continuous sequence of images including the patient's treatment area and facial region;

[0015] Adjust the viewing angle and focal length of the camera device so that the continuous image sequence can simultaneously cover the body surface contour area and the facial expression area that reacts to the treatment;

[0016] The frame rate of the camera device is set to capture the dynamic process of the patient's physiological response;

[0017] The continuous image sequence is arranged and packaged in chronological order to generate the video of the treated patient.

[0018] Furthermore, S2 specifically includes:

[0019] Each frame of the video of the treated patient is input into the input layer of the first linear sparse autoencoder network;

[0020] The encoding layer of the first linear sparse autoencoder performs a convolution operation on the input image to extract primary image features and generate a first set of feature maps;

[0021] The hidden layer of the first linear sparse autoencoder network applies a first sparsity constraint to the first set of feature maps. The first sparsity constraint is used to suppress the activation of non-contour-related feature neurons and enhance the response of contour edge feature neurons.

[0022] The decoding layer of the first linear sparse autoencoder performs a deconvolution operation on the sparsified features to reconstruct an output image that mainly contains contour information.

[0023] The difference between the output image and the original input image is calculated, and continuous contour lines are identified from the difference using an edge detection algorithm;

[0024] The identified contour lines are vectorized and the geometric properties of each contour line are calculated, including line length, curvature distribution, and enclosed area.

[0025] Based on the time sequence, the vectorized contour lines and their geometric attributes corresponding to each frame of the image are arranged to form the first matrix;

[0026] Each row of the first matrix corresponds to a video frame, and each column corresponds to a feature value of a contour geometric attribute. The first matrix is ​​the contour information.

[0027] Furthermore, step S3 performs a second linear sparse autoencoder network process on the treated patient video to collect facial expression change information, specifically including:

[0028] From each frame of the video of the treated patient, locate and crop out a sub-image region containing the patient's face;

[0029] The facial sub-image region is input into the input layer of the second linear sparse autoencoder network;

[0030] The first-level encoding layer of the second linear sparse autoencoder network uses small-sized convolutional kernels to extract features from the input facial sub-image, obtaining a second set of feature maps containing subtle texture variations;

[0031] The second-level encoding layer of the second linear sparse autoencoder performs deeper convolution and pooling operations on the second set of feature maps to extract higher-order semantic features.

[0032] The hidden layer of the second linear sparse autoencoder network applies a second sparsity constraint to the higher-order semantic features. The strength of the second sparsity constraint is higher than that of the first sparsity constraint. This constraint is used to focus on specific movement patterns in facial muscle groups that are related to significant facial expression changes under conditions of low signal-to-noise ratio, while filtering out non-expression-related muscle tremors and image noise.

[0033] The decoding layer of the second linear sparse autoencoder network reconstructs the high-order semantic features with strong sparsity constraints into an expression change intensity distribution map through multiple upsampling and deconvolution operations.

[0034] In the facial expression intensity distribution map, multiple predefined facial action unit regions are divided, and each region is associated with a specific facial muscle group.

[0035] For each facial motion unit region, the intensity change process in consecutive video frames is analyzed, and temporal dynamic features are extracted. The temporal dynamic features include the peak value, mean value, frequency of change, and duration of intensity changes.

[0036] The temporal dynamic features of all facial motion unit regions are organized and encoded frame by frame to form a second matrix;

[0037] Each row of the second matrix corresponds to a video frame, and each column corresponds to the temporal dynamic feature value of a facial motion unit region. The second matrix represents the expression change information.

[0038] Furthermore, the process of fusing the contour information and the facial expression change information to collect video data specifically includes:

[0039] Receive the first matrix and the second matrix as input;

[0040] Establish a time synchronization relationship between the first matrix and the second matrix to ensure that contour features and facial features from the same point in time can be associated.

[0041] A feature fusion layer is constructed, which concatenates the contour feature vector and the expression feature vector at the same time point to generate an initial fused feature vector;

[0042] Calculate the mutual information between the contour feature vector and the expression feature vector within the historical time window, and dynamically adjust the weight coefficients of the two types of features during splicing based on the mutual information, so that the feature combination with strong information complementarity can obtain a higher fusion weight.

[0043] The weighted concatenation of the initial fused feature vector is input into a dimension reduction module, which extracts the main components through linear transformation to generate a dimension-reduced fused feature vector.

[0044] The fused feature vectors, after reducing the dimensionality of all time points, are arranged in chronological order to form a three-dimensional tensor;

[0045] The first dimension of the three-dimensional tensor represents the time series, the second dimension represents the various dimensions of the fused feature vector, and the third dimension is reserved as an extended dimension.

[0046] Add metadata to the three-dimensional tensor, the metadata including at least the start time of video capture, the total duration, and the original dimension information of the first matrix and the second matrix;

[0047] The video data is encapsulated into a three-dimensional tensor with accompanying metadata.

[0048] Furthermore, S5 specifically includes:

[0049] From the three-dimensional tensor of the video data, extract the fused feature vector for each moment in chronological order;

[0050] The fused feature vector at each time step is input into a predefined classification model;

[0051] The classification model maps the input feature vector to a multidimensional comfort space and divides the comfort space into at least three non-overlapping decision regions, each corresponding to a different comfort level.

[0052] The classification model outputs the comfort level at the current moment as a preliminary judgment result;

[0053] Based on the preliminary judgment results of multiple consecutive moments preceding the current moment, a time series consistency analysis is performed.

[0054] If the time-series consistency analysis detects a continuous and unidirectional trend in comfort level, a trend warning signal is generated.

[0055] Based on the preliminary judgment results at the current moment and the trend warning signal, the structured comfort information is generated;

[0056] The comfort information includes at least the current comfort level, trend warning status, and judgment confidence level.

[0057] Furthermore, S6 specifically includes:

[0058] Set multiple preset conditions associated with different comfort levels and trend warning states;

[0059] Monitor the comfort information and compare it with the preset conditions;

[0060] If the comfort information indicates that the current comfort level is the first level and there is no trend warning, then parameter adjustment will not be triggered;

[0061] If the comfort information indicates that the current comfort level has dropped to the second level, or triggers a warning of an increasing trend of discomfort, then the first preset condition is determined to be met.

[0062] When the first preset condition is met, a first adjustment instruction set is generated. The first adjustment instruction set is used to guide the radiofrequency therapy device to reduce the output power of radiofrequency energy and shorten the dwell time at the same treatment point.

[0063] If the comfort information indicates that the current comfort level has further decreased to the third level, then the second preset condition is determined to be met;

[0064] When the second preset condition is met, a second adjustment instruction set is generated. The second adjustment instruction set is used to guide the radiofrequency therapy device to pause the radiofrequency energy output and perform the retraction or movement operation of the treatment head.

[0065] The generated set of adjustment instructions is sent to the control unit of the radiofrequency therapy device to perform adjustments to the radiofrequency therapy comfort parameters.

