Individualized back acupuncture point positioning method

Abstract

The invention relates to an individualized back acupuncture point positioning method. The method comprises the following steps: identifying key positioning points on an individual image; globally aligning the identified key positioning points with corresponding point positions on a standard model by adopting a similarity transformation method; if the alignment error after global alignment is greater than a threshold value, adjusting the position of the key positioning point on the individual image by adopting a thin plate spline deformation method; and generating a traditional Chinese medicine cun and image pixel conversion scale based on a traditional Chinese medicine bone degree reference, thereby mapping acupoint distribution on the standard model to an individual image by taking the key positioning point as an anchor point. According to the method, the stability and accuracy of key point detection can be improved, the labeling cost is low, and the individual adaptability is high.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and digital TCM technology, and in particular to a personalized method for locating acupoints on the back. Background Technology

[0002] Precise location of acupuncture points is a crucial aspect of clinical and rehabilitation treatment in Traditional Chinese Medicine. Currently, automatic location technology for back acupuncture points mainly falls into two categories:

[0003] 1. Geometric estimation methods based on human key points: These methods typically use human pose estimation models (such as OpenPose, HRNet, etc.) to obtain the standard anatomical proportions of human key point locations and estimate individual acupoint locations. However, human key point detection has certain errors under different poses, lighting, and occlusion conditions, and these errors are further propagated to the acupoint estimation results, resulting in insufficient accuracy for some acupoints. In addition, relying solely on simple proportional relationships cannot effectively adapt to the non-uniform differences in individual skeletal structure or body shape, and the estimation results have limited consistency with the actual individual.

[0004] 2. Direct acupoint detection methods based on deep learning: These methods attempt to use deep learning models (such as convolutional neural networks, Transformers, etc.) to directly regress acupoint coordinates from images. Although theoretically they can bypass the key point detection process, in practical applications, they suffer from high annotation costs, poor standardization, and thus affect model convergence and generalization. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an individualized method for locating acupoints on the back, which can improve the stability and accuracy of key point detection, and has low annotation cost and strong individual adaptability.

[0006] The technical solution adopted by this invention to solve its technical problem is: to provide an individualized method for locating acupoints on the back, comprising:

[0007] Identify key localization points on individual images;

[0008] The similarity transformation method is used to globally align the identified key positioning points with the corresponding points on the standard model;

[0009] If the alignment error after global alignment is greater than the threshold, the position of key positioning points on the individual image is adjusted using the thin plate spline deformation method.

[0010] Based on the TCM bone measurement benchmark, a conversion scale between TCM inches and image pixels is generated, thereby mapping the distribution of acupoints on the standard model to individual images using key positioning points as anchor points.

[0011] Furthermore, the method of adjusting the position of key positioning points on an individual image using thin-plate spline deformation includes:

[0012] Using key localization points on individual images as target points and corresponding points on the standard model as control points, a thin plate spline deformation function containing a global affine part and a local deformation part is constructed.

[0013] Construct a system of linear equations that satisfy energy minimization and control point constraints, and solve it to obtain the parameters of the global affine part and the local deformation part.

[0014] Furthermore, the key positioning points include skeletal or muscular features in the scapular and spinal regions that can be visually identified.

[0015] Furthermore, the step of generating a conversion scale between traditional Chinese medicine inches and image pixels based on traditional Chinese medicine bone measurement benchmarks includes:

[0016] Using the distance between the inner edges of the left and right scapulae as a reference distance, calculate the first conversion scale based on the scapular region skeletal measurement benchmark;

[0017] Using the distance from the spinous process of the cervical vertebrae to the sacrum of the lumbar vertebrae as a reference distance, a second conversion scale based on the longitudinal bone measurement benchmark of the spine is calculated.

[0018] Calculate the weighted sum of the first and second transformation scales.

[0019] Furthermore, the weights of the first and second conversion scales are adjusted according to the different back regions where the acupoints are located.

