Visual detection and positioning method and system for acupuncture points

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an acupuncture point visual detection and positioning method and system, and the method comprises the steps: synchronously collecting multi-view three-dimensional image data and flexible multi-parameter deformation data of a target area, and the multi-view three-dimensional image data comprises body surface textures and dynamic tracks of anatomical mark points; according to the method, multi-view three-dimensional image data and flexible multi-parameter deformation data of a target area are synchronously collected, an individual deformation feature model is constructed through mark point track purification processing, then a comprehensive feature set containing real-time state and trend prediction is generated through time sequence multi-mode fusion, and after initial acupuncture point coordinates are output through an AI basic matching model, the multi-view three-dimensional image data and flexible multi-parameter deformation data of the target area are obtained. Point-by-point correction is carried out by combining an AI self-adaptive correction model with a differential deformation compensation coefficient, irregular offset and predicted offset, and model parameters are dynamically adjusted through real-time feedback calibration, so that the problem that in the prior art, only static body shape proportional positioning is relied on, and dynamic deformation of soft tissue is not considered, so that acupuncture point offset is caused is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for visual detection and localization of acupuncture points. Background Technology

[0002] The core efficacy of acupuncture and physiotherapy depends on the accuracy of acupoint location. The actual location of acupoints can change with subtle variations in human physiological characteristics, and traditional methods relying on experience or surface landmarks are no longer sufficient to meet the standardized requirements of modern diagnosis and treatment. With the increasing penetration of artificial intelligence technology into the medical field, AI-assisted acupoint location has become an important direction for overcoming the limitations of traditional methods. Through image acquisition and algorithm analysis, it is hoped that precise and automated acupoint location can be achieved.

[0003] The currently published invention patent, CN119745685A, entitled "AI-Based Automated Acupoint Recognition and Positioning System and Method," proposes a scheme to establish a sample model by acquiring three-dimensional images of the human body, comparing the target image with the sample model in terms of size ratio, and locating acupoints and projecting them onto the human skin surface after a successful match. This scheme improves the convenience of positioning through visualization and is suitable for people of different body types, but it does not consider key influencing factors in human physiological characteristics.

[0004] Human soft tissues undergo dynamic deformation due to factors such as muscle tension, subtle adjustments in body posture, and differences in local fat distribution. This deformation is not simply a change in size or proportion; rather, it causes irregular, minute shifts in the actual location of acupoints relative to surface reference landmarks. Current technologies rely solely on static body size proportions for acupoint matching and positioning, failing to quantify and compensate for the locational deviations caused by these soft tissue deformations. Consequently, even with successful body matching, acupoint positioning cannot achieve the millimeter-level precision required for clinical application, thus affecting the stability of acupuncture and physiotherapy efficacy. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a visual detection and positioning method and system for acupuncture points, which solves the problem that existing technologies rely solely on static body proportions for positioning and fail to consider acupuncture point offset caused by dynamic deformation of soft tissues, thus failing to achieve millimeter-level positioning accuracy.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for visual detection and localization of acupuncture points, comprising the following steps: S1: Synchronously acquire multi-view three-dimensional image data and flexible multi-parameter deformation data of the target area. The multi-view three-dimensional image data includes the dynamic trajectory of body surface texture and anatomical landmarks. The flexible multi-parameter deformation data includes the pressure distribution, tensile amount and elastic coefficient of soft tissue. S2: Extract and track multiple highly stable anatomical landmarks based on multi-view 3D image data, generate dynamic trajectories of the landmarks, and purify the dynamic trajectories of the landmarks by combining flexible multi-parameter deformation data; S3: Construct an individual deformation feature model based on the dynamic trajectory of the purified marker points and flexible multi-parameter deformation data; S4: Perform temporal multimodal fusion on the multi-view 3D image data, flexible multi-parameter deformation data, and purified marker dynamic trajectory to generate a comprehensive feature set; S5: Call the preset sample model library and the individual deformation feature model. Based on the body shape features and purified marker point data in the comprehensive feature set, output the preliminary acupoint coordinates through the AI ​​basic matching model. Then, based on the multi-parameter deformation data and trend features in the comprehensive feature set, correct the preliminary acupoint coordinates point by point through the AI ​​adaptive correction model and output the dynamically optimized acupoint coordinates. S6: Based on the dynamically optimized acupoint coordinates, perform visualization projection and collect the fitting deviation data between the projected position and the actual body surface. Combined with the real-time updated flexible multi-parameter deformation data, dynamically adjust the parameters of the AI ​​adaptive correction model. When the fitting deviation data is greater than a preset threshold, trigger the re-execution of S1 to S5.

[0007] Furthermore, the implementation of S1 includes the following specific steps: S11: Through multiple array-distributed acquisition terminals, the three-dimensional contour, surface texture and dynamic trajectory of anatomical landmarks of the human target area are captured synchronously from different angles to generate continuous frame image data. S12: By covering the target area with an adhesive array sensor, the pressure distribution, tensile amount and elastic coefficient data of the soft tissue are collected simultaneously. S13: Align the continuous frame image data and multi-parameter deformation data according to a unified time reference to form a synchronized and aligned fused data stream.