[0066] Furthermore, the specific structure of the video data is as follows:

[0067] The video data is a data object with a logical hierarchy, the top of which contains a data header and a main data part.

[0068] The data header is used to store descriptive metadata, which includes a unique identifier for the video of the treated patient, the total number of frames in the video, the start and end timestamps of the acquisition, the camera device parameter identifier used for acquisition, and the identification code of the radiofrequency therapy device in the current treatment session.

[0069] The main data section is organized using a hierarchical index structure. The first layer is the time index layer, which is an ordered list. Each entry in the list corresponds to a specific point in time or time period and contains a pointer to the data block in the second layer.

[0070] The second layer of the main data portion is a feature data layer, which consists of multiple feature data blocks, each of which is associated with an entry in the time index layer.

[0071] Each feature data block further contains three logical partitions:

[0072] The first logical partition stores a subset of contour information extracted from the first matrix that corresponds to the current time point or time period. The subset includes a sequence of control point coordinates of vectorized contour lines and a set of derived geometric attributes calculated from the control point coordinate sequence.

[0073] The second logical partition stores a subset of facial expression change information extracted from the second matrix and corresponding to the current time point or time period. The subset includes the activation intensity value of each facial action unit region and a dynamic descriptor calculated based on the activation intensity of consecutive frames.

[0074] The third logical partition stores the fused feature data, which is the fused feature vector after the dimensionality reduction of the current time point or time period, processed by the feature fusion layer and the dimensionality reduction module.

[0075] The main data section also includes an auxiliary index layer, which is established based on the numerical range of contour geometric attributes or facial expression dynamic descriptors, and is used to support fast querying and retrieval of video data segments under specific feature conditions.

[0076] The video data organically integrates time, contours, expressions, and fusion features to form a self-contained composite data entity.

[0077] Furthermore, the method also includes treatment process recording and feedback steps:

[0078] The user interface of the radiofrequency therapy device dynamically displays a visual status indicator generated based on the aforementioned comfort information;

[0079] The visual status indicators include at least an icon reflecting the current comfort level and a simple chart reflecting recent trends in comfort.

[0080] Create and maintain a process log file bound to the current treatment session;

[0081] The process log file records the following events in chronological order: the initial settings of the radiofrequency therapy comfort parameters, each parameter adjustment event automatically triggered by the method and its details, and the sequence of changes in the comfort information over time.

[0082] When each parameter adjustment event is triggered, keyframe images are extracted from the current patient treatment video, and the keyframe images, along with the corresponding comfort information and adjustment instructions, are saved to the process log file.

[0083] At the end of this treatment session, a treatment process summary report is automatically generated based on the contents of the process log file.

[0084] Furthermore, the method adapts to different radiofrequency treatment application scenarios by configuring the contour region of interest of the first linear sparse autoencoder network and the facial action unit region of interest of the second linear sparse autoencoder network.

[0085] When applied to radiofrequency therapy targeting body parts, the first linear sparse autoencoder network is adjusted to focus on changes in the body surface contour of the treatment site and its surrounding area.

[0086] When applied to radiofrequency treatment targeting the face, the first linear sparse autoencoder network is adjusted to focus on the overall or specific facial contours, while the second linear sparse autoencoder network is enabled to perform fine analysis of facial expression changes.

[0087] When the treatment environment makes the patient's face invisible or the video quality is insufficient for expression analysis, the method described above mainly relies on the contour information for comfort judgment and parameter adjustment.

[0088] This invention discloses an intelligent control method based on comfort parameters in radiofrequency therapy. It involves acquiring video of the patient undergoing treatment and processing it in parallel using two complementary linear sparse autoencoder networks: the first network focuses on extracting contour information representing changes in the patient's posture from the video to generate a first matrix; the second network, with deeper layers and stronger low signal-to-noise ratio robustness, focuses on extracting subtle facial expression changes to generate a second matrix. This second matrix integrates contour and expression information to form video data, which is then used to comprehensively assess the patient's real-time comfort level. When the assessed comfort level meets preset conditions, the system automatically adjusts comfort parameters such as the output power of the radiofrequency therapy device and the speed of the treatment head movement. This invention achieves objective, real-time, and multi-dimensional assessment of the patient's physiological state, driving automated intelligent control of treatment parameters, significantly improving the safety, accuracy, and patient experience of radiofrequency therapy. Attached Figure Description

[0089] Figure 1 A flowchart illustrating the workflow of an intelligent control method based on radiofrequency therapy comfort parameters claimed in an embodiment of the present invention;

[0090] Figure 2 The second flowchart is shown for a method for intelligent control of comfort parameters in radiofrequency therapy as claimed in an embodiment of the present invention.

[0091] Figure 3 The third flowchart is a method for intelligent control of comfort parameters in radiofrequency therapy, as claimed in an embodiment of the present invention. Detailed Implementation

[0092] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0093] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of those features. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications in the embodiments of this application, such as up, down, left, right, front, back, etc., are only used to explain the relative positional relationships and movements between components in a specific orientation as shown in the accompanying drawings. If the specific orientation changes, the directional indications will change accordingly. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0094] References to embodiments herein mean that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0095] According to a first embodiment of the present invention, the present invention claims protection for an intelligent control method based on radiofrequency therapy comfort parameters, referring to... Figure 1 ,include:

[0096] S1, capturing videos of patient treatment;

[0097] S2, Perform first linear sparse autoencoder network processing on the video of the treated patient to collect contour information;

[0098] S3, the video of the treated patient is processed by a second linear sparse autoencoder network to collect facial expression change information. The second linear sparse autoencoder network is deeper in level than the first linear sparse autoencoder network, and the signal-to-noise ratio of the second linear sparse autoencoder network is lower than that of the first linear sparse autoencoder network. The second linear sparse autoencoder network is complementary to the first linear sparse autoencoder network. The contour information and the facial expression change information are complementary information.

[0099] S4, fuse the contour information and the facial expression change information to collect video data, wherein the contour information is a first matrix representing contour information, the facial expression change information is a second matrix representing facial expression change information, and the video data includes contour information and facial expression change information;

[0100] S5. Based on the video data, determine the patient's comfort information;

[0101] S6, when the comfort information meets the preset conditions, adjust the radiofrequency treatment comfort parameters.

[0102] In this embodiment, the complete application process of an intelligent control method based on radiofrequency therapy comfort parameters is described in a typical facial radiofrequency skin tightening treatment scenario.

[0103] Scene setup: The patient lies supine on the treatment bed, ready to receive radiofrequency skin tightening treatment targeting the lower face, including the jawline and double chin area. The treatment device integrates a visual analysis system.

[0104] Implementation steps: Acquiring video of the patient being treated: After treatment is started, the camera integrated into the device begins to work, continuously capturing a synchronous video stream that includes the treatment area on the lower face of the patient as well as their complete frontal face.