[0020] Furthermore, for acupoints on the upper back, the weight of the first conversion scale is greater than the weight of the second conversion scale; for acupoints on the lower back, the weight of the first conversion scale is less than the weight of the second conversion scale.

[0021] Furthermore, the step of mapping the acupoint distribution on the standard model to individual images using key positioning points as anchor points includes:

[0022] The rules of traditional Chinese medicine meridians are quantified into pixel offsets on individual images using a transformation scale.

[0023] Using key positioning points as anchor points, the position of the corresponding acupoint on the individual image is obtained based on the calculated pixel offset.

[0024] Furthermore, the step of using a similarity transformation method to globally align the identified key positioning points with their corresponding points on the standard model includes:

[0025] With the goal of minimizing transformation error, the corresponding points on the standard model are transformed to the positions of key localization points in the image by adjusting scaling, rotation and translation.

[0026] Solve for the rotation matrix, scaling factor, and translation matrix that satisfy the objective.

[0027] Furthermore, the identification of key localization points in an individual image includes:

[0028] Image enhancement is performed on the collected individual images, and heat maps are generated for each enhanced image.

[0029] The generated heatmaps are merged into a single detection map, which is then input into the target recognition model to predict key location points.

[0030] Furthermore, the standard model is constructed based on the standard human back contour and defines key positioning points, normalized coordinates of acupoints, and geometric relationships between acupoints and their corresponding key positioning points.

[0031] Furthermore, the alignment error is the average value of the error between each key positioning point and the corresponding point on the standard model.

[0032] Beneficial effects

[0033] Due to the adoption of the above-mentioned technical solutions, this invention has the following advantages and positive effects compared with the prior art: This invention utilizes the geometric relationships of the standard acupoint model on the back and achieves adaptive alignment between the model and individual human images through similarity transformation and local deformation methods (such as thin plate spline TPS). Simultaneously, it incorporates the traditional Chinese medicine "bone measurement method" and introduces a pixel-to-inch conversion mechanism, ensuring that the calculated results not only conform to individual body shape characteristics but also to the framework of traditional medical theory. This invention, through geometric and symmetry constraints after the candidate acupoints are calculated, can effectively reduce initial calculation errors, making the final acupoint coordinates more accurate. This invention, through ViTPose… Introducing Test-Time Enhancement (TTA) during keypoint detection effectively reduces prediction fluctuations caused by pose, lighting, and occlusion, improving the stability and accuracy of keypoint detection and thus reducing the cumulative effect of errors in subsequent acupoint estimation. This invention does not require a large amount of detailed and difficult-to-standardize acupoint annotation data. It utilizes features of the back that are relatively obvious and can be identified by the naked eye, such as bones and muscles, as keypoints, significantly reducing the annotation difficulty. In addition, because it mainly relies on the geometric relationship between keypoint detection and the standard model, it avoids the model instability problem caused by excessive annotation variance, significantly reducing the cost of data collection and annotation. The method of this invention is applicable to the back, limbs, and trunk regions of different individuals and has good generalization ability. The standard acupoint model can be extended to more meridians and acupoints, and can be quickly adapted by simply supplementing the model library. Attached Figure Description

[0034] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0035] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0036] The embodiments of the present invention relate to an individualized method for locating acupoints on the back, such as... Figure 1 As shown, it includes:

[0037] Identify key localization points on individual images;

[0038] The similarity transformation method is used to globally align the identified key positioning points with the corresponding points on the standard model;

[0039] If the alignment error after global alignment is greater than the threshold, the position of key positioning points on the individual image is adjusted using the thin plate spline deformation method.

[0040] Based on the TCM bone measurement benchmark, a conversion scale between TCM inches and image pixels is generated, thereby mapping the distribution of acupoints on the standard model to individual images using key positioning points as anchor points.