[0008] Furthermore, the implementation of S2 includes the following specific steps: S21: Identify and extract multiple highly stable anatomical landmarks from the continuous frame image data; S22: Apply the inter-frame association algorithm to filter out interference points caused by temporary wrinkles or hair occlusion on the body surface based on the positional continuity of the marker points between adjacent frames, and generate preliminary dynamic trajectories of the marker points. S23: Perform correlation analysis between the preliminary dynamic trajectory of the marker points and the elastic coefficient in the flexible multi-parameter deformation data, remove trajectory abrupt change points caused by abnormal deformation of local soft tissue, and obtain the purified dynamic trajectory of the marker points.

[0009] Furthermore, the implementation of S3 includes the following specific steps: S31: Receive the purified marker point dynamic trajectory and flexible multi-parameter deformation data, and establish an individual deformation feature database; S32: In the individual deformation feature database, the mapping relationship between soft tissue deformation parameters and corresponding acupoint offsets under different gradients of muscle tension and different angle ranges of body posture is stored. S33: The mapping relationship is iteratively optimized by collecting historical data multiple times to form a dynamically updatable individual deformation feature model.

[0010] Furthermore, the implementation of S4 includes the following specific steps: S41: Extract spatial features from the multi-view 3D image data to obtain 3D spatial coordinate features; S42: Extract physical features from the flexible multi-parameter deformation data to obtain pressure, tensile, and elastic feature vectors; S43: Extract trajectory features from the dynamic trajectory of the purified marker points to obtain trajectory velocity and acceleration features; S44: Unify the dimensions of the three-dimensional spatial coordinate features, physical feature vectors and trajectory features, and perform weighted fusion according to preset weight rules to generate feature-level fusion results; S45: Introduce historical time-series data, use the sliding window algorithm to perform time-series analysis on the feature-level fusion results, extract deformation trend features, and generate a comprehensive feature set that includes real-time status and trend prediction.

[0011] Furthermore, in step S5, the process by which the AI ​​adaptive correction model corrects the initial acupoint coordinates point by point includes: S51: Assign differentiated deformation compensation coefficients to acupoints in different body regions. ,in Acupoint index; S52: Based on the multi-parameter deformation data in the comprehensive feature set, calculate the irregular offset of each acupoint caused by real-time deformation. ; S53: Based on the deformation trend characteristics, calculate the predicted offset for future times using a deformation prediction algorithm. ; S54: Initial acupoint coordinates Irregular offset and predicted offset Weighted summation is performed to obtain dynamically optimized acupoint coordinates. The calculation formula is as follows:

[0012] in, This is a trend prediction weighting factor used to adjust the contribution of the prediction offset in the final correction.

[0013] Furthermore, in S54, the irregular offset amount The calculation is based on the following formula:

[0014] in, For the region adaptive compensation function, To synthesize the feature vectors related to deformation in the feature set, For the first The body region identifier to which each acupoint belongs.

[0015] Furthermore, the implementation of S6 includes the following specific steps: S61: Project the dynamically optimized acupoint coordinates onto the human body surface using laser or optical projection equipment; S62: Collect data on the fit deviation between the projected light spot and the preset body surface reference point through an auxiliary vision sensor; S63: The fitting deviation data and the real-time collected flexible multi-parameter deformation data are used together as feedback signals and input into the AI ​​adaptive correction model; S64: Based on the feedback signal, dynamically adjust the deformation compensation coefficient in the AI ​​adaptive correction model. and trend prediction weighting factors ; S65: Determine whether the fitting deviation data is greater than a preset threshold. If so, send a re-sampling command to the multi-view three-dimensional image acquisition unit and the flexible multi-parameter deformation sensing unit, and re-execute S1 to S5.

[0016] Furthermore, in S5, the AI ​​basic matching model and the AI ​​adaptive correction model adopt a serial collaborative architecture, wherein the output of the AI ​​basic matching model is directly used as the input of the AI ​​adaptive correction model, and the training process of the AI ​​adaptive correction model depends on the historical deformation-offset mapping relationship provided by the individual deformation feature model.

