An intelligent and accurate positioning method for eight-liao points based on joint driving of clinical knowledge and multi-modal image data

Abstract

This invention relates to the field of acupuncture point location technology, specifically a method for intelligent and precise location of the eight Liao acupoints based on a combination of clinical knowledge and multimodal image data. The method involves collecting multimodal image data of the patient's sacral region; preprocessing the multimodal image data; extracting multimodal fusion features through a teacher-student network architecture in a trained location model, and generating spatial constraint parameters by combining this with a clinical knowledge-based constraint perception module; obtaining the coordinates of the eight Liao acupoints through the teacher-student network architecture in the trained location model; post-processing the coordinates of the eight Liao acupoints; and integrating them into a mobile app system. This invention achieves a location error ≤10mm through the combined driving force of multimodal image data and traditional Chinese medicine clinical knowledge.

Description

Technical Field

[0001] This invention relates to the field of acupuncture point location technology, specifically a method for intelligent and precise location of the eight Liao points based on a combination of clinical knowledge and multimodal image data. Background Technology

[0002] Acupuncture, as an important part of traditional Chinese medicine, is widely used in the treatment of pelvic diseases, urinary system dysfunction, and chronic pain. The accurate location of the eight acupoints (Baliao) is crucial for therapeutic efficacy. Traditional methods for locating these points mainly rely on palpation and visual assessment of anatomical landmarks (such as the sacral angle and posterior superior iliac spine) by the physician, combined with clinical experience. This method has the following limitations:

[0003] 1. Low operational efficiency: The location of each acupoint requires repeated touching and measurement, which is time-consuming and affects the efficiency of diagnosis and treatment.

[0004] 2. High degree of subjectivity: The experience level of different physicians may lead to inconsistent localization results;

[0005] 3. Susceptible to individual differences: Patient body size, degree of obesity, or skeletal variations can increase the difficulty of localization;

[0006] 4. Environmental factors: Light conditions or changes in body position may further reduce reliability.

[0007] In recent years, computer vision and deep learning technologies have made progress in medical image analysis, such as using convolutional neural networks to detect key points on the human body. However, directly applying these technologies to the location of the eight acupoints (Baliao) still faces challenges:

[0008] 1. Acupoints lack obvious visual features, making it difficult to accurately map deep skeletal structures using only a single image modality (such as RGB images);

[0009] 2. Registration and fusion of multimodal data (such as CT and RGB images) are highly complex and prone to introducing errors;

[0010] 3. The model needs to adapt to the anatomical variations of different populations, while traditional algorithms have limited generalization ability and are difficult to integrate TCM clinical knowledge (such as the spatial distribution pattern of the posterior sacral foramen).

[0011] Therefore, to address the issues of insufficient positioning accuracy, low efficiency, and lack of multimodal data collaboration in existing technologies, it is necessary to develop an intelligent positioning method that integrates clinical knowledge and multimodal imaging to improve the accuracy and practicality of locating the Baliao acupoints. Summary of the Invention

[0012] To address the current problems of insufficient accuracy, low efficiency, and lack of multimodal data collaboration in acupoint location, this invention provides an intelligent and precise location method and system for the eight acupoints (Baliao) based on a combination of clinical knowledge and multimodal imaging data.

[0013] This invention is achieved through the following technical solution: a method for intelligent and precise localization of the eight acupoints (Baliao) based on clinical knowledge and multimodal imaging data, comprising the following steps:

[0014] S1. Collect multimodal image data of the patient's sacral region to obtain a multimodal image database, which includes pelvic CT images and body surface RGB images;

[0015] S2. Preprocess the multimodal image data obtained in S1;

[0016] S3. Extract multimodal fusion features from the teacher-student network architecture in the trained localization model, and generate spatial constraint parameters by combining them with the clinical knowledge constraint perception module;

[0017] S4. The trained teacher-student network model is integrated into the mobile APP system.

[0018] As a further improvement to the technical solution of the present invention, the multimodal acupoint database includes basic acupoint data and spatial positioning data, and the process of acquiring the basic acupoint data is as follows:

[0019] A1. Acquire multiple sets of pelvic CT images and body surface RGB images;

[0020] A2. Preprocess and label acupoint information on the pelvic CT images and body surface RGB images obtained in A1. The labeled acupoint information includes the standard name, anatomical location description, main functions and indications of the Eight Liao Acupoints.