[0041] Key location points can be identified by visually recognizable skeletal or muscular features in the scapular and spinal regions, such as the shoulder, hip, cervical spinous processes, acromion, inferior angle of the scapula, and posterior superior iliac spine. This implementation method is also applicable to the back, limbs, and trunk regions of different individuals. The standard acupoint model can be expanded to include more meridians and acupoints, requiring only the addition of a model library for rapid adaptation.

[0042] The core objective of similarity transformation is to transform the individual key points of the standard model (such as standard shoulder and hip coordinates) to the positions of individual key points (such as the actual detected shoulder and hip coordinates) by adjusting scaling, rotation, and translation, thereby minimizing the positional error between the two.

[0043] Thin Plate Spline (TPS) deformation and similarity transformation alignment are two different methods in geometric registration. The core difference lies in the flexibility of the transformation and the ability to handle local deformations. Specifically, these include:

[0044] Using key localization points on individual images as target points and corresponding points on the standard model as control points, a thin plate spline deformation function containing a global affine part and a local deformation part is constructed.

[0045] Construct a system of linear equations that satisfy energy minimization and control point constraints, and solve it to obtain the parameters of the global affine part and the local deformation part.

[0046] After the preliminary alignment of the standard model and the individual is achieved through similarity transformation, according to the evaluation result of the alignment error, TPS is used for local non-linear deformation to handle the non-uniform differences, obtaining more accurate key positioning points on the individual image, thereby reducing the cumulative deviation caused by the key point detection error. Further, by calculating pixel values, the traditional Chinese medicine "cun" is converted into image pixels, and the three together ensure that the acupoint positioning not only conforms to the traditional theory but also adapts to individual differences.

[0047] A preferred embodiment 1 of the present invention provides a method based on ViTPose human key point detection combined with test-time augmentation (TTA), aiming to improve the stability and accuracy of key point detection. This embodiment combines the geometric relationship of the standard acupoint model on the back, and maps the standard acupoints to the individual human body image through similarity transformation and local deformation methods. At the same time, the bone measurement conversion mechanism of "cun" is introduced to realize the correspondence between image pixels and the traditional Chinese medicine cun method, so as to ensure that the acupoint calculation conforms to the traditional medical theory. Based on the preliminary calculation result, a local optimization and fine-tuning mechanism is adopted to further reduce the cumulative deviation caused by the key point detection error. Finally, an automatic acupoint positioning method for the human back with low annotation cost, strong individual adaptability, conforming to traditional Chinese medicine theory, and having high precision and stability is realized.

[0048] The overall system architecture consists of the following functional modules:

[0049] Key point detection module: Based on the ViTPose model and combined with TTA technology, output the high-precision coordinates of human key points;

[0050] Standard back acupoint model library: Store the standard acupoint distribution on the human back, including standard coordinates and geometric relationships;

[0051] Geometric alignment module: Use the matching of human key points and standard anchor points to achieve the alignment and deformation of the standard model and the individual image;

[0052] Cun method estimation module: Calculate the conversion ratio between image pixels and "cun" according to the human bone measurement scale relationship;

[0053] Acupoint positioning module: Combine the standard model mapping and cun method conversion to calculate the acupoint coordinates in the test human body image;

[0054] Result visualization and interaction module: Display the acupoint points and support manual fine-tuning and global linkage update.

[0055] The specific process includes:

[0056] Step S1: Key point detection and enhancement;

[0057] Step S2: Establishment of the standard back acupoint model;

[0058] Step S3: Individualized geometric alignment;

[0059] Step S4: Calculate the pixel value using the inch method;

[0060] Step S5: Calculation of candidate acupoints;

[0061] Step S6: Output and interaction of results.

[0062] The following section provides a detailed explanation of each step.