[0017] The present invention also provides a visual detection and positioning system for acupuncture points, comprising: A multi-view 3D image acquisition unit is used to simultaneously capture the 3D contours, surface textures, and dynamic trajectories of anatomical landmarks of the target human body area. A flexible multi-parameter deformation sensing unit is used to cover the target area in an adhesive array form and simultaneously collect data on the pressure distribution, tensile amount and elastic coefficient of soft tissue. The surface landmark dynamic tracking module is used to extract and track highly stable anatomical landmarks from multi-view 3D images, and to refine the trajectory by combining deformation data. The individual deformation feature modeling module is used to construct and maintain an individual deformation feature model based on the purified trajectory data and deformation data. The temporal multimodal fusion unit is used to perform feature-level and temporal-level dual fusion of 3D image features, deformation physical features, and marker point trajectory features to generate a comprehensive feature set; The AI ​​basic matching model is used to call the preset sample model library and individual deformation feature model, and output the preliminary acupoint coordinates based on the comprehensive feature set. The AI ​​adaptive correction model is used to dynamically correct the initial acupoint coordinates point by point by combining deformation data and trend features in the comprehensive feature set. The acupoint precise positioning module is used to receive the corrected acupoint coordinates and control the visualization projection device for display. The real-time feedback calibration unit is used to collect projection fitting deviations, dynamically adjust the parameters of the AI ​​adaptive correction model by combining real-time deformation data, and trigger the system to re-collect and reposition when the deviation exceeds the limit.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention simultaneously acquires multi-view 3D image data and flexible multi-parameter deformation data of the target area, constructs an individual deformation feature model through marker point trajectory purification processing, and then generates a comprehensive feature set containing real-time state and trend prediction through temporal multimodal fusion. After outputting preliminary acupoint coordinates through an AI basic matching model, an AI adaptive correction model performs point-by-point correction by combining differential deformation compensation coefficients, irregular offsets, and predicted offsets. With real-time feedback calibration, the model parameters are dynamically adjusted, effectively solving the problem that existing technologies rely solely on static body proportions for positioning and do not consider the dynamic deformation of soft tissues that leads to acupoint offset. It achieves millimeter-level acupoint positioning accuracy and can adapt to changes in human muscle tension and body posture adjustments that cause dynamic deformation of soft tissues, making the positioning results more closely match the actual acupoint location and improving the accuracy and automation of acupuncture positioning. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a flowchart of the data synchronization acquisition and processing sub-process of the present invention; Figure 3 This is a flowchart of the marker point trajectory purification sub-process of the present invention; Figure 4 This is a flowchart of the temporal multimodal fusion sub-process of the present invention; Figure 5This is a flowchart of the AI ​​adaptive correction model of the present invention; Figure 6 This is a flowchart illustrating the system architecture and real-time feedback calibration process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0021] Please see Figure 1-5 This invention provides a method for visual detection and localization of acupuncture points, which includes the following steps: First, perform step S1: synchronously acquire data, namely, simultaneously acquire multi-view 3D image data and flexible multi-parameter deformation data of the target area. The multi-view 3D image data includes surface texture and dynamic trajectories of anatomical landmarks, while the flexible multi-parameter deformation data includes pressure distribution, stretching, and elasticity coefficient of soft tissue. Specifically, multiple arrayed acquisition terminals synchronously capture the 3D contour, surface texture, and dynamic trajectories of anatomical landmarks of the target area from different angles, generating continuous frame image data. The acquisition terminals can use eight 1920×1080 resolution 3D depth cameras, distributed around the target area at angles of front, back, left and right sides, upper left, lower left, upper right, and lower right. The overlap rate of adjacent camera views is set to 30% to ensure no blind spots, and the frame rate is set to 60fps to capture the dynamic changes of landmarks. Simultaneously, a flexible array sensor covers the target area, synchronously acquiring data on the pressure distribution, stretching, and elasticity coefficient of soft tissue. The flexible piezoelectric array sensor has a sensor unit spacing of 2mm, a pressure measurement range of 0-50kPa, a tensile measurement accuracy of 0.01mm, and an elastic coefficient measurement range of 0.1-10MPa. The sensor is fixed to the target area of ​​the body surface using a medical-grade adhesive patch, ensuring close contact with the skin during the acquisition process without affecting human activity. Subsequently, the continuous frame image data and multi-parameter deformation data are aligned according to a unified time series benchmark. Using the system timestamp of the acquisition terminal as a benchmark, the timestamp of the flexible multi-parameter deformation data is calibrated, with the calibration error controlled within ±1ms, forming a synchronously aligned fused data stream to avoid subsequent analysis errors caused by inconsistent data timing.

[0022] Next, step S2 is executed: Multiple highly stable anatomical landmarks are extracted and tracked based on multi-view 3D image data, generating dynamic trajectories for these landmarks. These dynamic trajectories are then purified using flexible multi-parameter deformation data. Multiple highly stable anatomical landmarks are identified and extracted from continuous frame image data. These landmarks are selected from bony prominences, joint spaces, and other areas less prone to displacement due to soft tissue deformation, such as the spinous process of the seventh cervical vertebra, the inferior angle of the scapula in the shoulder, and the spinous processes of the thoracic vertebrae in the back. The continuous frame images are processed using a deep learning-based target detection algorithm (such as YOLOv8) to extract the pixel coordinates of the landmarks and convert them into 3D spatial coordinates. An inter-frame correlation algorithm is applied to filter out interference points caused by temporary folds or hair occlusion based on the positional continuity of landmarks between adjacent frames, generating preliminary dynamic trajectories for the landmarks. The inter-frame correlation algorithm uses the KLT optical flow method to calculate the position of landmarks in adjacent frames. The displacement vector is used to filter out interference points when the Euclidean distance of the displacement vector is greater than a preset inter-frame displacement threshold. The inter-frame displacement threshold is set according to the acquisition frame rate and the normal human activity speed, for example, 5 mm / frame. The preliminary dynamic trajectory of the marker point is correlated with the elastic coefficient in the flexible multi-parameter deformation data to remove trajectory abrupt changes caused by abnormal deformation of local soft tissue, thus obtaining the purified dynamic trajectory of the marker point. Specifically, the mean elastic coefficient of the region where the marker point is located is calculated. When the mean elastic coefficient exceeds the normal physiological range (e.g., 0.3-5 MPa), the corresponding trajectory point is determined to be an abrupt change point and is removed to ensure that the trajectory of the marker point can truly reflect the dynamic position of the human anatomical structure.

