A Machine Learning-Based Method and System for Outputting Case Studies in Traditional Chinese Medicine Acupuncture

Abstract

This invention discloses a method and system for analyzing and outputting TCM acupuncture cases based on machine learning, belonging to the field of medical image processing technology. The method acquires facial image sequences from two pathological cycles of the subject, extracts spatiotemporal feature maps, and obtains local receptive field feature matrix pairs through sliding windowing. It calculates the cross-cycle joint entropy gradient and its matrix, and generates a two-dimensional deformation vector using optical flow. Using the coordinates of the first cycle as a reference, the deformation vector is set to zero when the joint entropy gradient does not exceed the rigidity threshold; otherwise, the joint entropy gradient matrix is ​​used as a weight matrix through a sigmoid activation function, and a Hadamard product is performed with the deformation vector to obtain a denoised deformation vector. The denoised vectors are superimposed to obtain corrected pixel coordinates, which are mapped to the target localization coordinate matrix, and the acupuncture case analysis results are output. This scheme achieves precise spatial decoupling of local tissue evolution, eliminates the contamination of rigid regions by global deformation, and significantly improves the robustness and accuracy of cross-cycle acupuncture target localization.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically, this application relates to a method and system for outputting TCM acupuncture case analysis based on machine learning. Background Technology

[0002] In the field of long-term facial image feature analysis of TCM acupuncture cases, the core task is to extract key features from non-standard facial image sequences of the subject at different time periods and accurately map them into spatial coordinate parameters to guide downstream execution agencies to complete the spatial positioning of target points.

[0003] Current conventional data processing solutions typically employ global spatiotemporal feature extraction and deformation compensation algorithms, whose underlying data flow logic is based on the rigid assumption that facial topological deformation has "spatial homogeneity" and "global continuity".

[0004] However, in actual long-term non-rigid deformation evolution, changes in facial tissue tension exhibit a high degree of local spatial heterogeneity. Specific areas undergoing local deformation may be experiencing drastic topological reshaping and feature displacement, while bony reference areas such as the forehead or bridge of the nose maintain high physical rigidity and morphological stability throughout the entire cycle. Existing image data processing techniques lack the ability to perceive the fine-grained topological activity of local features. Once the system enters the feature difference and coordinate mapping stage, it applies indiscriminate elastic calculation logic to the entire facial feature map. This forced calculation method forcibly involves the features of rigid regions that are originally in a static steady state into complex deformation compensation calculations, introducing unnecessary computational jitter and high-frequency mapping noise into the data features of stable regions.

[0005] Because existing technologies cannot identify and protect rigid feature anchors with spatial reference value during feature transfer, false computational features generated by local deformation will quickly pollute the global coordinate system under long-term continuous image processing conditions, causing continuous random oscillations and drifts of the static coordinate reference, ultimately causing the target point positioning parameters output by the model to deviate significantly from the real physical spatial location. Summary of the Invention

[0006] To address the aforementioned technical issues, this paper provides a method and system for outputting TCM acupuncture case analysis based on machine learning. This technical solution resolves the problems mentioned in the background section.

[0007] In a first aspect, embodiments of this application provide a method for analyzing and outputting TCM acupuncture cases based on machine learning, comprising the following steps: acquiring original facial image sequences of the target object in the first and second pathological cycles, and inputting them into a preset three-dimensional convolutional neural network for feature extraction, respectively obtaining spatiotemporal feature maps of the first and second pathological cycles, and performing sliding window cropping according to a preset step size to obtain feature matrix pairs of local receptive fields; performing probability density statistics on the pixel values ​​of the feature matrix pairs, and calculating the joint entropy gradient and joint entropy gradient matrix across cycles accordingly; calculating the pixel value displacement field of the feature matrix pairs using an optical flow algorithm, extracting the displacement direction and displacement amplitude to generate a two-dimensional deformation vector; and extracting the pixel coordinates of the spatiotemporal feature map of the first pathological cycle in the local receptive field as static... Based on the static coordinate reference, if the joint entropy gradient of the local receptive field is not greater than the rigidity threshold, then all two-dimensional deformation vectors contained in the local receptive field are reset to zero vectors; if the joint entropy gradient of the local receptive field is greater than the rigidity threshold, then the joint entropy gradient matrix of the local receptive field is calculated and constructed into a weight matrix through the sigmoid activation function, and the weight matrix is ​​performed with the two-dimensional deformation vectors in the local receptive field to obtain the denoised deformation vector; traversing all local receptive fields, the corresponding zero vectors or denoised deformation vectors are superimposed onto the static coordinate reference through vector addition to obtain the corrected pixel coordinate system, and the corrected pixel coordinate system is mapped through the preset camera intrinsic parameter matrix to obtain the target positioning coordinate matrix, and the TCM acupuncture case analysis results representing the acupuncture target coordinates are output accordingly.