[0063] Step S1: Keypoint Detection and Enhancement

[0064] 1) Input a human body RGB image

[0065] Keypoint detection using the ViTPose model yields a set of coordinates for keypoints including the shoulder, hip, cervical spinous processes, acromion, inferior angle of the scapula, posterior superior iliac spine, and medial border of the scapula. Besides ViTPose, other human pose estimation models can also be used, such as HRNet and various keypoint detection networks based on the MMPose framework. Keypoint detection results can be geometrically mapped and converted to proportions using standard acupoint models to obtain individualized acupoint locations. In addition to the keypoints described above, the following can also be used:

[0066] Along the spine: C7, T3, T7, T12, L4 / L5;

[0067] Scapula: superior angle, lateral angle, midpoint of the spine;

[0068] Sacrum / pelvis: Sacral cleft, PIIS;

[0069] 2) To enhance robustness, the Test-Time Enhancement (TTA) method is adopted:

[0070] ① Process the input image

[0071] a. Rotation: In most public datasets, the human body is standing in the image, while the human body in physical therapy is horizontal in the image, which is not conducive to annotation and pre-training using public datasets.

[0072] b. Horizontal flip: This is very important for improving the symmetry and stability of key points;

[0073] c. Scaling: Adjust the image to the input size that the model can accept;

[0074] ② Input the enhanced images into ViTPose to obtain multiple heatmaps.

[0075] ③ Align heatmaps, for example, rotate the heatmap obtained from the rotated enhanced image back to the original orientation to obtain a new heatmap; average the heatmaps at the same location from multiple heatmaps and decode them to obtain prediction results, including key point coordinates and key point confidence;

[0076] ④ In addition to horizontal flipping, scaling, and rotation, color dithering and multi-scale input can be used to enhance the stability of keypoint detection; model ensemble can also be used to replace TTA to improve the robustness of keypoint prediction.

[0077] Step S2: Establishment of a standard back acupoint model

[0078] 1) Construct a standard human acupoint model (Atlas) and store it in JSON or database format:

[0079] ① Construct the standard human body silhouette from the back;

[0080] ② Define standard anchor points (such as shoulder, hip, cervical spinous process, acromion, inferior angle of scapula, posterior superior iliac spine, etc.);

[0081] ③ Define the normalized coordinates of each common acupoint and their geometric relationship with the anchor point (e.g., "1.5 cun lateral to the center" or "3 cun lateral to the center").

[0082] Step S3: Individualized Geometric Alignment

[0083] 2) Extract the corresponding anchor points (shoulder, hip, cervical spinous process, acromion, inferior angle of scapula, posterior superior iliac spine, etc.) from the human body key points output by ViTPose.

[0084] 3) Use similarity transformations (rotation, scaling, translation) to align the standard model anchor points to the individual key points:

[0085] ① X={ } is the anchor point for human body keypoint detection, Y={ } are the anchor points of the standard human acupoint model. Calculate y = sRx + t, where R is the orthogonal rotation matrix, s is the scaling scalar, and t is the translation vector. Solve for the objective function of the similarity transformation:

[0086]

[0087] The goal is to find R, s, and t such that the predicted location is as close as possible. ;

[0088] ② Calculate the covariance:

[0089] a. Calculate the mean:

[0090]

[0091] It is the average position of the standard anchor points. It is the average position of individual key points;

[0092] b. Data centralization:

[0093]

[0094] c. Covariance calculation:

[0095]

[0096] ③ Solving the objective function:

[0097] a. Calculate the rotation matrix:

[0098]

[0099]

[0100]

[0101] D is a diagonal matrix, and S is a sign matrix;

[0102] b. Calculate the scaling factor:

[0103]

[0104] The variance of X after centering is used to describe the dispersion of the anchor points in the standard model. diagonal matrix The trace, s, represents scaling the dispersion of the standard model to match the dispersion of individual keypoints;

[0105] c. Calculate the translation matrix:

[0106]

[0107] t is a translation vector used to move the transformed standard model as a whole to the position of individual key points;

[0108] 4) If the pose differences are large, thin plate spline (TPS) deformation can be used to flexibly map the global acupoint distribution of the standard model onto the individual image; in addition to similarity transformation and thin plate spline (TPS) deformation, affine transformation, cubic spline interpolation or deformation field estimation methods based on deep learning can also be used; a 3D standard human body model can also be introduced, and more accurate individual acupoint mapping can be obtained through three-dimensional registration and projection;

[0109] ① Evaluate the similarity transformation error:

[0110] a. Calculate the residual (Euclidean distance) between the standard human back anchor point and the target key point after alignment. If the average residual is less than the threshold (10 px or a certain absolute / relative threshold), then only similarity transformation is used.