[0105] The first network processes the acquired contour information: The video stream is fed into a first linear sparse autoencoder network for processing in real time. This network is pre-designed to be adept at steadily extracting the body's edge contours from complex backgrounds. In this scenario, it ignores details such as skin texture and gloss, focusing instead on capturing the shape and position of the skin contour lines in treatment-related areas such as the jawline and the junction of the neck and jaw. After processing, the system obtains a series of quantified contour morphology data that changes over time, i.e., contour information.

[0106] The second network processes and collects facial expression change information: The same video stream is fed in parallel into a second linear sparse autoencoder network. This network is more complex in design, with deeper layers, and is more robust to the reduction in signal-to-noise ratio caused by changes in lighting and slight motion blur. It ignores large-scale contours and focuses on analyzing pixel-level textures and motion patterns in subtle facial areas such as the brows, corners of the eyes, and corners of the mouth to identify minute muscle contractions and relaxations. After processing, the system obtains a series of time-varying, quantified facial muscle activity data, i.e., facial expression change information. These two networks and the information they extract are functionally complementary, together forming a more comprehensive digital description of the patient's response.

[0107] The video data is fused by combining contour information (e.g., a first matrix describing the curvature of the mandible) and facial expression information (e.g., a second matrix describing the intensity of corrugator supercilii muscle activity) at the same moment. The fusion process is not a simple superposition, but rather establishes a correlation between the two types of information. For example, it identifies the co-occurrence pattern of contour information of sudden tightening of the mandibular muscles and the simultaneous slight frowning facial expression information. After fusion, a structured video data packet is generated, which contains the original contour and expression matrices, as well as the correlation between them.

[0108] To determine comfort level, the built-in assessment module analyzes the aforementioned video data. It identifies specific pattern combinations; for example, persistent jawline contraction accompanied by frequent squinting is categorized as a significant discomfort pattern, while a stable jawline with only occasional slight twitching of the corners of the mouth is categorized as a comfortable pattern. Based on this, the system outputs a structured comfort level report in real time, including the current level and its trend.

[0109] Adjusting radiofrequency therapy comfort parameters: The system is preset to automatically trigger adjustments when the comfort information shows moderate discomfort for a certain period of time. For example, under this condition, the system sends a command to the radiofrequency therapy device to reduce the current output radiofrequency energy intensity by one level and slightly increase the speed at which the treatment head moves across the skin surface. After adjustment, the system continues to monitor, forming an adaptive treatment based on visual feedback.

[0110] Furthermore, referring to Figure 2 S1 specifically includes:

[0111] During radiofrequency treatment, a camera device is simultaneously activated, and the camera device is configured to capture a continuous sequence of images including the patient's treatment area and facial region;

[0112] Adjust the viewing angle and focal length of the camera device so that the continuous image sequence can simultaneously cover the body surface contour area and the facial expression area that reacts to the treatment;

[0113] The frame rate of the camera device is set to capture the dynamic process of the patient's physiological response;

[0114] The continuous image sequence is arranged and packaged in chronological order to generate the video of the treated patient.

[0115] In this embodiment, the camera device is activated and configured. When the therapist starts the treatment program on the operation panel, the video acquisition module is activated simultaneously. In this embodiment, a high-definition camera with zoom and wide-angle functions is used. The camera is physically mounted on the treatment arm and located behind the treatment head.

[0116] Adjusting the viewing angle and focus, the camera automatically adjusts to the lower face of the treatment area. First, the viewing angle is adjusted to simultaneously frame the area from the patient's chin to the center of the eyebrows. Then, automatic focusing ensures that the skin of the jaw in the treatment area and the eyebrows, eyes, and mouth in the expression observation area are clearly imaged. This setting ensures that a single camera can cover both the contour and expression information sources.

[0117] Setting the capture frame rate: Considering that facial expressions can change very rapidly, such as the instantaneous expression caused by pain, the system sets the capture frame rate to a high value to ensure that these fast dynamic processes can be captured and to avoid losing key reaction information frames due to a low sampling rate.

[0118] The process generates a video of the patient's treatment. The acquired continuous image sequence is buffered and encapsulated according to a strict timestamp order, with each frame marked with a precise timecode relative to the start of treatment. Finally, these timecode-encoded frames are encoded into a coherent video file, i.e., the patient's treatment video, for use by subsequent analysis modules. The entire acquisition process is completed silently in the background, without affecting the foreground treatment operations.

[0119] Furthermore, referring to Figure 3 S2 specifically includes:

[0120] Each frame of the video of the treated patient is input into the input layer of the first linear sparse autoencoder network;

[0121] The encoding layer of the first linear sparse autoencoder performs a convolution operation on the input image to extract primary image features and generate a first set of feature maps;

[0122] The hidden layer of the first linear sparse autoencoder network applies a first sparsity constraint to the first set of feature maps. The first sparsity constraint is used to suppress the activation of non-contour-related feature neurons and enhance the response of contour edge feature neurons.

[0123] The decoding layer of the first linear sparse autoencoder performs a deconvolution operation on the sparsified features to reconstruct an output image that mainly contains contour information.

[0124] The difference between the output image and the original input image is calculated, and continuous contour lines are identified from the difference using an edge detection algorithm;

[0125] The identified contour lines are vectorized and the geometric properties of each contour line are calculated, including line length, curvature distribution, and enclosed area.

[0126] Based on the time sequence, the vectorized contour lines and their geometric attributes corresponding to each frame of the image are arranged to form the first matrix;

[0127] Each row of the first matrix corresponds to a video frame, and each column corresponds to a feature value of a contour geometric attribute. The first matrix is ​​the contour information.

[0128] In this embodiment, each frame of the image is extracted sequentially from the video of the treated patient and fed into the network input layer. The images are first converted to grayscale to simplify processing and then normalized to a fixed size.

[0129] The encoding layer extracts primary features. The encoding layer after the input layer contains multiple convolutional filters that scan the entire image and respond to edges and lines in different directions. For example, some filters are sensitive to horizontal edges, which helps to capture the horizontal part of the jawline; others are sensitive to oblique or vertical edges, which helps to capture the contour of the side of the neck. The output of this layer is a collection of multiple primary edge features, called the first set of feature maps.

[0130] The hidden layer is subject to sparsity constraints, which is the core of the network. The hidden layer imposes sparsity constraints on a large number of feature maps from the encoding layer. This means that the network is trained before application to activate only a few feature neurons that are most relevant to the skin contour of the treatment area, while strongly suppressing neurons that respond to irrelevant textures such as skin spots, hair, and background fabric wrinkles. In this embodiment, this constraint ensures that the network will not misclassify tiny moles or hairs on the patient's skin as contour edges.

[0131] The decoding layer reconstructs the contour image. The highly refined features retained after sparse filtering are fed into the decoding layer. The decoding layer's job is to try to reconstruct an image using these features that only represent the contour. Ideally, this reconstructed image should only contain clear, continuous lines, i.e., the skin contour of the patient's treatment area, while missing original details such as skin texture.