[0008] Secondly, embodiments of this application provide a machine learning-based TCM acupuncture case analysis output system, including: a local receptive field extraction module: used to acquire the original facial image sequences of the target object in the first and second pathological cycles, and input them into a preset three-dimensional convolutional neural network for feature extraction, respectively obtaining the spatiotemporal feature maps of the first and second pathological cycles, and performing sliding window truncation according to a preset step size to obtain feature matrix pairs of local receptive fields; a joint entropy gradient extraction module: used to perform probability density statistics on the pixel values ​​of the feature matrix pairs, and calculate the joint entropy gradient and joint entropy gradient matrix across cycles accordingly; a two-dimensional deformation vector processing module: used to calculate the pixel value displacement field of the feature matrix pairs through an optical flow algorithm, extract the displacement direction and displacement amplitude to generate a two-dimensional deformation vector; and a static coordinate processing module: used to extract the local receptive field of the spatiotemporal feature map of the first pathological cycle. The pixel coordinates of the receptive field are used as the static coordinate reference. If the joint entropy gradient of the local receptive field is not greater than the rigid judgment threshold, all two-dimensional deformation vectors contained in the local receptive field are reset to zero vectors. The noise reduction deformation vector processing module is used to calculate and construct a weight matrix of the joint entropy gradient matrix of the local receptive field through the S-shaped activation function if the joint entropy gradient of the local receptive field is greater than the rigid judgment threshold. The weight matrix is ​​then used to perform a Hadamard product operation with the two-dimensional deformation vectors in the local receptive field to obtain the noise reduction deformation vector. The output module is used to traverse all local receptive fields, and to superimpose the corresponding zero vectors or noise reduction deformation vectors onto the static coordinate reference through vector addition to obtain the corrected pixel coordinate system. The corrected pixel coordinate system is then mapped through the preset camera intrinsic parameter matrix to obtain the target positioning coordinate matrix, and the TCM acupuncture case analysis results representing the acupuncture target coordinates are output accordingly.

[0009] Thirdly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described machine learning-based method for analyzing and outputting TCM acupuncture cases.

[0010] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0011] 1. To address the phenomenon of numerous rigid and stable regions in facial tissues during pathological evolution, this solution extracts the joint entropy gradient through probability density statistics and uses it as a criterion for local topological activity. When the joint entropy gradient is not greater than the rigidity threshold, the system actively resets the two-dimensional deformation vector contained within that local receptive field to a zero vector. This mechanism completely eliminates the meaningless computational noise introduced by the global deformation algorithm in stable regions, fundamentally avoiding the contamination of the static coordinate reference by high-frequency mapping noise, ensuring the stability of the underlying reference system for subsequent coordinate mapping, accurately isolating computational jitter, and guaranteeing the absolute stability of the rigid reference system.

[0012] 2. For active regions undergoing substantial topological reshaping, this scheme does not simply transmit the original displacement field. Instead, it utilizes a sigmoid activation function to transform the joint entropy gradient matrix into a weight matrix with spatial correspondence. By performing a Hadamard product operation between this weight matrix and the two-dimensional deformation vector within the local receptive field, this scheme achieves pixel-level nonlinear scaling and suppression based on the actual intensity of local deformation. This accurately filters out spurious feature associations generated by the optical flow algorithm at the reshaping edge, outputting a high-purity, denoised deformation vector. Consequently, it significantly improves the localization accuracy of complex pathological evolution features, achieves nonlinear feature suppression, and enhances the realism of deformation in active regions.

[0013] 3. This scheme abandons the traditional assumption of global consistency deformation and instead employs a stitching method that traverses all local receptive fields and independently superimposes zero vectors or denoised deformation vectors onto a static coordinate reference to construct a corrected pixel coordinate system that highly restores the true physiological characteristics. Subsequently, combined with a pre-defined camera intrinsic parameter matrix mapping, the decoupled image-level coordinates are accurately transformed into a physical-level 3D target localization coordinate matrix. This effectively solves the problem of overall coordinate oscillation and drift caused by accumulated errors in long-term pathological monitoring, ensuring that the final output of TCM acupuncture case analysis results representing acupuncture target coordinates has extremely high security and reliability in physical space, eliminating cross-period semantic drift, and improving the physical reliability of target output. Attached Figure Description

[0014] Figure 1 A schematic diagram illustrating the steps of the machine learning-based TCM acupuncture case analysis output method provided in this application embodiment;

[0015] Figure 2 A schematic diagram of the logical flow of the machine learning-based TCM acupuncture case analysis output method provided in the embodiments of this application;

[0016] Figure 3 A schematic diagram of the structure of the machine learning-based TCM acupuncture case analysis output system provided in the embodiments of this application. Detailed Implementation

[0017] This application's embodiments address the technical problem in existing technologies where, in long-term facial image feature analysis scenarios, the algorithm's inability to decouple local deformation from rigid regions leads to computational jitter and reference coordinate system drift in global deformation calculations on stable regions. This is achieved through a machine learning-based method and system for analyzing and outputting TCM acupuncture cases.

[0018] In machine learning-based case analysis of TCM acupuncture across pathological cycles, the core challenge lies in the strong local heterogeneity of facial topological deformation. Facial paralysis causes severe tissue remodeling in certain affected areas, but maintains absolute physical rigidity in the bony reference region. Traditional deformation analysis methods treat the entire face as a homogeneous elastic rubber sheet, blindly applying global feature difference compensation calculations to all pixels. This forces the rigid reference region, which should be absolutely static, to participate in complex deformation calculations, introducing strong computational jitter and high-frequency mapping noise, ultimately leading to global oscillations and drifts in the target point spatial coordinate system.

[0019] To completely resolve this systemic technical deficiency, this solution abandons the traditional assumption of global consistency and instead seeks the underlying basis for spatial decoupling from information theory. This solution first acquires the original facial image sequences of the target object in the first and second pathological cycles. Then, it extracts the spatiotemporal feature maps of the first and second pathological cycles using a three-dimensional convolutional neural network. Finally, it performs sliding window cropping according to a preset step size, decomposing the macroscopic face into feature matrix pairs of finer-grained local receptive fields.

[0020] At this local microscale, this scheme introduces probability density statistics to calculate the joint entropy gradient and joint entropy gradient matrix across cycles. The joint entropy gradient essentially acts as a detector of local tissue topological activity: lower values ​​indicate a more stable local pixel distribution during the pathological cycle evolution, belonging to an absolutely rigid region; higher values ​​signify a drastic disorder migration of internal features, belonging to an active reshaping region. Simultaneously, an optical flow algorithm is used to calculate the pixel value displacement field of the feature matrix pairs, extracting a two-dimensional deformation vector containing the original displacement direction and magnitude.