[0021] The process of acquiring the spatial positioning data is as follows:

[0022] B1. Obtain three-dimensional data of the human sacral region using a three-dimensional scanning device;

[0023] B2. Label the three-dimensional coordinates of the eight acupoints, their relative distances and angular relationships with anatomical landmarks, and the three-dimensional data.

[0024] As a further improvement to the technical solution of the present invention, in the teacher-student network architecture, the teacher network is based on pelvic CT images using a ResNet-50 backbone, integrating the CBAM attention mechanism, focusing on learning the precise spatial location features of the posterior sacral foramen, and the output features are synchronously transmitted to the anatomical constraint submodule of the clinical knowledge constraint perception module for spatial parameter verification; the student network is based on RGB images using a MobileNetV3 lightweight architecture, and optimizes the inference speed through depthwise separable convolution, and its output features need to meet the individualized body size constraints of the clinical rule submodule; the teacher network and the student network achieve feature fusion through multi-level knowledge distillation.

[0025] As a further improvement to the technical solution of the present invention, the constraint mechanism of the clinical knowledge constraint perception module is as follows:

[0026] (1) Spatial symmetry and sequence constraint. This constraint maintains the left-right symmetry of human anatomical structures and the basic vertical arrangement order of the upper, middle, and lower liao points by multiplying the average symmetry deviation and sequence integrity. Its mathematical expression is as follows:

[0027] (1)

[0028] In the formula, Represents the averaging coefficient. This indicates the summation over the four acupoints. This indicates the coordinate position of the i-th acupoint on the left. This indicates the coordinate position of the i-th acupoint on the right. Indicates the sacral midline. This represents the symmetry tolerance threshold. Indicates an indicator function, This represents the ordinate of the i-th acupoint. This represents a sequence constraint condition. This represents the distance from the i-th acupoint on the left to the midline. This represents the distance from the i-th acupoint on the right side to the midline. This indicates a series of multiplication operations;

[0029] (2) Geometric relationship constraint between adjacent acupoints. This constraint, through a combination of standardized adjacent spacing deviation and normalized collinearity deviation, specifically constrains the spatial relationship between adjacent acupoints, promoting a reasonable linear arrangement of the eight acupoints and ensuring that the adjacent spacing basically conforms to anatomical statistical laws. Its mathematical expression is as follows:

[0030] (2)

[0031] In the formula, Represents the averaging coefficient. This indicates the summation of three adjacent acupoint pairs. This indicates the summation over the four acupoints. This represents the coordinates of the i-th acupoint. This represents the coordinate position of the (i+1)th acupoint. The actual Euclidean distance between adjacent acupoints. This represents the reference spacing between the i-th adjacent acupoint pairs. Indicates the spacing ratio. Indicates the standardized spacing deviation. This represents the best-fitting straight line. This represents the distance from the i-th acupoint to the fitted straight line. This indicates the maximum permissible collinearity deviation. Collinearity weighting coefficients Indicates the threshold for consistency in adjacency relationships;

[0032] (3) Individualized body measurement constraint, which ensures that the absolute positional relationship between each acupoint and the sacral midline conforms to individual anatomical characteristics, and its mathematical expression is:

[0033] (3)

[0034] In the formula, This represents the normalized importance weight of the i-th acupoint. Indicates an indicator function, This represents the actual distance from the i-th acupoint to the midline. This represents the reference distance from the i-th acupoint to the midline under standard body type. Indicates the scaling factor for body size. Indicates the individualized reference distance. Indicates absolute deviation. This represents the relative deviation rate. This represents the weighted voting score. This represents the tolerance threshold for individualized bias. This indicates the threshold for consistency of body dimensions.

[0035] As a further improvement to the technical solution of this invention, the trained localization model is converted into a format supported by the mobile inference framework through a model conversion tool and deployed on a mobile APP; real-time RGB images are captured by the mobile camera, and the locations of the eight acupoints are obtained through model inference and displayed on the real-time images, outputting the names of the eight acupoints, location information and acupuncture depth prompts.