[0023] Then, execute step S3: Based on the purified marker dynamic trajectory and flexible multi-parameter deformation data, construct an individual deformation feature model. Receive the purified marker dynamic trajectory and flexible multi-parameter deformation data, and establish an individual deformation feature database. The database uses MySQL and stores the timestamp, three-dimensional coordinates of the marker, pressure distribution matrix, stretching matrix, and elastic coefficient matrix for each data point. The individual deformation feature database stores the mapping relationship between soft tissue deformation parameters and corresponding acupoint offsets under different gradients of muscle tension and different body angle ranges. Muscle tension is divided into four gradients: relaxed, mildly tense, moderately tense, and highly tense. This is determined by the mean pressure distribution and mean elastic coefficient in the flexible multi-parameter deformation data. For example, a mean pressure distribution of 0-10 kPa and a mean elastic coefficient of 0.3-1 MPa correspond to a relaxed state; a mean pressure distribution of 10-20 kPa and a mean elastic coefficient of 1-2 MPa correspond to a mildly tense state; and a mean pressure distribution of 20-30 kPa and a mean elastic coefficient of 1-2 MPa correspond to a highly tense state. A mean pressure distribution and elasticity coefficient of 2-3 MPa correspond to a moderate tension state, while a mean pressure distribution and elasticity coefficient of 30 kPa or higher correspond to a high tension state. Postural angles are divided according to the joint movement angles of the target area, such as the flexion, extension, and lateral flexion angles of the neck, with each angle range set in 15°, such as flexion 0-15°, 15-30°, etc. Soft tissue deformation parameters include the standard deviation of pressure distribution, the maximum value of stretching, and the variance of elasticity coefficient. Acupoint offset is obtained by comparing with the reference position of acupoints in a relaxed and neutral posture. The mapping relationship is iteratively optimized through multiple historical data collections to form a dynamically updatable individual deformation feature model. The iterative optimization adopts the gradient descent method, using the prediction error of acupoint offset as the loss function, and continuously adjusts the parameters in the mapping relationship. The model is automatically updated after each data collection to improve the model's adaptability to individual deformation features.

[0024] Subsequently, step S4 is executed: temporal multimodal fusion is performed on multi-view 3D image data, flexible multi-parameter deformation data, and the cleaned marker dynamic trajectory to generate a comprehensive feature set. Spatial feature extraction is performed on the multi-view 3D image data to obtain 3D spatial coordinate features. The PointNet algorithm is used to encode the features of the multi-view 3D point cloud data, outputting a 256-dimensional 3D spatial coordinate feature vector. This vector includes the contour features, texture features, and relative position features of the marker points in the target area. Physical feature extraction is performed on the flexible multi-parameter deformation data to obtain pressure, tension, and elasticity feature vectors. Principal component analysis is performed on the pressure distribution data, and the top 50 principal components are extracted as pressure features. The mean, variance, maximum, and minimum values ​​of the tension data are calculated. This process generates 16-dimensional tensile features; similar statistical analysis is performed on the elastic coefficient data to generate 16-dimensional elastic features. The pressure, tensile, and elastic features are then concatenated to obtain an 82-dimensional physical feature vector. Trajectory features are extracted from the cleaned marker dynamic trajectory to obtain trajectory velocity and acceleration features. Instantaneous velocity is obtained by calculating the difference between the three-dimensional coordinates of markers in adjacent frames, and acceleration is obtained by calculating the difference between the instantaneous velocities. Moving average processing is performed on the velocity and acceleration data (sliding window size is 5 frames). Each marker corresponds to 6-dimensional trajectory features (three-dimensional velocity, three-dimensional acceleration ... The trajectory features of all marker points are concatenated to form a trajectory feature matrix of dimension N×6 (N being the number of marker points). The dimensions of the three-dimensional spatial coordinate features, physical feature vectors, and trajectory features are unified. The three-dimensional spatial coordinate feature vectors (256-dimensional), physical feature vectors (82-dimensional), and trajectory feature matrix (N×6) are converted to the same dimension (e.g., 512-dimensional) through a fully connected layer. Then, they are weighted and fused according to a preset weighting rule to generate a feature-level fusion result. The weighting rule is set according to the degree of influence of each feature on acupoint location. For example, the weight of the three-dimensional spatial coordinate features is 0. 4. The weight of the physical feature vector is 0.3, and the weight of the trajectory feature is 0.3. The weight values ​​are determined by expert experience and cross-validation. Historical time series data is introduced, and a sliding window algorithm is used to perform time series analysis on the feature-level fusion results to extract deformation trend features and generate a comprehensive feature set containing real-time status and trend prediction. The sliding window size is set to 30 frames. By linearly fitting the feature-level fusion results within the window, the deformation trend slope is obtained. Combined with the variance of the features within the window, a trend feature vector is formed. After being concatenated with the real-time feature-level fusion results, a comprehensive feature set with a dimension of 1024 is obtained.