[0021] Subsequently, this scheme constructs a spatial hard decoupling and soft dynamic suppression mechanism based on entropy gradient. The pixel coordinates of the spatiotemporal feature map of the first pathological cycle are extracted as a static coordinate reference. For each local receptive field, when the joint entropy gradient is not greater than the rigidity threshold, all two-dimensional deformation vectors contained within that local receptive field are completely reset to zero vectors. This action directly cuts off the feature contamination of stable regions by the deformation algorithm at the source, achieving absolute protection of the rigid reference anchor point. For active regions where the joint entropy gradient is greater than the threshold, this scheme uses a sigmoid activation function to transform the joint entropy gradient matrix into a smoothly transitioning weight matrix, and performs a Hadamard product operation with the internal two-dimensional deformation vectors. This nonlinear weighted multiplication based on tensor in-situ mapping achieves edge-preserving denoising of drastic local deformations, obtaining high-fidelity denoised deformation vectors.

[0022] Finally, by traversing all local receptive fields, the cleaned zero vector and the denoised deformation vector are precisely superimposed onto the static coordinate reference through pure vector addition, and a corrected pixel coordinate system that perfectly matches the real physiological state is assembled. This system is then converted into a physical space target positioning coordinate matrix through camera intrinsic parameter matrix mapping, outputting highly accurate TCM acupuncture case analysis results.

[0023] This scheme uses the joint entropy gradient of information as the evaluation anchor point for local topological activity, and creatively constructs a differential discrete activation and nonlinear suppression architecture in the feature space. It completely breaks the dilemma of excessive contamination of rigid regions caused by global operations in traditional image alignment, and solves the common problem of oscillation and drift of the positioning reference frame under long-period high distortion conditions from the data bottom layer.

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] like Figure 1 The diagram illustrates the steps of the machine learning-based TCM acupuncture case analysis output method provided in this embodiment. This machine learning-based TCM acupuncture case analysis output method is mainly used in machine vision data processing scenarios involving long-term facial image feature analysis and non-rigid deformation tracking and localization. The specific data stream processing steps are as follows:

[0026] The original facial image sequences of the target object in the first and second pathological cycles are obtained and respectively input into a preset three-dimensional convolutional neural network for feature extraction. The spatiotemporal feature maps of the first and second pathological cycles are obtained respectively, and sliding window truncation is performed according to a preset step size to obtain the feature matrix pairs of the local receptive field.

[0027] In the above technical steps, the preset 3D convolutional neural network adopts the standard C3D architecture, namely the Convolutional3D model. The training dataset of this model comes from a dynamic 3D structured light video library containing multiple stages of facial nerve damage evolution, annotated by professional physicians. The network loss function is a weighted combination of the cross-entropy loss function and the smoothing L1 loss function, and the convergence termination condition is set to the validation set loss decreasing by less than 0.0001 for 10 consecutive iterations. The input tensor of the preset 3D convolutional neural network is the original facial image sequence uniformly normalized to a size of 256×256×16, and the output tensor is the spatiotemporal feature map of the first pathological period and the spatiotemporal feature map of the second pathological period, containing deep spatiotemporal semantic information. The preset stride is a fixed spatial step parameter, with a typical engineering value set to 8 pixels. The feature matrix pairs of local receptive fields obtained by traversing the feature maps constitute the basic data units for all subsequent microscopic topological analyses.

[0028] Probability density statistics are performed on the pixel values ​​of the feature matrix pairs, and the joint entropy gradient and joint entropy gradient matrix across periods are calculated accordingly.

[0029] The probability density statistics employ kernel density estimation algorithms, such as Kernel Density Estimation (KDE), which uses a Gaussian kernel function to continuously and smoothly map the discrete pixel value distribution of the feature matrix pair, outputting a continuous probability distribution function curve. Based on the subsequent information entropy formula, the joint entropy gradient and joint entropy gradient matrix across periods are derived.

[0030] The pixel displacement field of the feature matrix pairs is calculated using an optical flow algorithm, and the displacement direction and magnitude are extracted to generate a two-dimensional deformation vector. The optical flow algorithm employs a dense optical flow tracing model, such as the Lucas-Kanade model. The model input consists of local receptive field image patches from the first and second pathological cycles. Based on the assumption of local spatial consistency, the difference equation of the Taylor expansion is solved using the least squares method, outputting the horizontal and vertical motion components of each pixel, and then synthesizing the corresponding two-dimensional deformation vector.

[0031] The pixel coordinates of the spatiotemporal feature map of the first pathological cycle in the local receptive field are extracted as the static coordinate reference. If the joint entropy gradient of the local receptive field is not greater than the rigidity judgment threshold, then all two-dimensional deformation vectors contained in the local receptive field are reset to zero vectors.

[0032] The rigidity threshold is defined as β. β is obtained as follows: during calibration initialization, for the rigid skeletal regions of a healthy subject's face, such as the root of the nose and the glabella, corresponding sample data is extracted. The joint entropy gradient background noise distribution under static conditions is calculated, and the upper limit of its confidence interval is taken as the β value. A typical engineering value range is [0.05, 0.08]. This step constitutes a decoupled judgment branch in the data flow, completely intercepting invalid computational jitter in rigid regions.

[0033] If the joint entropy gradient of the local receptive field is greater than the rigidity threshold, the joint entropy gradient matrix of the local receptive field is constructed into a weight matrix by using the sigmoid activation function, and the weight matrix is ​​then subjected to the Hadamard product operation with the two-dimensional deformation vector in the local receptive field to obtain the noise-reduced deformation vector.

[0034] Traversing all local receptive fields, the corresponding zero vector or noise reduction deformation vector is superimposed onto the corresponding static coordinate reference through vector addition to obtain the corrected pixel coordinate system. The corrected pixel coordinate system is then mapped through a preset camera intrinsic parameter matrix to obtain the target positioning coordinate matrix, and the TCM acupuncture case analysis results representing the acupuncture target coordinates are output accordingly.