[0111] ② Solve for 2D TPS, using key localization points on individual images as target points Y={ }, with the corresponding points on the standard model as control points C={ }, construct the TPS mapping expression:

[0112]

[0113] in It is a translation term, A is an analogous linear term. Let U(r) be the weight vector, and U(r) be the basis functions. Commonly used in two dimensions:

[0114]

[0115] b. Definition These are the basis function values ​​between control points, used to describe the mutual influence between them:

[0116] c. Define P to constrain the degrees of freedom of the affine part: P=[1, , The first column is all 1s, and the last two columns are the x and y coordinates of the control points, respectively.

[0117] d. Construct a large matrix L:

[0118]

[0119] λ is a regularization parameter used to improve the stability of matrix calculations, I is the identity matrix (a square matrix with all diagonal elements being 1 and all other elements being 0), and V is a vector used to constrain the translation and rotation of the affine part.

[0120] e. Solving linear systems:

[0121] f. Unknown parameter vector:

[0122]

[0123] Once W is obtained, f(x) can be calculated for any standard acupoint x to obtain the current human acupoint being detected.

[0124] Step S4: Calculation of pixel values ​​in inch method

[0125] 1) Calculate the pixel / inch (ppc) value based on the bone measurement benchmark in the "Jingchuancun Method":

[0126] ① Based on the scapular region's bony benchmarks:

[0127] a. The distance between the inner edges of the left and right scapulae is used as a reference distance: 6 inches;

[0128] b. Calculate pixel distance using the left and right scapular medial points obtained during the keypoint detection phase. ;

[0129] c. = ;

[0130] ② Based on the longitudinal bone measurement benchmark of the spine:

[0131] a. Reference distance from the spinous process of the cervical vertebrae to the sacrum of the lumbar vertebrae: 18 inches;

[0132] b. Detect the pixel distance from the highest point of the neck to the midpoint of the line connecting the posterior superior iliac spine: ;

[0133] c. ;

[0134] ③ Weighted calculation:

[0135] a. A single benchmark may be affected by individual pose or detection errors, therefore a weighted average can be used:

[0136] b. , α∈[0,1];

[0137] ④ In addition to directly converting "pixels / inch" through the distance of key points, you can also:

[0138] a. 3D human point cloud obtained by depth camera, converted by actual physical scale;

[0139] b. Combine wearable sensors or ranging devices to obtain individual bone length, and then map it onto an image;

[0140] ⑤ The weight α is selected based on the task area:

[0141] a. Acupoints on the upper back (such as Fengmen and Feishu) are more dependent on the interscapular distance, α≈0.7;

[0142] b. Lower back acupoints (such as Shenshu and Dachangshu) are more dependent on the longitudinal direction of the spine, α≈0.3.

[0143] Step S5: Calculation of candidate acupoints

[0144] 1) Map the standard model onto the individual image to obtain the initial acupoint coordinates;

[0145] 2) For the acupoints of the Bladder Meridian on the back, the following geometric calculations are used:

[0146] ① Using the midline of the spine as a reference, calculate the position of the corresponding horizontal line of the thoracic vertebrae;

[0147] ② On the horizontal line, using the rule of "1.5 inches / 3 inches to the side", Calculate the pixel offset;

[0148] ③ Obtain symmetrical acupoints on both sides;

[0149] 3) For acupoints on the shoulder, such as Jianjing, Bingfeng, and Naoshu, use key points and geometric relationships to calculate their candidate locations;

[0150] 4) In addition to optimization based on geometric constraints, the following can also be used:

[0151] ① Use a small regression network based on deep learning to perform local fine regression near candidate points;

[0152] ② Graph Convolutional Networks (GCNs) are used to model the topological relationships between adjacent acupoints and optimize the acupoint locations as a whole.