[0132] The system calculates the difference between the original input frame and the pure contour frame reconstructed by the network. Since the network only reconstructs the contour, the region where the contour line itself is located in the difference map will also have significant difference values ​​due to imperfect reconstruction. By applying edge detection algorithms such as finding the maximum value in the gradient direction to the difference map, the system can accurately locate continuous and smooth contour lines.

[0133] Vectorization and geometric attribute calculation: The identified pixel-level contour lines are converted into vector graphics such as Bézier curves defined by a series of control points. The system calculates the geometric attributes of each vector contour: length (total length of contour line segments), curvature distribution (changes in the curvature of the contour, smooth or abrupt turns), and enclosing area (area of ​​the region enclosed by the contour). For example, when a patient clenches their teeth due to discomfort, the mandibular contour line may become shorter and straighter, the curvature distribution may have more abrupt changes, and the local enclosing area may shrink.

[0134] The system constructs the first matrix. For each frame of the video, it generates a feature vector containing the aforementioned geometric attribute values. Arranging these feature vectors from all frames in chronological order, the first matrix is ​​formed. The rows of this matrix represent the time frame sequence, and the columns represent various contour geometric attributes. This first matrix is ​​the final, purely quantized contour information extracted by the system, after removing redundant visual information.

[0135] Furthermore, step S3 performs a second linear sparse autoencoder network process on the treated patient video to collect facial expression change information, specifically including:

[0136] From each frame of the video of the treated patient, locate and crop out a sub-image region containing the patient's face;

[0137] The facial sub-image region is input into the input layer of the second linear sparse autoencoder network;

[0138] The first-level encoding layer of the second linear sparse autoencoder network uses small-sized convolutional kernels to extract features from the input facial sub-image, obtaining a second set of feature maps containing subtle texture variations;

[0139] The second-level encoding layer of the second linear sparse autoencoder performs deeper convolution and pooling operations on the second set of feature maps to extract higher-order semantic features.

[0140] The hidden layer of the second linear sparse autoencoder network applies a second sparsity constraint to the higher-order semantic features. The strength of the second sparsity constraint is higher than that of the first sparsity constraint. This constraint is used to focus on specific movement patterns in facial muscle groups that are related to significant facial expression changes under conditions of low signal-to-noise ratio, while filtering out non-expression-related muscle tremors and image noise.

[0141] The decoding layer of the second linear sparse autoencoder network reconstructs the high-order semantic features with strong sparsity constraints into an expression change intensity distribution map through multiple upsampling and deconvolution operations.

[0142] In the facial expression intensity distribution map, multiple predefined facial action unit regions are divided, and each region is associated with a specific facial muscle group.

[0143] For each facial motion unit region, the intensity change process in consecutive video frames is analyzed, and temporal dynamic features are extracted. The temporal dynamic features include the peak value, mean value, frequency of change, and duration of intensity changes.

[0144] The temporal dynamic features of all facial motion unit regions are organized and encoded frame by frame to form a second matrix;

[0145] Each row of the second matrix corresponds to a video frame, and each column corresponds to the temporal dynamic feature value of a facial motion unit region. The second matrix represents the expression change information.

[0146] In this embodiment, face localization and cropping first involves face detection in each frame of video, accurately locating and cropping the facial region, and eliminating interference from hair, ears, and background to obtain a standardized facial sub-image.

[0147] The first level of encoding captures subtle textures. Facial sub-images are fed into the network's second-level encoding layer, which uses very small convolutional kernels to scan tiny areas of the skin like a microscope. This layer can capture subtle changes in skin texture caused by muscle contractions, such as the fine wrinkles on the skin around the eyebrows caused by the contraction of the corrugator supercilii muscle, or the deepening of fine lines at the corners of the eyes caused by squinting. The output is a second set of feature maps containing a large number of underlying texture features.

[0148] The second level of encoding extracts higher-order semantic features, and the aforementioned texture features are fed into a deeper encoding layer. This layer maps the underlying texture changes into higher-order, physiologically meaningful semantic features through more complex combinations and abstractions. For example, it identifies specific texture change patterns in the glabella region as corrugator supercilii muscle activity, and texture changes on the outer corner of the eye as orbicularis oculi muscle outer canthus activity.

[0149] The hidden layers undergo strong sparsity and noise reduction processing. The hidden layers of this network are subject to stronger sparsity constraints than the first network and integrate noise reduction mechanisms. The aim is to stably extract only persistent and significant facial muscle movement patterns related to treatment-induced pain or discomfort, even amidst noise caused by uneven lighting, slight tremors, or fleeting facial expressions such as normal blinking. For example, it can effectively distinguish between transient, filterable noise such as momentary squinting caused by strong light stimulation and persistent, attention-grabbing signals such as persistent pain squinting caused by overheating from radiofrequency energy.

[0150] The decoding process generates an intensity map of facial expression changes. After strong filtering, the high-order semantic features are upsampled through the decoding layer to reconstruct an intensity distribution map of facial expression changes. This map is not an ordinary image, but a heatmap. The brightness value of each pixel in the map represents the activity intensity of facial muscles at that location in recent frames. Pain-related areas, such as furrowed brows and taut lips, will be highlighted.

[0151] Facial movement units were segmented and analyzed. Based on facial movement coding theory, the heatmap was divided into predefined facial movement unit regions, such as the brow depression area, eyelid tightening area, and upper lip lifting area. For each region, the system analyzed the change curve of its heat value within a time window to extract temporal dynamic characteristics: peak pain response intensity, mean tension level, frequency of muscle twitching, and whether the sustained tension state was instantaneous or persistent.

[0152] A second matrix is ​​constructed. For each frame, the temporal dynamic feature values ​​of all facial motion units are organized into a feature vector. The feature vectors of all frames are arranged in chronological order to obtain the second matrix. The rows of this matrix represent time, and the columns represent the various dynamic features of each facial motion unit. This second matrix is ​​the system-extracted, interference-resistant, and quantified information on facial expression changes.

[0153] Furthermore, the process of fusing the contour information and the facial expression change information to collect video data specifically includes:

[0154] Receive the first matrix and the second matrix as input;

[0155] Establish a time synchronization relationship between the first matrix and the second matrix to ensure that contour features and facial features from the same point in time can be associated.

[0156] A feature fusion layer is constructed, which concatenates the contour feature vector and the expression feature vector at the same time point to generate an initial fused feature vector;

[0157] Calculate the mutual information between the contour feature vector and the expression feature vector within the historical time window, and dynamically adjust the weight coefficients of the two types of features during splicing based on the mutual information, so that the feature combination with strong information complementarity can obtain a higher fusion weight.

[0158] The weighted concatenation of the initial fused feature vector is input into a dimension reduction module, which extracts the main components through linear transformation to generate a dimension-reduced fused feature vector.