[0035] The preset camera intrinsic parameter matrix is ​​a 3×3 matrix containing the camera focal length, principal point coordinates, and radial distortion coefficients, obtained during the device manufacturing process using the Zhang Zhengyou checkerboard calibration method. The mapping process converts the two-dimensional corrected pixel coordinate system into a three-dimensional target positioning coordinate matrix in the physical world.

[0036] Figure 2 This is a schematic diagram of the logical flow of the machine learning-based TCM acupuncture case analysis output method provided in this application embodiment. Through the above scheme, this embodiment introduces the joint entropy gradient as an evaluation anchor point for local topological activity. Combined with hard decoupling reset and soft nonlinear Hadamard product operation, it solves the core technical problem in long-period feature analysis where the algorithm's inability to decouple local deformation from rigid regions causes computational jitter and reference coordinate system drift in stable regions due to global deformation calculation. This achieves the unique technical effect of absolute stability of the underlying spatial coordinate system benchmark.

[0037] Furthermore, in cases of asymmetrical paralysis of facial tissues or severe unevenness of tissue tension, an extreme physical sub-scenario is proposed, where the initial fixed sliding window sampling results in severe spatial structural misalignment before and after pathological progression.

[0038] In this sub-scenario, relying solely on the fixed preset step size in the core basic solution for sliding window truncation will result in the extracted feature matrix actually containing two completely different anatomical tissues. In this case, subsequent optical flow algorithms will forcibly calculate the displacement between these two unrelated tissues, producing physically meaningless topological tearing artifacts, causing the entire system output to completely fail.

[0039] To further address the technical issues of semantic misalignment and matching failure in fixed-step sliding window feature extraction caused by drastic deformation of non-standard faces, this step, as a conditionally triggered branch, operates after obtaining the feature matrix pairs of the local receptive field and completes the loop closure before performing probability density statistics on the pixel values ​​of the feature matrix pairs, ensuring the homogeneity and legality of downstream statistical data.

[0040] Specifically, it includes the following technical steps: obtaining a preset step size scaling factor, a preset neighbor pixel length, and a preset neighbor pixel width; defining the preset step size scaling factor as a global variable. Its range is Typical engineering value Define the preset neighborhood pixel length as The preset neighborhood pixel width is Both are set based on the maximum possible muscle stretch limit, with typical values ​​being [missing value]. 1 pixel.

[0041] The feature matrix is ​​a first feature matrix containing the spatiotemporal feature map of the first pathological cycle in the local receptive field, and a second feature matrix containing the spatiotemporal feature map of the second pathological cycle in the local receptive field; the mutual information of the first and second feature matrices is calculated; the mutual information is defined as a global variable. The calculation model is the sum of edge information entropy minus joint information entropy.

[0042] If the mutual information is not lower than the preset association matching threshold, no processing is performed; the preset association matching threshold is defined as a global variable. Based on the normalized mutual information criterion, the typical value is obtained as follows: .

[0043] If the mutual information is lower than the preset association matching threshold, the preset step size of the current sliding window is reduced by the preset step size scaling factor, and a candidate offset coordinate set containing multiple random discrete coordinate points is established within the local neighborhood range determined by the preset neighborhood pixel length and the preset neighborhood pixel width, with the geometric center of the current sliding window's intercepted area as the center point.

[0044] Traverse the candidate offset coordinate set, take each random discrete coordinate point as the cut center, cut the feature matrix to be matched from the spatiotemporal feature map of the second pathological cycle, and calculate the mutual information between each feature matrix to be matched and the feature matrix in the spatiotemporal feature map of the first pathological cycle.

[0045] The random discrete coordinate point corresponding to the maximum value of mutual information is selected as the target offset point, and the feature matrix is ​​re-trunculated based on the target offset point to update the feature matrix pair, thus obtaining a new feature matrix pair.

[0046] Through the aforementioned dynamic adaptive neighborhood search resampling mechanism driven by mutual information, this scheme successfully retrieved the feature association points with the highest similarity under spatial misalignment conditions, outputting new feature matrix pairs with high confidence, completely overcoming the alignment deviation caused by local deformation, and providing a reliable homogeneous benchmark for subsequent difference operations.

[0047] Furthermore, in the process of acquiring long-term medical data, there are adverse data sub-scenarios where isolated high-frequency noise points appear within the extracted feature matrix due to changes in ambient lighting, sensor thermal noise, and facial oil reflection. Although the spatial alignment problem has been solved, if high-frequency isolated noise exists within the feature matrix, this noise will be severely amplified during the kernel density estimation stage, leading to an artificially high calculated joint entropy gradient. The system may mistakenly identify pure optical noise as severe deformation, causing a misjudgment and collapse of the rigid reference frame.

[0048] To address the technical problem of low signal-to-noise ratio in local feature activation values, leading to failure in joint entropy gradient calculation and misjudgment of states, this step, as a serial interception and filtering node, operates strictly between the acquisition of the new feature matrix pair and the performance of probability density statistics.

[0049] Specifically, it includes the following technical steps:

[0050] Calculate the statistical variance of the pixel values ​​contained in each feature matrix of the new feature matrix pair; define the statistical variance as a global variable. It quantifies the degree of dispersion of local eigenvalues.

[0051] If the statistical variance exceeds the preset noise interference threshold, the corresponding feature matrix is ​​divided into multiple overlapping local sub-blocks according to the preset two-dimensional grid size parameters, the local histogram of each sub-block is calculated, and the number of pixels in the histogram is redistributed according to the preset shearing threshold. The equalized pixel values ​​are obtained using the bilinear interpolation algorithm.