[0036] The intelligent and precise location method for the eight acupoints based on clinical knowledge and multimodal imaging data provided in this invention has the following advantages compared with the prior art:

[0037] 1. This invention achieves a positioning error of ≤10mm by combining multimodal imaging data (CT and RGB) with traditional Chinese medicine clinical knowledge (such as the symmetry of the posterior sacral foramen).

[0038] 2. This invention employs a lightweight network design and a cascade optimization strategy, controlling the processing time to within 3 seconds;

[0039] 3. This invention combines a mobile AR interface to reduce the difficulty of operation and improve acupuncture efficiency;

[0040] 4. This invention supports multicenter validation (300 subjects) and maintains stability across different body size groups (error coefficient of variation <15%).

[0041] 5. This invention reduces hardware costs (annual rental fee ≤ 50,000 yuan) by deploying cloud computing resources, providing a feasible technical solution for the standardization of acupuncture treatment. Attached Figure Description

[0042] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This diagram illustrates the architecture of an intelligent and precise acupoint localization method for the Baliao acupoints, driven jointly by clinical knowledge and multimodal imaging data.

[0045] Figure 2 The left image shows the architecture of ResNet-50, and the right image shows the architecture of MobileNetV3. Detailed Implementation

[0046] To better understand the above-mentioned objectives, features, and advantages of the present invention, the solutions of the present invention will be further described below. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0048] The specific embodiments of the present invention will be described in detail below.

[0049] like Figures 1 to 2 As shown, a method for intelligent and precise localization of the eight acupoints (Baliao) based on clinical knowledge and multimodal imaging data includes the following steps:

[0050] S1. Collect multimodal image data of the patient's sacral region to obtain a multimodal image database, which includes pelvic CT images and body surface RGB images.

[0051] CT images are acquired using medical imaging equipment (such as a CT scanner), covering the sacral region with a resolution of at least 0.5 mm. Surface RGB images are acquired using a high-resolution camera (e.g., 1080p or higher), ensuring coverage of relevant surface landmarks of the eight acupoints (such as the posterior superior iliac spine and sacral angle). Simultaneously, musculoskeletal ultrasound verification data can be optionally acquired as the gold standard for subsequent model training and validation. This step ensures the diversity and clinical applicability of data sources, providing a foundation for multimodal fusion.

[0052] Specifically, the CT image acquisition parameters are as follows: tube voltage 120KV, tube current SmartmA 300-600, Protected Dose (lowest dose) scanning protocol, FOV of 350mm×350mm, matrix 512×512, and reconstruction slice thickness of 0.6mm. RGB image acquisition of the body surface requires standard lighting conditions, with the patient in a prone position to fully expose the lumbosacral region, and the sacral region captured using a digital camera with at least 12 megapixels in the anteroposterior view.

[0053] S2. Preprocess the multimodal image data obtained in S1;

[0054] The preprocessing in step S2 is as follows: CT images are reconstructed in three dimensions using the AW VolumeShare7 platform, and interfering structures such as soft tissues and blood vessels are cut and removed, while preserving complete bony landmarks (sacral shape and spinous process of the fifth lumbar vertebra, posterior superior iliac spine, sacral horn, and sacral hiatus); the RGB images of the body surface are processed for color balance correction, Gaussian filtering for noise reduction, pose alignment, and size normalization; finally, spatial registration between the CT three-dimensional coordinate system and the RGB two-dimensional image coordinate system is achieved through affine transformation or perspective projection model.

[0055] S3. Multimodal fusion features are extracted from the teacher-student network architecture in the post-trained localization model, and spatial constraint parameters are generated by combining them with the clinical knowledge constraint perception module. The post-trained localization model is obtained by training the localization model with a multimodal acupoint database. The localization model consists of a teacher-student network and a clinical knowledge constraint perception module. The clinical knowledge constraint perception module includes an anatomical constraint submodule and a clinical rule submodule.

[0056] In this embodiment, the teacher-student network uses a multi-task convolutional neural network (CNN) backbone. In this teacher-student network architecture, the teacher network is based on CT images using a ResNet-50 backbone (approximately 25.6M parameters, see Table 1), integrating a CBAM attention mechanism to focus on learning the precise spatial location features of the posterior sacral foramen. The output features are synchronously transmitted to the anatomical constraint submodule of the clinical knowledge constraint perception module for spatial parameter verification. The student network is based on RGB images using a lightweight MobileNetV3 architecture (approximately 2.5M parameters, see Table 2), optimizing inference speed through depthwise separable convolutions. Its output features must meet the individualized body-size constraints (weighted voting score ≥ 0.8) of the clinical rule submodule. The teacher and student networks achieve feature fusion through multi-level knowledge distillation.