[0159] The fused feature vectors, after reducing the dimensionality of all time points, are arranged in chronological order to form a three-dimensional tensor;

[0160] The first dimension of the three-dimensional tensor represents the time series, the second dimension represents the various dimensions of the fused feature vector, and the third dimension is reserved as an extended dimension.

[0161] Add metadata to the three-dimensional tensor, the metadata including at least the start time of video capture, the total duration, and the original dimension information of the first matrix and the second matrix;

[0162] The video data is encapsulated into a three-dimensional tensor with accompanying metadata.

[0163] In this embodiment, the input is time-synchronized, receiving first matrix contour information and second matrix facial expression change information from the previous process. The primary task is to ensure that the two matrices are perfectly aligned on the timeline. The system utilizes a unified timestamp embedded during video capture to precisely map the Nth row of the first matrix to the Nth row of the second matrix at the same millisecond.

[0164] Feature splicing and dynamic weighting: For each aligned time point t, the system splices the contour feature vector in the first matrix and the facial expression feature vector in the second matrix at time t to form a longer initial fused feature vector. To reflect complementarity, it does not simply splice, but introduces a dynamic weighting mechanism. This mechanism analyzes the correlation strength and mutual information between changes in contour features and changes in facial expression features in the recent time period. If a strong correlation is found, for example, each time the jaw tightens, it is accompanied by frowning, then the fusion weight of these two types of features is appropriately increased during splicing, because they mutually confirm the same physiological response. Conversely, if a certain type of feature changes in isolation, the weight may be reduced.

[0165] Dimensionality reduction and principal component extraction result in a high-dimensional initial fused feature vector after concatenation. The system uses dimensionality reduction methods such as principal component analysis to extract the principal components that best represent the variations in the original data, generating a fused feature vector with lower dimension but more concentrated information.

[0166] A three-dimensional tensor is constructed by stacking the dimensionality-reduced fused feature vectors of each time point in the treatment process in chronological order. This tensor can be imagined as a data cube: the first dimension, length, is the timeline; the second dimension, width, is the dimensions of the fused feature vectors at each time point; and the third dimension, height, can be reserved for future expansion, such as adding data channels for other modalities.

[0167] Metadata and encapsulation are added. A data header is attached to this 3D tensor to record its birth certificate, including the start and end times of the original video, the total duration, the original size of the first and second matrices, the dimensions after dimensionality reduction, and other metadata. Finally, this 3D tensor structure with attached metadata is packaged as a whole as the final, structured video data output, providing a standardized input format for subsequent comfort judgment.

[0168] Furthermore, S5 specifically includes:

[0169] From the three-dimensional tensor of the video data, extract the fused feature vector for each moment in chronological order;

[0170] The fused feature vector at each time step is input into a predefined classification model;

[0171] The classification model maps the input feature vector to a multidimensional comfort space and divides the comfort space into at least three non-overlapping decision regions, each corresponding to a different comfort level.

[0172] The classification model outputs the comfort level at the current moment as a preliminary judgment result;

[0173] Based on the preliminary judgment results of multiple consecutive moments preceding the current moment, a time series consistency analysis is performed.

[0174] If the time-series consistency analysis detects a continuous and unidirectional trend in comfort level, a trend warning signal is generated.

[0175] Based on the preliminary judgment results at the current moment and the trend warning signal, the structured comfort information is generated;

[0176] The comfort information includes at least the current comfort level, trend warning status, and judgment confidence level.

[0177] In this embodiment, the fusion feature vector is extracted by sequentially reading the fusion feature vector corresponding to each moment along the time dimension from the encapsulated three-dimensional tensor of the video data.

[0178] The classification model maps each fused feature vector into a pre-trained classification model in real time. This model internally defines a comfort space, which may be spanned by multiple feature dimensions. The model maps the input vector into this space and, based on its position, classifies it into a predefined category, such as comfortable, mild discomfort, or significant discomfort. This is the initial assessment result.

[0179] Temporal consistency analysis does not view single-frame judgments in isolation. It maintains a short-term historical buffer and analyzes the continuous changes in comfort level over the most recent few seconds. If a clear trend is detected in the level continuously decreasing from comfortable to mild discomfort and continuing to develop towards significant discomfort, the system will generate a trend warning signal. If an intensifying trend of discomfort is detected, this avoids decision fluctuations caused by misjudgment of a single frame or instantaneous reaction.

[0180] Structured comfort information is generated by combining the preliminary judgment results of the current frame (e.g., mild discomfort) with trend warning signals generated by time-series analysis (e.g., no warning or worsening trend), and attaching a score reflecting the confidence level of this judgment based on the distance from the feature vector's position in the classification space to the decision boundary. This is then packaged into a structured comfort information report. This report is the core basis for subsequent parameter adjustment decisions.

[0181] Furthermore, S6 specifically includes:

[0182] Set multiple preset conditions associated with different comfort levels and trend warning states;

[0183] Monitor the comfort information and compare it with the preset conditions;

[0184] If the comfort information indicates that the current comfort level is the first level and there is no trend warning, then parameter adjustment will not be triggered;

[0185] If the comfort information indicates that the current comfort level has dropped to the second level, or triggers a warning of an increasing trend of discomfort, then the first preset condition is determined to be met.

[0186] When the first preset condition is met, a first adjustment instruction set is generated. The first adjustment instruction set is used to guide the radiofrequency therapy device to reduce the output power of radiofrequency energy and shorten the dwell time at the same treatment point.

[0187] If the comfort information indicates that the current comfort level has further decreased to the third level, then the second preset condition is determined to be met;

[0188] When the second preset condition is met, a second adjustment instruction set is generated. The second adjustment instruction set is used to guide the radiofrequency therapy device to pause the radiofrequency energy output and perform the retraction or movement operation of the treatment head.

[0189] The generated set of adjustment instructions is sent to the control unit of the radiofrequency therapy device to perform adjustments to the radiofrequency therapy comfort parameters.

[0190] In this embodiment, preset conditions are set: multiple trigger conditions are preset in the system policy library. For example:

[0191] Condition A: The comfort level remains comfortable for a certain period of time without any warning.

[0192] Condition B: Warning that the comfort level has changed to mild discomfort, or that there is a trend of increasing discomfort.

[0193] Condition C: The comfort level becomes significantly uncomfortable.

[0194] Real-time monitoring and condition comparison: The system continuously receives comfort information streams from the comfort judgment module and compares them with preset conditions in real time.

[0195] The first level of adjustment is triggered when the system detects that the comfort information meets the first preset condition B. The system determines that intervention is needed and generates a first set of adjustment instructions. This set of instructions may contain two specific instructions: a) adjust the current output power value of the radiofrequency therapy device downward by a preset step amount; b) instruct the robotic arm to control the treatment head to move slightly faster at the current treatment point to reduce the accumulation time of heat energy in the local area.

[0196] If the discomfort is not relieved after the adjustment, or if the comfort information directly meets the second preset condition C, the system determines that stronger intervention is needed and generates a second set of adjustment instructions. This set of instructions may include: a) immediately pausing the radiofrequency energy output; b) controlling the treatment head to be completely lifted off the skin or quickly moved to a non-treatment area, giving the patient a clear pause.