[0052] Define the preset noise interference threshold as Based on multiple periods of no-load noise floor testing, the typical value is the mean. Times; Define the preset 2D grid size parameters as Typical value Define the preset shear threshold as: Typical value This processing method follows the standard procedure of the contrast-limited adaptive histogram equalization algorithm, using bilinear interpolation to eliminate the block stitching effect between sub-blocks. Based on the reduced preset step size, the probability density of the equalized pixel values ​​is recalculated to obtain a new joint entropy gradient and a new joint entropy gradient matrix across periods.

[0053] By using variance monitoring and local histogram redistribution operations, this embodiment successfully suppressed the peaks of isolated noise points, improved the effective feature contrast, ensured that the data source input to the entropy calculation had high purity, and avoided a systemic avalanche in the state quantification evaluation stage.

[0054] Furthermore, for homogeneous texture areas on the face, the aperture effect of the optical flow algorithm causes chaotic local jumps and oscillations in the calculated two-dimensional deformation vector directions between adjacent local receptive fields, resulting in a refined sub-scene.

[0055] While the preceding methods ensure alignment and noise-free operation, optical flow naturally lacks edge constraints in homogeneous regions. Directly using these transition vectors for target point coordinate mapping can lead to jagged, high-frequency jitter in the acupuncture planning path, posing a serious medical safety hazard.

[0056] To further address the technical issues of deformation direction oscillations and spatial consistency loss caused by ambiguities in local optical flow calculations on manifold surfaces, this step is performed after obtaining the new joint entropy gradient and the new joint entropy gradient matrix across periods. It must be executed after obtaining the two-dimensional deformation vector through the optical flow algorithm and generating the denoised deformation vector using the Hadamard product, serving as the final smoothing layer before final coordinate synthesis.

[0057] The specific technical steps and formulas are as follows: Extract the vector angle between the current local receptive field and the denoised deformation vectors between its adjacent local receptive fields, and calculate the cosine similarity of the vector angle; define the vector angle between the denoised deformation vectors as a global variable. Cosine similarity is defined as The formula is the inner product of two vectors divided by the product of their magnitudes.

[0058] If the cosine similarity of the included vector angle is lower than the preset consistency threshold, and the mutual information of the corresponding feature matrix pair is not lower than the association matching threshold, then the joint entropy gradient of the current local receptive field is multiplied by the preset variance adjustment coefficient, and the product result is added to the preset basic variance to calculate the variance parameter of the Gaussian smoothing kernel.

[0059] Define the preset consistency threshold as Typical value The preset variance adjustment coefficient is defined as follows: Define the preset basic variance as The variance parameter of the Gaussian smoothing kernel is The specific linear mapping equation established in this step is as follows: ,in, This is the joint entropy gradient. This formula ensures that the more severe the deformation, the more... The larger the region, the greater the generated smoothing variance.

[0060] Substituting the variance parameter into a preset two-dimensional Gaussian distribution function, the corresponding weight matrix is ​​calculated and used as the Gaussian smoothing kernel; the preset two-dimensional Gaussian distribution function is defined as follows: Its standard formula is:

[0061] ,in, The coordinates are two-dimensional position indexes within the Gaussian kernel space.

[0062] A new denoised deformation vector is obtained by performing spatial weighted convolution filtering on the denoised deformation vector within the current local receptive field and its preset neighborhood using a Gaussian smoothing kernel.

[0063] This scheme utilizes mutual information and cosine similarity to implement double-reset reliability verification, and creatively reuses joint entropy gradient to adaptively generate Gaussian variance parameters. It performs strict manifold constraint filtering on the deformation direction, thoroughly corrects local jumps, and ensures the anatomical rationality of the endpoint positioning coordinate matrix.

[0064] Furthermore, the key information theory mathematical model in the core basic scheme is white-boxed and decoupled from its parameters: the probability distribution values ​​of the pixel values ​​of the feature matrix are obtained;

[0065] Calculate the information entropy of the first and second feature matrices of the feature matrix pair in the corresponding period according to the following information entropy formula;

[0066] ,in, Represents information entropy. Represents the pixel values ​​of the feature matrix. This represents the probability distribution of pixel values; the formula requires summing all unique discrete pixel values ​​in the feature matrix. The absolute deviation of the information entropy of the first and second feature matrices in the corresponding period is calculated, denoted as the cross-period joint entropy gradient; the joint entropy gradient is a single scalar value, and the global variable is denoted as... .

[0067] Extract the spatial row and column structure of the feature matrix pairs of the same local receptive field, construct an empty matrix consistent with the spatial row and column structure, and uniformly replace all elements in the matrix with the joint entropy gradient to generate a joint entropy gradient matrix that retains the spatial topology of the local receptive field.

[0068] This embodiment transforms macroscopic statistical scalars into microscopic spatial topological tensors without loss, completely eliminating the black box of dimensional breakage when multiplying data of different dimensions.

[0069] Furthermore, the operational logic of the sigmoid activation function and the tensor in-situ mapping process are clarified:

[0070] The joint entropy gradient is calculated into a weight matrix using the sigmoid activation function, specifically including:

[0071] Using the joint entropy gradient matrix as the input tensor, we iterate through each element of the joint entropy gradient matrix; then, we input the joint entropy gradient within each element into a sigmoid activation function for nonlinear mapping, obtaining the weight values. The sigmoid activation function is as follows:

[0072] ,in, These are the weight values ​​in the weight matrix. Represents the joint entropy gradient. Represents the natural constant. The preset slope adjustment coefficient is used. A rigid threshold for judgment;

[0073] global variables The degree of nonlinear steepness of the controlled mapping response is determined by the optimal smoothing performance on the validation set, with a typical value of [value missing]. .