[0153] Step S6: Result Output and Interaction

[0154] 1) Overlay the final acupoint coordinates onto the original image for visualization;

[0155] 2) Provide interactive operation:

[0156] ① Users can drag anchor points or acupoints;

[0157] ② The system automatically updates the positions of other acupoints based on TPS or meridian geometry;

[0158] 3) In addition to image visualization, acupoints can also be displayed in real time through three-dimensional virtual human body models or augmented reality (AR) devices; users can interactively adjust the position of acupoints through voice commands or body-sensing operations, and the system will automatically complete global linkage updates.

Claims

1. A personalized method for locating acupoints on the back, characterized in that, include: Identify key localization points on individual images; The similarity transformation method is used to globally align the identified key positioning points with the corresponding points on the standard model; If the alignment error after global alignment is greater than the threshold, the position of key positioning points on the individual image is adjusted using the thin plate spline deformation method. Based on the TCM bone measurement benchmark, a conversion scale between TCM inches and image pixels is generated, thereby mapping the distribution of acupoints on the standard model to individual images using key positioning points as anchor points.

2. The method according to claim 1, characterized in that, The method of adjusting the position of key positioning points on an individual image using thin plate spline deformation includes: Using key localization points on individual images as target points and corresponding points on the standard model as control points, a thin plate spline deformation function containing a global affine part and a local deformation part is constructed. Construct a system of linear equations that satisfy energy minimization and control point constraints, and solve it to obtain the parameters of the global affine part and the local deformation part.

3. The method according to claim 1, characterized in that, The key positioning points include skeletal or muscular features in the scapular and spinal regions that can be visually identified.

4. The method according to claim 3, characterized in that, The method for generating a conversion scale between traditional Chinese medicine inches and image pixels based on traditional Chinese medicine bone measurement benchmarks includes: Using the distance between the inner edges of the left and right scapulae as a reference distance, calculate the first conversion scale based on the scapular region skeletal measurement benchmark; Using the distance from the spinous process of the cervical vertebrae to the sacrum of the lumbar vertebrae as a reference distance, a second conversion scale based on the longitudinal bone measurement benchmark of the spine is calculated. Calculate the weighted sum of the first and second transformation scales.

5. The method according to claim 4, characterized in that, The weights of the first and second conversion scales are adjusted according to the different back regions where the acupoints are located.

6. The method according to claim 5, characterized in that, For acupoints on the upper back, the weight of the first conversion scale is greater than that of the second conversion scale; for acupoints on the lower back, the weight of the first conversion scale is less than that of the second conversion scale.

7. The method according to claim 1, characterized in that, The process of mapping the distribution of acupoints on the standard model to individual images using key positioning points as anchor points includes: The rules of traditional Chinese medicine meridians are quantified into pixel offsets on individual images using a transformation scale. Using key positioning points as anchor points, the position of the corresponding acupoint on the individual image is obtained based on the calculated pixel offset.

8. The method according to claim 1, characterized in that, The step of using a similarity transformation method to globally align the identified key positioning points with their corresponding points on the standard model includes: With the goal of minimizing transformation error, the corresponding points on the standard model are transformed to the positions of key localization points in the image by adjusting scaling, rotation and translation. Solve for the rotation matrix, scaling factor, and translation matrix that satisfy the objective.

9. The method according to claim 1, characterized in that, The identification of key localization points in individual images includes: Image enhancement is performed on the collected individual images, and heat maps are generated for each enhanced image. The generated heatmaps are merged into a single detection map, which is then input into the target recognition model to predict key location points.

10. The method according to claim 1, characterized in that, The alignment error is the average value of the error between each key positioning point and the corresponding point on the standard model.

Images