[0057] Table 1: Overview of ResNet-50 + CBAM Teacher Network Architecture

[0058]

[0059] Table 2: Overview of MobileNetV3-Small Student Network Architecture

[0060]

[0061] The two networks achieve feature fusion through multi-level knowledge distillation, including:

[0062] Feature layer distillation: The similarity between student network feature maps and teacher network feature maps is constrained by L2 loss.

[0063] Output layer distillation: KL divergence loss is used to align the output distributions of the two networks.

[0064] Relationship layer distillation: Maintaining consistency in the relative spatial relationships of acupoints output by the two networks.

[0065] Specifically, the clinical knowledge constraint perception module includes an anatomical constraint submodule and a clinical rule submodule. The clinical knowledge constraint perception module is the core constraint unit of the positioning model. The anatomical constraint submodule is constructed based on spatial positioning data from a multimodal acupoint database, integrating the spatial distribution patterns of the posterior sacral foramen, symmetry constraints, and the relative distance / angle relationship between acupoints and anatomical landmarks (posterior superior iliac spine, sacral angle). The clinical rule submodule integrates acupuncture depth thresholds for the eight acupoints (superficial < 20mm, middle 20-40mm, deep ≥ 40mm), positioning adjustment coefficients for special populations (obese / skeletal variation patients), and contraindication association rules (e.g., deep positioning is prohibited for patients with sacral fractures).

[0066] The constraint mechanism of the clinical knowledge constraint perception module is as follows:

[0067] (1) Spatial symmetry and sequence constraint. This constraint maintains the left-right symmetry of human anatomical structures and the basic vertical arrangement order of the upper, middle, and lower liao points by multiplying the average symmetry deviation and sequence integrity. Its mathematical expression is as follows:

[0068] (1)

[0069] In the formula, This represents the averaging coefficient, which is the average value of the symmetry deviations of the four acupoints. This indicates the summation over the four acupoints (liao). This represents the coordinates of the i-th acupoint on the left, which is one of the eight acupoints on the left (i=1,2,3,4 correspond to Shangliao, Ciliao, Zhongliao, and Xialiao respectively). This represents the coordinates of the i-th acupoint on the right side, which is one of the eight acupoints on the right side (i=1,2,3,4 correspond to Shangliao, Ciliao, Zhongliao, and Xialiao respectively). The midline of the sacrum is the midline of the sacral region of the human body, serving as a reference line for left-right symmetry. This represents the symmetry tolerance threshold, the maximum allowable symmetry deviation, typically set to 2.5-5 mm. This indicates an indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise. This represents the ordinate of the i-th acupoint, which is the coordinate value of the acupoint in the vertical direction (head-to-foot direction). This represents a sequence constraint condition, ensuring the order from top to bottom: Upper Liao → Secondary Liao → Middle Liao → Lower Liao, with the vertical axis decreasing in value; This represents the distance from the i-th acupoint on the left to the midline. The Euclidean distance is used to calculate the vertical distance from the acupoint to the midline. This represents the distance from the i-th acupoint on the right side to the midline. The Euclidean distance is used to calculate the vertical distance from the acupoint to the midline. This indicates that the multiplication operation is performed on adjacent acupoint pairs to ensure that all adjacent acupoints satisfy the order constraint.

[0070] (2) Geometric relationship constraint between adjacent acupoints. This constraint, through a combination of standardized adjacent spacing deviation and normalized collinearity deviation, specifically constrains the spatial relationship between adjacent acupoints, promoting a reasonable linear arrangement of the eight acupoints and ensuring that the adjacent spacing basically conforms to anatomical statistical laws. Its mathematical expression is as follows:

[0071] (2)