[0074] The weight values ​​generated by the mapping are overwritten and saved in their original positions to the matrix element positions of the joint entropy gradient matrix, thereby directly converting the entire joint entropy gradient matrix into a weight matrix, ensuring that the generated weight matrix corresponds one-to-one with the spatial position nodes of the corresponding two-dimensional deformation vector.

[0075] This embodiment ensures that the generated weight matrix strictly inherits the physical coordinate attributes of the feature map, providing a mathematical basis for point-to-point noise reduction and suppression in space.

[0076] Furthermore, this embodiment refines the underlying computation execution level: the specific process of obtaining the noise reduction deformation vector includes: extracting the weight values ​​of each spatial location node in the weight matrix, and performing Hadamard product operation on each weight value with the horizontal axis displacement component and the vertical axis displacement component of the same spatial location index in the two-dimensional deformation vector to obtain the noise reduction deformation vector.

[0077] The Hadamard product requires that the matrix dimensions be completely consistent.

[0078] Since the preceding scheme has ensured that the weight matrix and the displacement vector have the same dimension, this element-wise multiplication operation can adjust the magnitude intensity by weights without destroying the original two-dimensional deformation vector topological orientation. Adaptive compression is used to achieve precise noise reduction.

[0079] Furthermore, this implementation method addresses the mapping and fusion process of the final macroscopic coordinate space:

[0080] The specific process of obtaining the corrected pixel coordinate system includes:

[0081] Obtain the horizontal and vertical displacements of the zero vector or the noise reduction deformation vector in the pixel coordinate system, respectively.

[0082] The horizontal axis displacement is arithmetically added to the x-coordinate of the static coordinate reference, and the vertical axis displacement is arithmetically added to the y-coordinate of the static coordinate reference to obtain the corrected pixel value coordinate points. The corrected pixel value coordinate system is formed by traversing all the corrected pixel value coordinate points obtained from the local receptive field.

[0083] By extracting independent orthogonal components and performing bottom-level rigid arithmetic translation accumulation, the cleaned discrete micro-deformation features are perfectly connected to the bottom-level reference coordinate map, eliminating the system's cumulative splicing error.

[0084] Further considering complex scenarios where overlapping pixel regions exist between different local receptive fields due to the preset step size being smaller than the spatial size of the local receptive field, the following approach is taken: The horizontal and vertical displacements of the zero vector or the denoised deformation vector in the pixel value coordinate system are obtained. The horizontal displacement is arithmetically added to the x-coordinate value of the static coordinate reference, and the vertical displacement is arithmetically added to the y-coordinate value of the static coordinate reference to obtain the corrected pixel value coordinate points. The process of constructing the corrected pixel value coordinate system from all corrected pixel value coordinate points obtained by traversing the local receptive fields is executed through the following spatial weight fusion logic to eliminate coordinate overlap conflicts:

[0085] Determine the overlap coverage depth: For any pixel to be corrected in the corrected pixel value coordinate system, retrieve the total number of times it is covered by the local receptive field, and denote it as the coverage depth parameter. .like If so, then perform a single vector addition directly; if Then it enters the multi-dimensional fusion process.

[0086] Calculate the spatial contribution weights: for the first A local receptive field covering the pixel to be corrected is defined, and the corresponding spatial contribution weight is calculated based on the Euclidean distance of the pixel from the geometric center of the local receptive field using a preset Gaussian decay function. The calculation formula is as follows:

[0087] ,in, This represents the original position of the pixel to be corrected on the static coordinate reference. For the first The geometric center coordinates of a local receptive field This is a preset spatial smoothing factor used to control the rate at which the weights decay with distance.

[0088] Perform weighted vector fusion: extract the corresponding vectors of the pixel to be corrected. A zero vector or denoised deformation vector is used, and a normalized fusion operation is performed in combination with the corresponding spatial contribution weights to obtain a unique fused displacement vector. The calculation logic is as follows:

[0089] ,in, For the first The zero vector or noise-reduced deformation vector output by a local receptive field.

[0090] Coordinate system reconstruction: The horizontal and vertical components of the calculated fused displacement vector are superimposed on the static coordinate reference of the pixel to be corrected, generating the final unique corrected pixel value coordinate point.

[0091] Through the above technical solution, this embodiment solves the coordinate update conflict and pixel ghosting problem caused by the overlap of local receptive fields during the sliding window cropping process by introducing the spatial contribution weight based on the Gaussian decay function and performing weighted vector fusion. It achieves the unique technical effect of outputting a corrected pixel value coordinate system with global spatial consistency while ensuring feature continuity.

[0092] Furthermore, the corrected pixel coordinate system is mapped through a preset camera intrinsic parameter matrix to obtain the target positioning coordinate matrix, and based on this, the results of acupuncture case analysis representing the coordinates of the acupuncture target point are output, including:

[0093] The camera intrinsic parameter matrix is ​​used to transform the two-dimensional coordinates in the corrected pixel coordinate system to the camera three-dimensional coordinate system, and combined with the preset extrinsic parameter matrix to perform rotation and translation transformations, mapping the coordinates to the world three-dimensional coordinate system to generate the target positioning coordinate matrix;

[0094] The three-dimensional coordinate points in the target positioning coordinate matrix are matched with the preset standard acupoint map of traditional Chinese medicine by spatial nearest neighbor matching to determine the name of the target acupoint and output the analysis results.