[0025] Then, the preset sample model library and individual deformation feature model are called. Based on the body shape features in the comprehensive feature set and the purified marker point data, the preliminary acupoint coordinates are output through the AI ​​basic matching model. Then, based on the multi-parameter deformation data and trend features in the comprehensive feature set, the preliminary acupoint coordinates are corrected point by point through the AI ​​adaptive correction model, and the dynamically optimized acupoint coordinates are output. The pre-set sample model library contains standard acupoint coordinate models for people of different ages and body types (thin, medium, and fat), totaling 10,000 samples. Each sample includes body type parameters (height, weight, body fat percentage), anatomical landmark coordinates, and corresponding standard acupoint coordinates. The AI ​​basic matching model uses a convolutional neural network (CNN), which takes body type features (derived from 3D spatial coordinate features) from the comprehensive feature set as input and purified landmark data as input. It performs similarity matching with samples in the sample model library and outputs preliminary acupoint coordinates. The similarity calculation uses cosine similarity. The AI ​​basic matching model and the AI ​​adaptive correction model adopt a cascaded collaborative architecture. The output of the AI ​​basic matching model is directly used as the input of the AI ​​adaptive correction model, and the training process of the AI ​​adaptive correction model depends on the historical deformation-offset mapping relationship provided by the individual deformation feature model.

[0026] In the process of the AI ​​adaptive correction model correcting the initial acupoint coordinates point by point, it first assigns differentiated deformation compensation coefficients to acupoints in different body regions. ,in For acupoint indexing, the deformation compensation coefficient is determined based on the soft tissue thickness and muscle distribution density of the acupoint area, such as acupoints in the neck region. Acupoints on the back acupoints in the shoulder area Acupoints in the limbs The coefficient values ​​are determined by statistically analyzing the degree to which acupoints in different regions are affected by deformation; based on multi-parameter deformation data in the comprehensive feature set, the irregular offset of each acupoint caused by real-time deformation is calculated. Its calculation formula is ,in The region-adaptive compensation function is implemented using the Gaussian process regression algorithm. This is a combination of feature vectors related to deformation in the comprehensive feature set (i.e., the concatenation of physical feature vectors and trend feature vectors). For the first The body region identifiers to which each acupoint belongs (e.g., R1 for the neck, R2 for the back, etc.); based on deformation trend characteristics, the predicted offset for future moments is calculated using a deformation prediction algorithm. The deformation prediction algorithm uses a Long Short-Term Memory (LSTM) network, taking historical time-series irregular offset data as input and outputting the predicted offset for the next frame; it also uses the initial acupoint coordinates. Irregular offset and predicted offset Weighted summation is performed to obtain dynamically optimized acupoint coordinates. The calculation formula is as follows:

[0027] in, This is a trend forecasting weighting factor used to adjust the contribution of the forecast bias to the final correction. The value of is determined based on the stability of the deformation trend; when the deformation trend fluctuates small... When the fluctuations are large The default value is 0.3. It is dynamically adjusted by monitoring the variance of the deformation trend in real time. When the variance is greater than 0.1, it is considered to have large fluctuations; otherwise, it is considered to have small fluctuations.

[0028] Finally, step S5 is executed: Visual projection is performed based on dynamically optimized acupoint coordinates, and data on the fit deviation between the projected position and the actual body surface is collected. Combined with real-time updated flexible multi-parameter deformation data, the parameters of the AI ​​adaptive correction model are dynamically adjusted. When the fit deviation data exceeds a preset threshold, the data acquisition and coordinate optimization process is re-executed. The dynamically optimized acupoint coordinates are projected onto the human body surface using a laser projection device with a projection accuracy of 0.1mm, a projection spot diameter of 1mm, and a red color for easy observation by medical personnel. A visual sensor (high-definition industrial camera, resolution 2592×1944, frame rate 30fps) is used to collect data on the fit deviation between the projected spot and a preset body surface reference point. The preset body surface reference point is a purified, highly stable anatomical landmark. The fit deviation data and the real-time collected flexible multi-parameter deformation data are used as feedback signals and input into the AI ​​adaptive correction model. Based on the feedback signals, the deformation compensation coefficient in the AI ​​adaptive correction model is dynamically adjusted. and trend prediction weighting factors The system adjusts the proportional-integral-derivative (PID) control algorithm to adjust parameter values ​​based on the magnitude and rate of change of the deviation. It then determines whether the fitting deviation data exceeds a preset threshold of 0.5 mm, which is determined based on the millimeter-level positioning accuracy requirements of traditional Chinese medicine acupuncture. If so, a re-sampling command is sent to the multi-view 3D image acquisition unit and the flexible multi-parameter deformation sensing unit to re-execute the data acquisition, marker point extraction and trajectory purification, individual deformation feature model construction, temporal multimodal fusion, and acupoint coordinate optimization processes, ensuring that the acupoint positioning accuracy meets clinical needs.

[0029] This embodiment, through the above process, deeply integrates multi-view three-dimensional image data with flexible multi-parameter deformation data, and combines AI model matching and correction to effectively compensate for acupoint offset caused by dynamic deformation of human soft tissue. It solves the problem of insufficient accuracy of traditional static body proportion positioning method, enabling acupoint positioning to adapt to subtle changes in human physiological characteristics. The positioning results are more in line with the actual acupoint location, providing technical support for the stability of acupuncture and physiotherapy efficacy. Example

[0030] Please see Figure 6 The present invention also provides a visual detection and positioning system for acupuncture points, comprising: a multi-view three-dimensional image acquisition unit, a flexible multi-parameter deformation perception unit, a body surface marker dynamic tracking module, an individual deformation feature modeling module, a temporal multimodal fusion unit, an AI basic matching model, an AI adaptive correction model, an acupuncture point precise positioning module, and a real-time feedback calibration unit.