[0072] In the formula, This represents the averaging coefficient, which is the average of the deviations in the spacing between three adjacent acupoint pairs. This represents the summation of three adjacent acupoint pairs; This indicates the summation over the four acupoints (liao). This represents the coordinates of the i-th acupoint, which is either on the left or right side (i=1,2,3,4). This represents the coordinate position of the (i+1)th acupoint; The actual Euclidean distance between adjacent acupoints is used to calculate the spatial distance between the i-th and (i+1)-th acupoints. This represents the reference spacing of the i-th adjacent acupoint pair, which is a standard adjacent spacing obtained based on anatomical statistics (the reference value is different for different acupoint pairs). This represents the spacing ratio, the ratio of the actual spacing to the reference spacing, ideally close to 1. This indicates the standardized spacing deviation, which measures the relative deviation (absolute value) between the actual spacing and the standard spacing. The best-fit line is the fitted line obtained by linear regression through all acupoints. This represents the distance from the i-th acupoint to the fitted straight line, and the perpendicular distance from the point to the straight line, which measures the degree to which the acupoint deviates from the linear arrangement. This represents the maximum allowable collinearity deviation, the maximum tolerable distance of the acupoint from the fitted straight line, used for normalization; The collinearity weighting coefficient, the weighting parameter for the balancing spacing constraint and the collinearity constraint, is usually set to 0.3; This represents the consistency threshold for adjacent relationships, which is the maximum allowable comprehensive deviation threshold for the entire constraint condition, and is usually set to 5.0-10mm.

[0073] (3) Individualized body measurement constraint, which ensures that the absolute positional relationship between each acupoint and the sacral midline conforms to individual anatomical characteristics, and its mathematical expression is:

[0074] (3)

[0075] In the formula, Let represent the normalized importance weight of the i-th acupoint, reflecting the clinical importance of different acupoints, and satisfying . Of these, the second thoracic vertebra has the highest weight. This indicates an indicator function that returns 1 if the acupoint satisfies the individualized body measurement constraint, and 0 otherwise. This represents the actual distance from the i-th acupoint to the midline, and the distance from the acupoint location predicted by the system to the sacral midline. This represents the reference distance from the i-th acupoint to the midline under standard body type, a reference distance based on standard anatomical data (reference values ​​differ for different acupoints); This represents the body-size scaling factor, an individualized scaling factor calculated based on the patient's body characteristics (such as the distance between the iliac crests); This represents the individualized reference distance, which is obtained by multiplying the standard reference distance by the individual scaling factor to get the reference distance adapted to the patient's body size. This represents the absolute deviation, the absolute difference between the actual distance and the individualized reference distance; The relative deviation rate is obtained by dividing the absolute deviation by the reference distance, resulting in a standardized relative deviation rate (dimensionless). This represents the weighted voting score, and the weighted proportion of acupoints that satisfy the constraints is calculated. This represents the tolerance threshold for individualized deviation, the maximum allowable relative deviation rate, typically set at 15%-20%. This represents the threshold for consistency of body measurements, which is the minimum weighted percentage required to meet the requirement. It is usually set to 0.8-0.9 (meaning that at least 80%-90% of the weighted acupoints must meet the requirement).

[0076] In this embodiment, the aforementioned multimodal acupoint database includes basic acupoint data and spatial positioning data.

[0077] Specifically, the process of obtaining the basic acupoint data is as follows:

[0078] A1. Acquire multiple sets of pelvic CT images and body surface RGB images;

[0079] A2. Preprocess and label acupoint information on the pelvic CT images and body surface RGB images obtained in A1. The labeled acupoint information includes the standard name, anatomical location description, main functions and indications of the Eight Liao Acupoints.

[0080] Preferably, the process for acquiring the above spatial positioning data is as follows:

[0081] B1. Obtain three-dimensional data of the human sacral region using a three-dimensional scanning device;

[0082] B2. Label the three-dimensional coordinates of the eight acupoints (Baliao) and their relative distances and angles with anatomical landmarks (such as the posterior superior iliac spine and sacral angle).

[0083] S4. The trained teacher-student network model is integrated into the mobile APP system.

[0084] Specifically, this involves mobile deployment and visualization integration: the trained student network model is converted to a format supported by a mobile inference framework (such as NCNN) via the ONNX intermediate format and deployed on a mobile app; real-time RGB images are captured through the mobile camera, and the coordinates of the eight acupoints are obtained through forward inference of the model. The coordinates are then overlaid on the real-time images through coordinate transformation, and the names, location information, and acupuncture depth prompts of the eight acupoints are output (shallow layers are marked in green, middle layers in yellow, and deep layers in red).