[0095] Figure 3This is a schematic diagram of the structure of the machine learning-based TCM acupuncture case analysis output system provided in this application embodiment. The machine learning-based TCM acupuncture case analysis output system includes: a local receptive field extraction module: used to acquire the original facial image sequences of the target object in the first and second pathological cycles, and input them into a preset three-dimensional convolutional neural network for feature extraction, obtaining spatiotemporal feature maps of the first and second pathological cycles respectively, and performing sliding window truncation according to a preset step size to obtain feature matrix pairs of local receptive fields; a joint entropy gradient extraction module: used to perform probability density statistics on the pixel values ​​of the feature matrix pairs, and calculate the joint entropy gradient and joint entropy gradient matrix across cycles accordingly; a two-dimensional deformation vector processing module: used to calculate the pixel value displacement field of the feature matrix pairs using an optical flow algorithm, extracting the displacement direction and displacement amplitude to generate a two-dimensional deformation vector; and a static coordinate processing module: used to extract the pixel coordinates of the spatiotemporal feature map of the first pathological cycle in the local receptive field. The local receptive field is used as a static coordinate reference. If the joint entropy gradient of the local receptive field is not greater than the rigid judgment threshold, all two-dimensional deformation vectors contained in the local receptive field are reset to zero vectors. The noise reduction deformation vector processing module is used to calculate and construct a weight matrix of the joint entropy gradient matrix of the local receptive field through the S-shaped activation function if the joint entropy gradient of the local receptive field is greater than the rigid judgment threshold. The weight matrix is ​​then used to perform a Hadamard product operation with the two-dimensional deformation vectors in the local receptive field to obtain the noise reduction deformation vector. The output module is used to traverse all local receptive fields, and then superimpose the corresponding zero vectors or noise reduction deformation vectors onto the static coordinate reference through vector addition to obtain the corrected pixel coordinate system. The corrected pixel coordinate system is then mapped through the preset camera intrinsic parameter matrix to obtain the target positioning coordinate matrix, and the TCM acupuncture case analysis results representing the acupuncture target coordinates are output accordingly.

[0096] This application also provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements a method for outputting TCM acupuncture case analysis based on machine learning.

[0097] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0098] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0099] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0100] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0101] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0102] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for analyzing and outputting TCM acupuncture cases based on machine learning, characterized in that, Includes the following steps: The original facial image sequences of the target object in the first and second pathological cycles are obtained and respectively input into a preset three-dimensional convolutional neural network for feature extraction. The spatiotemporal feature maps of the first and second pathological cycles are obtained respectively, and sliding window truncation is performed according to a preset step size to obtain the feature matrix pairs of the local receptive field. Probability density statistics are performed on the pixel values ​​of the feature matrix pairs, and the joint entropy gradient and joint entropy gradient matrix across periods are calculated accordingly. The pixel value displacement field of the feature matrix pair is calculated by optical flow algorithm, and the displacement direction and displacement amplitude are extracted to generate a two-dimensional deformation vector. The pixel coordinates of the spatiotemporal feature map of the first pathological cycle in the local receptive field are extracted as the static coordinate reference. If the joint entropy gradient of the local receptive field is not greater than the rigidity judgment threshold, then all two-dimensional deformation vectors contained in the local receptive field are reset to zero vectors. If the joint entropy gradient of the local receptive field is greater than the rigidity threshold, the joint entropy gradient matrix of the local receptive field is constructed into a weight matrix by using the sigmoid activation function, and the weight matrix is ​​then subjected to the Hadamard product operation with the two-dimensional deformation vector in the local receptive field to obtain the noise-reduced deformation vector. Traversing all local receptive fields, the corresponding zero vector or noise reduction deformation vector is superimposed onto the static coordinate reference through vector addition to obtain the corrected pixel coordinate system. The corrected pixel coordinate system is then mapped through a preset camera intrinsic parameter matrix to obtain the target positioning coordinate matrix, and the results of TCM acupuncture case analysis representing the coordinates of the acupuncture target point are output accordingly.

2. The method for outputting TCM acupuncture case analysis based on machine learning according to claim 1, characterized in that, After obtaining the feature matrix pairs of the local receptive fields, the following is also included: Obtain the preset step size scaling factor, the preset neighbor pixel length, and the preset neighbor pixel width; The feature matrix pair contains the first feature matrix of the spatiotemporal feature map of the first pathological cycle in the local receptive field and the second feature matrix of the spatiotemporal feature map of the second pathological cycle in the local receptive field. Calculate the mutual information between the first and second characteristic matrices; If the mutual information is not lower than the preset association matching threshold, no processing will be performed; If the mutual information is lower than the preset association matching threshold, the preset step size of the current sliding window is reduced by the preset step size scaling factor, and a candidate offset coordinate set containing multiple random discrete coordinate points is established within the local neighborhood range determined by the preset neighborhood pixel length and the preset neighborhood pixel width, with the geometric center of the current sliding window's intercepted area as the center point. Traverse the candidate offset coordinate set, take each random discrete coordinate point as the cut center, cut the feature matrix to be matched from the spatiotemporal feature map of the second pathological cycle, and calculate the mutual information between each feature matrix to be matched and the feature matrix in the spatiotemporal feature map of the first pathological cycle. The random discrete coordinate point corresponding to the maximum value of mutual information is selected as the target offset point, and the feature matrix is ​​re-trunculated based on the target offset point to update the feature matrix pair, thus obtaining a new feature matrix pair.

3. The method for outputting TCM acupuncture case analysis based on machine learning according to claim 2, characterized in that, After obtaining the new feature matrix pairs, the following quality enhancement steps are also included: Calculate the statistical variance of the pixel values ​​contained in each feature matrix of the new feature matrix pair; If the statistical variance exceeds the preset noise interference threshold, the corresponding feature matrix is ​​divided into multiple overlapping local sub-blocks according to the preset two-dimensional grid size parameters, the local histogram of each sub-block is calculated, and the number of pixels in the histogram is redistributed according to the preset shearing threshold. The equalized pixel values ​​are obtained using the bilinear interpolation algorithm. Based on the reduced preset step size, the probability density statistics of the equalized pixel values ​​are re-performed to calculate the new joint entropy gradient and the new joint entropy gradient matrix across the period.