[0197] Once the instructions are executed, the generated set of adjustment instructions is sent in real time to the radio frequency energy generator and the treatment head motion control unit via the device's internal communication bus. The control unit parses and executes these instructions, thereby automatically adjusting the comfort parameters of radio frequency therapy, such as power, movement speed, and contact status.

[0198] Furthermore, the specific structure of the video data is as follows:

[0199] The video data is a data object with a logical hierarchy, the top of which contains a data header and a main data part.

[0200] The data header is used to store descriptive metadata, which includes a unique identifier for the video of the treated patient, the total number of frames in the video, the start and end timestamps of the acquisition, the camera device parameter identifier used for acquisition, and the identification code of the radiofrequency therapy device in the current treatment session.

[0201] The main data section is organized using a hierarchical index structure. The first layer is the time index layer, which is an ordered list. Each entry in the list corresponds to a specific point in time or time period and contains a pointer to the data block in the second layer.

[0202] The second layer of the main data portion is a feature data layer, which consists of multiple feature data blocks, each of which is associated with an entry in the time index layer.

[0203] Each feature data block further contains three logical partitions:

[0204] The first logical partition stores a subset of contour information extracted from the first matrix that corresponds to the current time point or time period. The subset includes a sequence of control point coordinates of vectorized contour lines and a set of derived geometric attributes calculated from the control point coordinate sequence.

[0205] The second logical partition stores a subset of facial expression change information extracted from the second matrix and corresponding to the current time point or time period. The subset includes the activation intensity value of each facial action unit region and a dynamic descriptor calculated based on the activation intensity of consecutive frames.

[0206] The third logical partition stores the fused feature data, which is the fused feature vector after the dimensionality reduction of the current time point or time period, processed by the feature fusion layer and the dimensionality reduction module.

[0207] The main data section also includes an auxiliary index layer, which is established based on the numerical range of contour geometric attributes or facial expression dynamic descriptors, and is used to support fast querying and retrieval of video data segments under specific feature conditions.

[0208] The video data organically integrates time, contours, expressions, and fusion features to form a self-contained composite data entity.

[0209] In this embodiment, the data structure is described as follows:

[0210] Video data is an elegantly designed, self-describing data object whose logical structure consists of two main parts:

[0211] The header section, like the title page of a file, stores global descriptive information. It includes:

[0212] session_id: A unique identifier for this treatment session.

[0213] total_frames: Total number of video frames.

[0214] start_time, end_time: The precise start and end timestamps for data collection.

[0215] camera_params: The intrinsic parameters of the camera used, such as focal length and extrinsic position identifiers, for possible coordinate inverse calculation.

[0216] device_id: The ID of the radiofrequency therapy device.

[0217] Original matrix dimensions: Records the number of rows and columns of the first and second matrices to ensure data traceability.

[0218] Main data section: This is the core of the data, organized using a hierarchical index structure to achieve efficient access.

[0219] The first layer is the time index layer, which is an array arranged in chronological order. Each element corresponds to a specific point in time or a short time interval and contains a key piece of information: a pointer to the location in memory or a file of the feature data block storing all detailed data at that moment. This allows the system to quickly locate data by time point.

[0220] The second layer: the feature data layer, consists of multiple feature data blocks that correspond one-to-one with the time index. Each data block is further divided into three logical partitions:

[0221] Contour information partitioning: Stores a subset of contour information extracted from the first matrix that is relevant to that time point, specifically including:

[0222] The vectorized sequence of control point coordinates for the contour lines.

[0223] The set of derived geometric properties calculated in real time from these coordinates, such as the total profile length, average curvature, and centroid position at that moment.

[0224] Facial Expression Information Partition: Stores a subset of facial expression changes extracted from the second matrix that are relevant to that point in time. Specifically, it includes:

[0225] The current activation intensity value of each predefined facial action unit (AU).

[0226] Dynamic descriptors calculated based on recent intensity values, such as the intensity rise slope of AU4 frown over the past second.

[0227] Feature fusion partitioning: Stores the final product after processing according to the process described in claim 5—the dimensionality-reduced fusion feature vector at this point in time, which is the direct input to the subsequent comfort judgment model.

[0228] Auxiliary index layer: In order to support complex queries, auxiliary indexes are also built in the main data part. For example, an index based on the magnitude of changes in the curvature of the jaw contour can be built, so that the system can quickly find the time points when all contours change drastically without scanning all the data.

[0229] This structure transforms the raw pixel stream of video data into a structured data entity with a clear timeline, well-defined feature hierarchy, and support for fast querying, greatly improving the efficiency and convenience of subsequent processing modules.

[0230] Furthermore, the method also includes treatment process recording and feedback steps:

[0231] The user interface of the radiofrequency therapy device dynamically displays a visual status indicator generated based on the aforementioned comfort information;

[0232] The visual status indicators include at least an icon reflecting the current comfort level and a simple chart reflecting recent trends in comfort.

[0233] Create and maintain a process log file bound to the current treatment session;

[0234] The process log file records the following events in chronological order: the initial settings of the radiofrequency therapy comfort parameters, each parameter adjustment event automatically triggered by the method and its details, and the sequence of changes in the comfort information over time.

[0235] When each parameter adjustment event is triggered, keyframe images are extracted from the current patient treatment video, and the keyframe images, along with the corresponding comfort information and adjustment instructions, are saved to the process log file.

[0236] At the end of this treatment session, a treatment process summary report is automatically generated based on the contents of the process log file.

[0237] In this embodiment, a patient status display area is set up on the operation screen of the radiofrequency therapy device. This area displays visual elements driven by comfort information in real time: a striking icon such as a green smiley face, a yellow neutral face, and a red pain face indicates the current comfort level; a scrolling trend chart shows the curve change of the comprehensive discomfort index in the last few tens of seconds, allowing the therapist to understand the trend at a glance.

[0238] A process log file is created. When treatment begins, the system automatically creates a log file that is bound to the current treatment session.

[0239] Event logging faithfully records key events in chronological order in the log:

[0240] When parameters are changed, the initial settings of all comfort parameters such as radio frequency power and mode are recorded. Whenever the system automatically triggers parameter adjustment as described in claim 7, an event is recorded, which includes the adjustment time, parameter value before adjustment, parameter value after adjustment, and a snapshot of the triggered comfort information.

[0241] The comfort trajectory records the current comfort level, trend, and confidence level at a configurable frequency, such as once per second.

[0242] Keyframe snapshot association: At each time parameter adjustment is automatically triggered, the system automatically captures several keyframe images from the original video stream before and after the adjustment. These images are associated with the record of this adjustment event and saved to provide intuitive visual evidence for post-event analysis, such as whether the patient did show signs of pain before the adjustment.