[0031] The multi-view 3D image acquisition unit consists of multiple arrayed 3D depth cameras. Its core function is to simultaneously capture the 3D contours, surface textures, and dynamic trajectories of anatomical landmarks of the target human body region, transmitting the acquired continuous frame image data to the surface landmark dynamic tracking module. The flexible multi-parameter deformation sensing unit is an integrated array sensor that, after covering the target area, simultaneously acquires data on the pressure distribution, tensile amount, and elastic coefficient of soft tissue, transmitting this data to the individual deformation feature modeling module and the time-series multimodal fusion unit. The surface landmark dynamic tracking module receives continuous frame image data, extracts and tracks highly stable anatomical landmarks, and combines the flexible multi-parameter deformation data to purify the trajectory, transmitting the purified trajectory data to the individual deformation feature modeling module and the time-series multimodal fusion unit. Based on the purified trajectory data and deformation data, the individual deformation feature modeling module constructs and maintains an individual deformation feature model, providing a historical deformation-offset mapping for the AI ​​adaptive correction model. The process involves several steps: The temporal multimodal fusion unit performs dual fusion of 3D image features, deformation physical features, and marker trajectory features at both the feature level and temporal level to generate a comprehensive feature set, which is then transmitted to the AI ​​basic matching model. The AI ​​basic matching model calls a preset sample model library and an individual deformation feature model, outputting preliminary acupoint coordinates to the AI ​​adaptive correction model based on the comprehensive feature set. The AI ​​adaptive correction model combines deformation data and trend features from the comprehensive feature set to dynamically correct the preliminary acupoint coordinates point-by-point, outputting optimized coordinates to the acupoint precision positioning module. The acupoint precision positioning module receives the corrected acupoint coordinates and controls a visualization projection device (laser projection device) for display. The real-time feedback calibration unit collects projection fitting deviations through an auxiliary visual sensor and dynamically adjusts the parameters of the AI ​​adaptive correction model based on real-time deformation data. When the deviation exceeds the limit, the multi-view 3D image acquisition unit and the flexible multi-parameter deformation sensing unit are triggered to re-acquire data.

[0032] During system operation, the multi-view 3D image acquisition unit and the flexible multi-parameter deformation sensing unit simultaneously initiate data acquisition, which is then transmitted to various functional modules after time-series alignment. The body surface landmark dynamic tracking module and the individual deformation feature modeling module process the data in parallel to generate purification trajectories and individual deformation feature models. The time-series multimodal fusion unit fuses multi-source features and outputs a comprehensive feature set. The AI ​​basic matching model and the AI ​​adaptive correction model work in series to complete the initial matching and dynamic correction of acupoint coordinates. The acupoint precise positioning module realizes visual projection, and the real-time feedback calibration unit continuously monitors deviations and dynamically adjusts parameters to ensure positioning accuracy. The entire operation process is seamlessly connected, efficiently completing acupoint detection and positioning. Example

[0033] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below in conjunction with a specific application scenario.

[0034] This embodiment selects the acupoint location scenario in cervical spondylosis rehabilitation physiotherapy to verify the application of the above-mentioned acupuncture acupoint visual detection and location method and system. The application target is cervical spondylosis patients, and the target acupoints include Jianjing (GB21), Dazhui (GV14), Fengchi (GB20), and Tianzong (SI11), with the location area being the neck and upper back. First, the patient is placed on the physiotherapy bed and adjusted to a relaxed state, ensuring that the target area is not obstructed by clothing; the eight 3D depth cameras of the multi-view three-dimensional image acquisition unit are distributed around the patient's neck and upper back at preset angles, and the camera focal length and viewing angle are adjusted to ensure complete coverage of the target area; the flexible piezoelectric array sensor is attached to the target area using a medical adhesive patch, with the sensor edge aligned with anatomical landmarks (spinous process of the seventh cervical vertebra, inferior angle of the scapula) to ensure the accuracy of the acquired data.

[0035] After the system is started, the multi-view 3D image acquisition unit captures the 3D contour, surface texture and dynamic trajectory of anatomical landmarks of the target area at a frame rate of 60fps, generating continuous frame image data; the flexible multi-parameter deformation sensing unit simultaneously acquires the pressure distribution, stretching amount and elastic coefficient data of soft tissue. The pressure distribution data reflects the tension state of the patient's neck muscles, and the stretching amount and elastic coefficient data reflect the deformation characteristics of the soft tissue of the upper back; the two types of data are time-series aligned by timestamp calibration to form a fused data stream.

[0036] The surface landmark dynamic tracking module extracts highly stable anatomical landmarks such as the spinous process of the seventh cervical vertebra and the inferior angle of the scapula from continuous frame image data. It uses KLT optical flow to track the landmark trajectories, filtering out interference points caused by surface folds and hair occlusion. Then, it combines the elastic coefficients from flexible multi-parameter deformation data to eliminate trajectory abrupt changes caused by slight postural adjustments by the patient, resulting in the purified dynamic trajectory of the landmarks. The individual deformation feature modeling module, based on the purified trajectory and deformation data, establishes an individual deformation feature database for the patient. It stores the mapping relationship between soft tissue deformation parameters and acupoint offsets under different muscle tension levels (relaxed, mildly tense) and different neck posture angles (neutral, 15° flexion, 15° extension). The mapping relationship is iteratively optimized through historical data (multiple acquisitions during this physiotherapy session) to form an individual deformation feature model.