4. The method for outputting TCM acupuncture case analysis based on machine learning according to claim 3, characterized in that, After obtaining the new joint entropy gradient and the new joint entropy gradient matrix across the period, the following is also included: Extract the vector angle between the current local receptive field and the denoised deformation vectors between its adjacent local receptive fields, and calculate the cosine similarity of the vector angle. If the cosine similarity of the included vector angle is lower than the preset consistency threshold, and the mutual information of the corresponding feature matrix pair is not lower than the association matching threshold, then the joint entropy gradient of the current local receptive field is multiplied by the preset variance adjustment coefficient, and the product result is added to the preset basic variance to calculate the variance parameter of the Gaussian smoothing kernel. Substitute the variance parameter into the preset two-dimensional Gaussian distribution function to calculate and generate the corresponding weight matrix as the Gaussian smoothing kernel; A new denoised deformation vector is obtained by performing spatial weighted convolution filtering on the denoised deformation vector within the current local receptive field and its preset neighborhood using a Gaussian smoothing kernel.

5. The method for outputting TCM acupuncture case analysis based on machine learning according to claim 1, characterized in that, The specific processing steps for the cross-period joint entropy gradient and joint entropy gradient matrix include: Obtain the probability distribution values ​​of the pixel values ​​in the feature matrix; Calculate the information entropy of the first and second feature matrices of the feature matrix pair in the corresponding period according to the following information entropy formula; ,in, Represents information entropy. Represents the pixel values ​​of the feature matrix. Represents pixel value The probability distribution value; Calculate the absolute deviation of the information entropy of the first feature matrix and the second feature matrix in the corresponding period, and denote it as the joint entropy gradient across the period; Extract the spatial row and column structure of the feature matrix pairs of the same local receptive field, construct an empty matrix consistent with the spatial row and column structure, and uniformly replace all elements in the empty matrix with the joint entropy gradient to generate a joint entropy gradient matrix that retains the spatial topology of the local receptive field.

6. The method for outputting TCM acupuncture case analysis based on machine learning according to claim 1, characterized in that, The joint entropy gradient is calculated into a weight matrix using the sigmoid activation function, specifically including: Take the joint entropy gradient matrix as the input tensor and iterate through each element of the joint entropy gradient matrix. The joint entropy gradient within each matrix element is input into a sigmoid activation function for nonlinear mapping, yielding the weight values. The sigmoid activation function is as follows: ,in, These are the weight values ​​in the weight matrix. Represents the joint entropy gradient. Represents the natural constant. The preset slope adjustment coefficient is used. A rigid threshold for judgment; The weight values ​​generated by the mapping are overwritten and saved in their original positions to the matrix element positions of the joint entropy gradient matrix, thereby directly converting the entire joint entropy gradient matrix into a weight matrix, ensuring that the generated weight matrix corresponds one-to-one with the spatial position nodes of the two-dimensional deformation vector.

7. The method for outputting TCM acupuncture case analysis based on machine learning according to claim 1, characterized in that, The specific process for obtaining the noise reduction deformation vector includes: Extract the weight values ​​of each spatial location node in the weight matrix, and perform Hadamard product operation on each weight value with the horizontal axis displacement component and the vertical axis displacement component with the same spatial location index in the two-dimensional deformation vector to obtain the noise-reduced deformation vector.

8. The method for outputting TCM acupuncture case analysis based on machine learning according to claim 1, characterized in that, The specific process for obtaining the corrected pixel coordinate system includes: Obtain the horizontal and vertical displacements of the zero vector or the noise reduction deformation vector in the pixel coordinate system, respectively. The horizontal axis displacement is arithmetically added to the x-coordinate of the static coordinate reference, and the vertical axis displacement is arithmetically added to the y-coordinate of the static coordinate reference to obtain the corrected pixel value coordinate points. The corrected pixel value coordinate system is formed by traversing all the corrected pixel value coordinate points obtained from the local receptive field.

9. A machine learning-based TCM acupuncture case analysis output system, characterized in that, include: Local receptive field extraction module: used to acquire the original facial image sequences of the target object in the first and second pathological cycles, and input them into a preset three-dimensional convolutional neural network for feature extraction, to obtain the spatiotemporal feature maps of the first and second pathological cycles respectively, and to perform sliding window cropping according to a preset step size to obtain the feature matrix pairs of the local receptive field; Joint entropy gradient extraction module: used to perform probability density statistics on the pixel values ​​of feature matrix pairs, and calculate the joint entropy gradient and joint entropy gradient matrix across periods accordingly; Two-dimensional deformation vector processing module: used to calculate the pixel value displacement field of feature matrix pairs through optical flow algorithm, extract displacement direction and displacement amplitude to generate two-dimensional deformation vector; Static coordinate processing module: used to extract the pixel coordinates of the spatiotemporal feature map of the first pathological cycle in the local receptive field as the static coordinate reference. If the joint entropy gradient of the local receptive field is not greater than the rigidity judgment threshold, then all two-dimensional deformation vectors contained in the local receptive field are reset to zero vectors. The noise reduction deformation vector processing module is used to calculate and construct a weight matrix from the joint entropy gradient matrix of the local receptive field if the joint entropy gradient of the local receptive field is greater than the rigidity threshold. Then, the weight matrix is ​​used to perform a Hadamard product operation with the two-dimensional deformation vector in the local receptive field to obtain the noise reduction deformation vector. Output module: Used to traverse all local receptive fields, superimpose the corresponding zero vector or noise reduction deformation vector onto the static coordinate reference through vector addition to obtain the corrected pixel coordinate system, map the corrected pixel coordinate system through the preset camera intrinsic parameter matrix to obtain the target positioning coordinate matrix, and output the TCM acupuncture case analysis results representing the acupuncture target coordinates.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-8.

Images