[0085] The mobile visualization and interaction module is used to display location results and facilitate human-computer interaction on the mobile app, and supports real-time guidance of acupuncture operations.

[0086] S4.1 Model Format Conversion: Export the trained student network model (MobileNetV3 architecture) to the ONNX universal format, optimize the ONNX model using a model simplification tool to remove redundant operators, and then convert the ONNX model to a format supported by mobile inference frameworks (such as NCNN's param and bin files) using a format conversion tool.

[0087] S4.2 Mobile App Development Environment Configuration: Configure the deep learning inference framework library files in the mobile development environment, and import the converted model parameter files and weight files; implement interactive calls between the Java layer and the C++ inference layer through the JNI interface.

[0088] S4.3 Image Acquisition and Preprocessing: Real-time RGB images are acquired through the mobile phone camera, the video stream is decomposed into image frames, and preprocessing operations such as size normalization and color space conversion are performed on the images to convert them into the data format required for model input.

[0089] S4.4 Model Inference and Coordinate Output: The preprocessed image data is input into the mobile inference framework for forward propagation calculation to obtain the prediction results of the location of the eight acupoints.

[0090] S4.5 Visual Overlay Display: The predicted coordinates of the eight acupoints are overlaid on the real-time RGB image through coordinate transformation, and different colors are used to distinguish each acupoint (Shangliao, Ciliao, Zhongliao, Xialiao). At the same time, the acupoint name, location information and acupuncture depth prompts are displayed (shallow layer is marked with green, middle layer with yellow, and deep layer with red).

[0091] S4.6 Real-time rendering and performance optimization: OpenGL ES is used for real-time rendering. Through optimization strategies such as model quantization and multi-threaded parallelism, the inference frame rate is ensured to meet the real-time requirements. Kalman filtering is used to eliminate coordinate jitter and ensure stable display of acupoint annotations.

[0092] When using the intelligent and precise positioning method for the eight acupoints described in this embodiment:

[0093] Training data came from subjects enrolled at Shanxi Provincial Hospital of Traditional Chinese Medicine. Exclusion criteria included variations in the number of sacral vertebrae (4 / 6 vertebrae) and fusion variations (lumbosacral fusion, sacrococcygeal fusion, etc.). Data annotation was performed by three acupuncture experts with the title of associate chief physician or above, who independently annotated the locations of the eight Liao acupoints according to the "National Standard for Acupoint Location (GB / T 12346-2021)". The consistent region of the annotation results was used as the gold standard. Data augmentation employed random horizontal flipping (probability 0.5), color jitter (brightness 0.2, contrast 0.2, saturation 0.2), and random cropping (scaling ratio 0.8-1.2) strategies.

[0094] Teacher network pre-training: Pre-training for 50 epochs on CT 3D reconstruction data using mean squared error loss.

[0095] Student network distillation training: The teacher network parameters are fixed, and the student network is trained by combining soft target loss and hard target loss.

[0096] Overall fine-tuning: The entire system is fine-tuned using a cosine annealing strategy with an initial learning rate of 0.001.

[0097] Performance verification and evaluation solution:

[0098] The system was validated using data from 300 multicenter cases, including 200 cases from Shanxi Provincial Hospital of Traditional Chinese Medicine and 100 cases from collaborating hospitals.

[0099] Clinical validation was conducted by 10 experts at the level of associate chief physician or above using a blinded method, with the traditional anatomical localization method as a control. An error of ≤10mm was considered clinically acceptable.

[0100] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been provided with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.