[0037] The temporal multimodal fusion unit performs PointNet feature extraction on multi-view 3D image data to obtain 3D spatial coordinate features; performs principal component analysis and statistical analysis on flexible multi-parameter deformation data to obtain physical feature vectors; calculates the velocity and acceleration of the dynamic trajectory of the cleaned marker points to obtain trajectory features; unifies the three types of feature dimensions to 512 dimensions, and performs weighted fusion with a weight of 0.4 for 3D spatial coordinate features, 0.3 for physical feature vectors, and 0.3 for trajectory features to obtain feature-level fusion results; uses a 30-frame sliding window to perform temporal analysis on the feature-level fusion results, extracts deformation trend features, and generates a comprehensive feature set.

[0038] The AI-based matching model calls a pre-defined sample model library to perform similarity matching between the body shape features from the comprehensive feature set and the purified marker data, outputting the preliminary acupoint coordinates of Jianjing (GB21), Dazhui (GV14), Fengchi (GB20), and Tianzong (SI11). The AI ​​adaptive correction model assigns differentiated deformation compensation coefficients to the four acupoints, with Jianjing (GB21) receiving the coefficients. Dazhui acupoint Fengchi acupoint , Tianzong point Based on the deformation-related feature vectors in the comprehensive feature set, the irregular offset of each acupoint is calculated through Gaussian process regression. Then, an LSTM network is used to predict the offset at future times. The trend prediction weight factor is used to calculate the dynamically optimized acupoint coordinates using a modified formula.

[0039] The acupoint precision positioning module controls the laser projection device to project optimized acupoint coordinates onto the patient's body surface, clearly marking the location of each acupoint with the projected light spot. The real-time feedback calibration unit uses a high-definition industrial camera to collect data on the fit deviation between the projected light spot and preset body surface reference points (spinous process of the seventh cervical vertebra, inferior angle of the scapula), and combines this with real-time updated flexible multi-parameter deformation data to dynamically adjust the AI ​​adaptive correction model. and Parameters; During this application, the fitting deviation data were all less than the preset threshold of 0.5mm, and the re-acquisition process was not triggered, and the positioning process remained stable.

[0040] During application, the system can adapt in real time to the dynamic deformation of soft tissues caused by changes in the degree of muscle relaxation and slight body posture adjustments. The acupoint projection position always maintains a high degree of fit with the actual acupoint position. Medical staff can directly perform acupuncture operations based on the projected light spot, which simplifies the positioning process, improves the efficiency of physiotherapy, and verifies the feasibility and practicality of the invention in the acupoint positioning scenario of cervical spondylosis rehabilitation physiotherapy.

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0042] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for visual detection and positioning of acupuncture points, characterized in that, The method comprises the following steps: S1: synchronously collecting multi-view three-dimensional image data and flexible multi-parameter deformation data of a target region, wherein the multi-view three-dimensional image data contains body surface texture and dynamic trajectories of anatomical landmarks, and the flexible multi-parameter deformation data contains pressure distribution, stretching amount and elastic coefficient of soft tissue; S2: extracting and tracking a plurality of high-stability anatomical landmarks based on the multi-view three-dimensional image data, generating dynamic trajectories of the landmarks, and purifying the dynamic trajectories of the landmarks in combination with the flexible multi-parameter deformation data; S3: constructing an individual deformation feature model based on the purified dynamic trajectories of the landmarks and the flexible multi-parameter deformation data; S4: performing time-series multi-modal fusion on the multi-view three-dimensional image data, the flexible multi-parameter deformation data and the purified dynamic trajectories of the landmarks to generate a comprehensive feature set; S5: calling a preset sample model library and the individual deformation feature model, outputting preliminary acupoint coordinates based on body shape features in the comprehensive feature set and the purified landmark data through an AI-based matching model, and performing point-by-point correction on the preliminary acupoint coordinates based on multi-parameter deformation data and trend features in the comprehensive feature set through an AI self-adaptive correction model to output dynamically optimized acupoint coordinates; S6: performing visual projection based on the dynamically optimized acupoint coordinates, collecting fitting deviation data of the projection position and the actual body surface, dynamically adjusting parameters of the AI self-adaptive correction model in combination with real-time updated flexible multi-parameter deformation data, and triggering re-execution of S1 to S5 when the fitting deviation data is greater than a preset threshold.

2. The method of claim 1, wherein, The implementation of S1 comprises the following specific steps: S11: synchronously capturing three-dimensional contours, body surface texture and dynamic trajectories of anatomical landmarks of a human target region from different angles through a plurality of array-distributed collection terminals to generate continuous frame image data; S12: covering the target region with a fitted array sensor to synchronously collect pressure distribution, stretching amount and elastic coefficient data of soft tissue; S13: aligning the continuous frame image data and the multi-parameter deformation data according to a unified time sequence reference to form a synchronous aligned fusion data stream.

3. The method of claim 1, wherein, The implementation of S2 comprises the following specific steps: S21: identifying and extracting a plurality of high-stability anatomical landmarks from the continuous frame image data; S22: applying an inter-frame correlation algorithm to filter interference points caused by temporary body surface wrinkles or hair obstruction according to the position continuity of adjacent landmarks to generate preliminary dynamic trajectories of the landmarks; S23: correlatively analyzing the preliminary dynamic trajectories of the landmarks and the elastic coefficient in the flexible multi-parameter deformation data to eliminate trajectory mutation points caused by local soft tissue abnormal deformation to obtain purified dynamic trajectories of the landmarks.