[0243] A treatment summary report is generated. After treatment, the system automatically analyzes the process log and generates a summary report. The report includes: the total treatment duration, the percentage of time spent at each comfort level, the number of automatic adjustment triggers, and the average comfort score during the treatment. This report, along with detailed log files and associated keyframes, constitutes a complete and traceable digital treatment record.

[0244] Furthermore, the method adapts to different radiofrequency treatment application scenarios by configuring the contour region of interest of the first linear sparse autoencoder network and the facial action unit region of interest of the second linear sparse autoencoder network.

[0245] When applied to radiofrequency therapy targeting body parts, the first linear sparse autoencoder network is adjusted to focus on changes in the body surface contour of the treatment site and its surrounding area.

[0246] When applied to radiofrequency treatment targeting the face, the first linear sparse autoencoder network is adjusted to focus on the overall or specific facial contours, while the second linear sparse autoencoder network is enabled to perform fine analysis of facial expression changes.

[0247] When the treatment environment makes the patient's face invisible or the video quality is insufficient for expression analysis, the method described above mainly relies on the contour information for comfort judgment and parameter adjustment.

[0248] In this embodiment, the flexibility lies in the attention configuration of the two core processing networks, the first and second linear sparse autoencoders, rather than in changing the core algorithm.

[0249] Adaptation to Body Treatment Scenarios: When used for body shaping such as radiofrequency treatments on the abdomen and thighs, patients are usually wearing treatment garments, and their faces may not be in the field of vision. In this case, the system is configured as follows:

[0250] The first linear sparse autoencoder network is adjusted so that the region of interest for its feature extraction is relabeled as the treatment site and the surrounding body surface contours; for example, when treating the abdomen, the network focuses on extracting the contour undulations and wrinkle changes of the abdominal skin surface, which may indicate muscle tension or spasm.

[0251] Since the face is invisible or does not require analysis, the weights of the second linear sparse autoencoder network can be disabled or reduced. In this case, the system's comfort judgment mainly, or even entirely, depends on the tension, twitching, and other patterns analyzed from the body surface contour information.

[0252] Facial treatment scenario adaptation: When used for facial rejuvenation such as full-face skin tightening treatment, the patient's front is fully exposed. In this case, the system performs full-function configuration:

[0253] The first network focuses on changes in the overall facial contour or specific treatment areas such as nasolabial folds and the jawline. For example, it monitors whether the corners of the mouth are pulled downwards due to discomfort, causing contour deformation.

[0254] The second network was activated to perform detailed analysis of facial expression changes, capturing micro-expressions in areas such as the eyebrows, eyes, and mouth caused by the healing energy.

[0255] The system integrates both contour and facial expression information to make a more accurate overall judgment.

[0256] In situations with limited environmental constraints, even when treating the face, the quality of facial video may be extremely poor due to patient positioning, dressing obstructions, or other factors. In such cases, the system can dynamically adjust its strategy, primarily relying on still reliable contour information for conservative comfort assessment and parameter control, ensuring that the system can still provide basic safety features even under the worst-case scenario.

[0257] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0258] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0259] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A method for intelligent control of comfort parameters in radiofrequency therapy, characterized in that, include: S1, capturing videos of patient treatment; S2, Perform first linear sparse autoencoder network processing on the video of the treated patient to collect contour information; S3, the video of the treated patient is processed by a second linear sparse autoencoder network to collect facial expression change information. The second linear sparse autoencoder network is deeper in level than the first linear sparse autoencoder network, and the signal-to-noise ratio of the second linear sparse autoencoder network is lower than that of the first linear sparse autoencoder network. The second linear sparse autoencoder network is complementary to the first linear sparse autoencoder network. The contour information and the facial expression change information are complementary information. S4, fuse the contour information and the facial expression change information to collect video data, wherein the contour information is a first matrix representing contour information, the facial expression change information is a second matrix representing facial expression change information, and the video data includes contour information and facial expression change information; S5. Based on the video data, determine the patient's comfort information; S6, when the comfort information meets the preset conditions, adjust the radiofrequency treatment comfort parameters.

2. The method according to claim 1, characterized in that, S1 specifically includes: During radiofrequency treatment, a camera device is simultaneously activated, and the camera device is configured to capture a continuous sequence of images including the patient's treatment area and facial region; Adjust the viewing angle and focal length of the camera device so that the continuous image sequence can simultaneously cover the body surface contour area and the facial expression area that reacts to the treatment; The frame rate of the camera device is set to capture the dynamic process of the patient's physiological response; The continuous image sequence is arranged and packaged in chronological order to generate the video of the treated patient.

3. The method according to claim 1, characterized in that, S2 specifically includes: Each frame of the video of the treated patient is input into the input layer of the first linear sparse autoencoder network; The encoding layer of the first linear sparse autoencoder performs a convolution operation on the input image to extract primary image features and generate a first set of feature maps; The hidden layer of the first linear sparse autoencoder network applies a first sparsity constraint to the first set of feature maps. The first sparsity constraint is used to suppress the activation of non-contour-related feature neurons and enhance the response of contour edge feature neurons. The decoding layer of the first linear sparse autoencoder performs a deconvolution operation on the sparsified features to reconstruct an output image that mainly contains contour information. The difference between the output image and the original input image is calculated, and continuous contour lines are identified from the difference using an edge detection algorithm; The identified contour lines are vectorized and the geometric properties of each contour line are calculated, including line length, curvature distribution, and enclosed area. Based on the time sequence, the vectorized contour lines and their geometric attributes corresponding to each frame of the image are arranged to form the first matrix; Each row of the first matrix corresponds to a video frame, and each column corresponds to a feature value of a contour geometric attribute. The first matrix is ​​the contour information.

4. The method according to claim 1, characterized in that, S3 performs second linear sparse autoencoder network processing on the treated patient video to collect facial expression change information, specifically including: From each frame of the video of the treated patient, locate and crop out a sub-image region containing the patient's face; The facial sub-image region is input into the input layer of the second linear sparse autoencoder network; The first-level encoding layer of the second linear sparse autoencoder network uses small-sized convolutional kernels to extract features from the input facial sub-image, obtaining a second set of feature maps containing subtle texture variations; The second-level encoding layer of the second linear sparse autoencoder performs deeper convolution and pooling operations on the second set of feature maps to extract higher-order semantic features. The hidden layer of the second linear sparse autoencoder network applies a second sparsity constraint to the higher-order semantic features. The strength of the second sparsity constraint is higher than that of the first sparsity constraint. This constraint is used to focus on specific movement patterns in facial muscle groups that are related to significant facial expression changes under conditions of low signal-to-noise ratio, while filtering out non-expression-related muscle tremors and image noise. The decoding layer of the second linear sparse autoencoder network reconstructs the high-order semantic features with strong sparsity constraints into an expression change intensity distribution map through multiple upsampling and deconvolution operations. In the facial expression intensity distribution map, multiple predefined facial action unit regions are divided, and each region is associated with a specific facial muscle group. For each facial motion unit region, the intensity change process in consecutive video frames is analyzed, and temporal dynamic features are extracted. The temporal dynamic features include the peak value, mean value, frequency of change, and duration of intensity changes. The temporal dynamic features of all facial motion unit regions are organized and encoded frame by frame to form a second matrix; Each row of the second matrix corresponds to a video frame, and each column corresponds to the temporal dynamic feature value of a facial motion unit region. The second matrix represents the expression change information.