Claims

1. An intelligent and accurate positioning method for eight-liao points based on clinical knowledge and multi-modal image data joint driving, characterized in that, Includes the following steps: S1. Collect multimodal image data of the patient's sacral region to obtain a multimodal image database, which includes pelvic CT images and body surface RGB images; S2. Preprocess the multimodal image data obtained in S1; S3. Extract multimodal fusion features from the teacher-student network architecture in the trained localization model, and generate spatial constraint parameters by combining them with the clinical knowledge constraint perception module; S4. The trained teacher-student network model is integrated into the mobile APP system; The constraint mechanism of the clinical knowledge constraint perception module is as follows: (1) Spatial symmetry and sequence constraint. This constraint maintains the left-right symmetry of human anatomical structures and the basic vertical arrangement order of the upper, middle, and lower liao points by multiplying the average symmetry deviation and sequence integrity. Its mathematical expression is as follows: (1) In the formula, Represents the averaging coefficient. This indicates the summation over the four acupoints. This indicates the coordinate position of the i-th acupoint on the left. This indicates the coordinate position of the i-th acupoint on the right. Indicates the sacral midline. This represents the symmetry tolerance threshold. Indicates an indicator function, This represents the ordinate of the i-th acupoint. This represents a sequence constraint condition. This represents the distance from the i-th acupoint on the left to the midline. This represents the distance from the i-th acupoint on the right side to the midline. This represents a series of multiplication operations; (2) Geometric relationship constraint between adjacent acupoints. This constraint, through a combination of standardized adjacent spacing deviation and normalized collinearity deviation, specifically constrains the spatial relationship between adjacent acupoints, promoting a reasonable linear arrangement of the eight acupoints and ensuring that the adjacent spacing basically conforms to anatomical statistical laws. Its mathematical expression is as follows: (2) In the formula, Represents the averaging coefficient. This indicates the summation of three adjacent acupoint pairs. This indicates the summation over the four acupoints. This represents the coordinates of the i-th acupoint. This represents the coordinate position of the (i+1)th acupoint. The actual Euclidean distance between adjacent acupoints. This represents the reference spacing between the i-th adjacent acupoint pairs. Indicates the spacing ratio. Indicates the standardized spacing deviation. This represents the best-fitting straight line. This represents the distance from the i-th acupoint to the fitted straight line. This indicates the maximum permissible collinearity deviation. Collinearity weighting coefficients Indicates the threshold for consistency in adjacency relationships; (3) Individualized body measurement constraint, which ensures that the absolute positional relationship between each acupoint and the sacral midline conforms to individual anatomical characteristics, and its mathematical expression is: (3) In the formula, This represents the normalized importance weight of the i-th acupoint. Indicates an indicator function, This represents the actual distance from the i-th acupoint to the midline. This represents the reference distance from the i-th acupoint to the midline under standard body type. Indicates the scaling factor for body size. Indicates the individualized reference distance. Indicates absolute deviation. This represents the relative deviation rate. This represents the weighted voting score. This represents the tolerance threshold for individualized bias. This indicates the threshold for consistency of body dimensions.

2. The method of claim 1, wherein the method is based on a combination of clinical knowledge and multi-modal image data to intelligently and precisely locate the eight Liao points. The multimodal image database includes basic acupoint data and spatial positioning data. The process for acquiring the basic acupoint data is as follows: A1. Acquire multiple sets of pelvic CT images and body surface RGB images; A2. Preprocess and label acupoint information on the pelvic CT images and body surface RGB images obtained in A1. The labeled acupoint information includes the standard name, anatomical location description, main functions and indications of the Eight Liao Acupoints. The process of acquiring the spatial positioning data is as follows: B1. Obtain three-dimensional data of the human sacral region using a three-dimensional scanning device; B2. Label the three-dimensional coordinates of the eight acupoints, their relative distances and angular relationships with anatomical landmarks, and the three-dimensional data.

3. The method of claim 1, wherein the method is characterized by, In the teacher-student network architecture, the teacher network is based on pelvic CT images using a ResNet-50 backbone and integrates the CBAM attention mechanism. It focuses on learning the precise spatial location features of the posterior sacral foramen, and the output features are synchronously transmitted to the anatomical constraint submodule of the clinical knowledge constraint perception module for spatial parameter verification. The student network is based on RGB images using a lightweight MobileNetV3 architecture. It optimizes inference speed through depthwise separable convolution, and its output features must meet the individualized dimensional constraints of the clinical rule submodule. The teacher network and the student network achieve feature fusion through multi-level knowledge distillation.

4. The method of claim 1, wherein, The trained localization model is converted into a format supported by the mobile inference framework using a model conversion tool and then deployed on a mobile app. Real-time RGB images are captured by a mobile camera, and the locations of the eight acupoints are obtained through model inference. These locations are then displayed on the real-time images, and the names, locations, and acupuncture depth of the eight acupoints are output.

Images