4. The method of claim 1, wherein, The implementation of S3 comprises the following specific steps: S31: receiving the purified dynamic trajectories of the landmarks and the flexible multi-parameter deformation data to establish an individual deformation feature database; S32: storing mapping relationships between soft tissue deformation parameters and corresponding acupoint offset amounts under different gradient muscle tension and different angle interval body posture angles in the individual deformation feature database; S33: iteratively optimize the mapping relationship through multiple historical acquisition data, and form an individual deformation feature model that can be dynamically updated.

5. The method of claim 1, wherein the method further comprises: The implementation of S4 includes the following specific steps: S41: spatial feature extraction is performed on the multi-view three-dimensional image data to obtain three-dimensional spatial coordinate features; S42: physical feature extraction is performed on the flexible multi-parameter deformation data to obtain pressure, stretching and elasticity feature vectors; S43: trajectory feature extraction is performed on the purified landmark point dynamic trajectory to obtain trajectory speed and acceleration features; S44: the three-dimensional spatial coordinate features, physical feature vectors and trajectory features are unified in dimension, and weighted fusion is performed according to a preset weight rule to generate a feature-level fusion result; S45: introduce historical time series data, and use a sliding window algorithm to perform time series analysis on the feature-level fusion result to extract deformation trend features and generate a comprehensive feature set containing real-time state and trend prediction.

6. The method of claim 1, wherein, In S5, the process of the AI adaptive correction model correcting the preliminary acupoint coordinates point by point includes: S51: assigning differentiating deformation compensation coefficients to acupoint sites of different body regions wherein is an acupoint index; S52: calculating, based on the multi-parameter deformation data in the comprehensive feature set, an irregular displacement amount of each acupoint caused by real-time deformation ; S53: According to the deformation trend feature, a predicted offset at a future time is calculated by a deformation prediction algorithm ; S54: performing weighted summation on the preliminary acupoint coordinates , irregular offset , and predicted offset to obtain dynamically optimized acupoint coordinates , and the calculation formula is: wherein, is a trend prediction weight factor, used to adjust the contribution of the predicted offset in the final correction.

7. The method of claim 6, wherein the method further comprises: In the S54, the irregularity offset is calculated based on the following equation: wherein, is a region adaptive compensation function, is a feature vector associated with the deformation in the comprehensive feature set, is a body region identifier to which the nth acupoint belongs.

8. The method of claim 1, wherein the method further comprises: The implementation of S6 includes the following specific steps: S61: project the dynamically optimized acupoint coordinates to the human body surface through laser or optical projection equipment; S62: collect the fitting deviation data between the projected light spots and the preset body surface reference points through an auxiliary visual sensor; S63: use the fitting deviation data and the real-time collected flexible multi-parameter deformation data as feedback signals, and input them into the AI adaptive correction model; S64: dynamically adjust a deformation compensation coefficient in the AI adaptive correction model based on the feedback signal and trend prediction weight factors ; S65: determine whether the fitting deviation data is greater than a preset threshold, and if so, send a re-sampling instruction to the multi-view three-dimensional image acquisition unit and the flexible multi-parameter deformation perception unit to re-execute S1 to S5.

9. The method of claim 1, wherein the method further comprises: In S5, the AI basic matching model and the AI adaptive correction model adopt a series cooperative architecture, wherein the output of the AI basic matching model is directly used as the input of the AI adaptive correction model, and the training process of the AI adaptive correction model depends on the historical deformation-offset mapping relationship provided by the individual deformation feature model.

10. A visual detection and positioning system for acupuncture points, for implementing a visual detection and positioning method for acupuncture points according to any one of claims 1 to 9, characterized in that, It includes: A multi-view three-dimensional image acquisition unit for synchronously capturing the three-dimensional contour, body surface texture and dynamic trajectory of anatomical landmarks of a human target region; A flexible multi-parameter deformation perception unit for covering the target region in a conformal array form and synchronously collecting pressure distribution, stretching amount and elasticity coefficient data of soft tissue; A body surface landmark dynamic tracking module for extracting and tracking high-stability anatomical landmark points from multi-view three-dimensional images, and purifying the trajectory in combination with deformation data; An individual deformation feature modeling module for constructing and maintaining an individual deformation feature model based on the purified trajectory data and deformation data; A time series multi-modal fusion unit for performing feature-level and time series-level double fusion on three-dimensional image features, deformation physical features and landmark point trajectory features to generate a comprehensive feature set; An AI basic matching model for calling a preset sample model library and an individual deformation feature model, and outputting preliminary acupoint coordinates based on the comprehensive feature set; An AI adaptive correction model is used to dynamically correct the preliminary acupoint coordinates point by point by combining the deformation data in the comprehensive feature set and the trend features; An acupoint precise positioning module is used to receive the corrected acupoint coordinates and control the visual projection device to display; A real-time feedback calibration unit is used to collect the projection fitting deviation, dynamically adjust the parameters of the AI adaptive correction model in combination with the real-time deformation data, and trigger the system to re-collect and position when the deviation exceeds the limit.

Images