5. The method according to claims 3 and 4, characterized in that, The process of fusing the contour information and the facial expression change information to collect video data specifically includes: Receive the first matrix and the second matrix as input; Establish a time synchronization relationship between the first matrix and the second matrix to ensure that contour features and facial features from the same point in time can be associated. A feature fusion layer is constructed, which concatenates the contour feature vector and the expression feature vector at the same time point to generate an initial fused feature vector; Calculate the mutual information between the contour feature vector and the expression feature vector within the historical time window, and dynamically adjust the weight coefficients of the two types of features during splicing based on the mutual information, so that the feature combination with strong information complementarity can obtain a higher fusion weight. The weighted concatenation of the initial fused feature vector is input into a dimension reduction module, which extracts the main components through linear transformation to generate a dimension-reduced fused feature vector. The fused feature vectors, after reducing the dimensionality of all time points, are arranged in chronological order to form a three-dimensional tensor; The first dimension of the three-dimensional tensor represents the time series, the second dimension represents the various dimensions of the fused feature vector, and the third dimension is reserved as an extended dimension. Add metadata to the three-dimensional tensor, the metadata including at least the start time of video capture, the total duration, and the original dimension information of the first matrix and the second matrix; The video data is encapsulated into a three-dimensional tensor with accompanying metadata.

6. The method according to claim 5, characterized in that, S5 specifically includes: From the three-dimensional tensor of the video data, extract the fused feature vector for each moment in chronological order; The fused feature vector at each time step is input into a predefined classification model; The classification model maps the input feature vector to a multidimensional comfort space and divides the comfort space into at least three non-overlapping decision regions, each corresponding to a different comfort level. The classification model outputs the comfort level at the current moment as a preliminary judgment result; Based on the preliminary judgment results of multiple consecutive moments preceding the current moment, a time series consistency analysis is performed; If the time-series consistency analysis detects a continuous and unidirectional trend in comfort level, a trend warning signal is generated. Based on the preliminary judgment results at the current moment and the trend warning signal, the structured comfort information is generated; The comfort information includes at least the current comfort level, trend warning status, and judgment confidence level.

7. The method according to claim 6, characterized in that, S6 specifically includes: Set multiple preset conditions associated with different comfort levels and trend warning states; Monitor the comfort information and compare it with the preset conditions; If the comfort information indicates that the current comfort level is the first level and there is no trend warning, then parameter adjustment will not be triggered; If the comfort information indicates that the current comfort level has dropped to the second level, or triggers a warning of an increasing trend of discomfort, then the first preset condition is determined to be met. When the first preset condition is met, a first adjustment instruction set is generated. The first adjustment instruction set is used to guide the radiofrequency therapy device to reduce the output power of radiofrequency energy and shorten the dwell time at the same treatment point. If the comfort information indicates that the current comfort level has further decreased to the third level, then the second preset condition is determined to be met; When the second preset condition is met, a second adjustment instruction set is generated. The second adjustment instruction set is used to guide the radiofrequency therapy device to pause the radiofrequency energy output and perform the retraction or movement operation of the treatment head. The generated set of adjustment instructions is sent to the control unit of the radiofrequency therapy device to perform adjustments to the radiofrequency therapy comfort parameters.

8. The method according to claim 5, characterized in that, The specific structure of the video data is as follows: The video data is a data object with a logical hierarchy, the top of which contains a data header and a main data part. The data header is used to store descriptive metadata, which includes a unique identifier for the video of the treated patient, the total number of frames in the video, the start and end timestamps of the acquisition, the camera device parameter identifier used for acquisition, and the identification code of the radiofrequency therapy device in the current treatment session. The main data section is organized using a hierarchical index structure. The first layer is the time index layer, which is an ordered list. Each entry in the list corresponds to a specific point in time or time period and contains a pointer to the data block in the second layer. The second layer of the main data section is the feature data layer, which consists of multiple feature data blocks, and each feature data block is associated with an entry in the time index layer. Each feature data block further contains three logical partitions: The first logical partition stores a subset of contour information extracted from the first matrix that corresponds to the current time point or time period. The subset includes a sequence of control point coordinates of vectorized contour lines and a set of derived geometric attributes calculated from the control point coordinate sequence. The second logical partition stores a subset of facial expression change information extracted from the second matrix and corresponding to the current time point or time period. The subset includes the activation intensity value of each facial action unit region and a dynamic descriptor calculated based on the activation intensity of consecutive frames. The third logical partition stores the fused feature data, which is the fused feature vector after the dimensionality reduction of the current time point or time period, processed by the feature fusion layer and the dimensionality reduction module. The main data section also includes an auxiliary index layer, which is established based on the numerical range of contour geometric attributes or facial expression dynamic descriptors, and is used to support fast querying and retrieval of video data segments under specific feature conditions. The video data organically integrates time, contours, expressions, and fusion features to form a self-contained composite data entity.

9. The method according to claim 1, characterized in that, The method also includes steps for recording and providing feedback on the treatment process: The user interface of the radiofrequency therapy device dynamically displays a visual status indicator generated based on the aforementioned comfort information; The visual status indicators include at least an icon reflecting the current comfort level and a simple chart reflecting recent trends in comfort level. Create and maintain a process log file bound to the current treatment session; The process log file records the following events in chronological order: the initial settings of the radiofrequency therapy comfort parameters, each parameter adjustment event automatically triggered by the method and its details, and the sequence of changes in the comfort information over time. When each parameter adjustment event is triggered, keyframe images are extracted from the current patient treatment video, and the keyframe images, along with the corresponding comfort information and adjustment instructions, are saved to the process log file. At the end of this treatment session, a treatment process summary report is automatically generated based on the contents of the process log file.

10. The method according to claim 1, characterized in that, The method adapts to different radiofrequency treatment application scenarios by configuring the contour region of the first linear sparse autoencoder network and the facial motion unit region of the second linear sparse autoencoder network. When applied to radiofrequency therapy targeting body parts, the first linear sparse autoencoder network is adjusted to focus on changes in the body surface contour of the treatment site and its surrounding area. When applied to radiofrequency treatment targeting the face, the first linear sparse autoencoder network is adjusted to focus on the overall or specific facial contours, while the second linear sparse autoencoder network is enabled to perform fine analysis of facial expression changes. When the treatment environment makes the patient's face invisible or the video quality is insufficient for expression analysis, the method described above mainly relies on the contour information for comfort judgment and parameter adjustment.

Images