Muscle tension detection method, system and device

Abstract

The present invention relates to the technical field of non-invasive muscle tension detection. Specifically, the present invention relates to a muscle tension detection method, system and device. The method acquires muscle surface micro-vibration signals through a millimeter-wave radar array and generates a time-frequency diagram. Feature parameters such as the main vibration frequency amplitude are extracted. The feature values ​​are fused and output through a pre-trained model to obtain a preliminary grade. The deviation is verified by combining with a Hill muscle model. If the deviation exceeds the limit, a three-dimensional muscle model is reconstructed based on CT images, and the radar wave attenuation coefficient is calculated for compensation. The final grade is output. The method solves the problems of insufficient accuracy and poor adaptability of traditional detection to individual differences, improves the accuracy and personalization of muscle tension detection, and provides a reliable basis for clinical diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of non-invasive muscle tension detection, and in particular to a muscle tension detection method, system and device. Background Art

[0002] Non-invasive muscle tension detection technology is an important technology. In rehabilitation medicine and neurological clinical diagnosis, muscle tension detection is a core means to evaluate muscle function and diagnose conditions such as spasticity or muscle weakness. Accurate detection methods quantify muscle tension levels, providing a scientific basis for treatment plan formulation and rehabilitation effect evaluation.

[0003] However, traditional muscle tension detection technology has the core problems of insufficient detection accuracy and poor adaptability to individual differences. The existing scheme only judges the tension level by a single vibration signal feature, and does not correct for differences in muscle anatomical structure. When the thickness of the fat layer at the detection site is different, the radar wave attenuation is not effectively compensated, resulting in large deviations in the detection results of the same muscle state in different individuals. Only a fixed threshold is used to divide the tension level, and the physiological model of muscle contraction is not associated. When the muscle micro-vibration signal is disturbed by involuntary tremors, it is impossible to distinguish between pathological tension changes and motion artifacts, resulting in misjudgment of grading. In addition, the traditional system lacks closed-loop verification of the detection results and the theoretical model. When the measured signal deviates significantly from the theoretical curve, the secondary calibration is not initiated, resulting in reduced reliability of the clinical diagnosis basis and increased misdiagnosis rate, which not only affects the effectiveness of the treatment plan, but may also delay the patient's recovery process, making it difficult to achieve accurate detection and grading of muscle tension. In order to solve this technical problem, we provide a muscle tension detection method, system and device. Summary of the Invention

[0004] The object of the present invention is to provide a muscle tension detection method, system and device to solve the problems raised in the above background technology.

[0005] 1. Because traditional testing does not compensate for individual anatomical differences, fat layers and other factors can cause signal attenuation, resulting in significant deviations in test results. Therefore, this case used CT images to reconstruct a three-dimensional model, calculate and compensate for the radar wave attenuation coefficient, and standardize the characteristic values ​​to adapt to individual differences and improve test accuracy.

[0006] 2. Because traditional testing lacks theoretical verification and is easily misjudged due to artifact interference, this case uses the Hill model to generate a theoretical correlation curve, verify the measured deviation, and activate secondary calibration, which can reduce misjudgments and improve classification reliability.

[0007] To achieve the above object, one of the objects of the present invention is to provide a muscle tension detection method, comprising the following steps:

[0008] S1. Scan the target individual muscle group area through the millimeter wave radar array to obtain the muscle surface micro-vibration signal, and generate a time-frequency graph based on the muscle surface micro-vibration signal;

[0009] S2. Extracting features from the time-frequency graph to obtain characteristic parameters, the characteristic parameters including the amplitude of the main vibration frequency, the proportion of low-frequency band energy, and the energy entropy of the mid-frequency band;

[0010] S3. Generate low-frequency band weights and mid-frequency band weights based on the ratio of the main vibration frequency amplitude to the preset spasm threshold, and calculate a fusion eigenvalue based on the low-frequency band weight, the mid-frequency band weight, the low-frequency band energy proportion, and the mid-frequency band energy entropy. Input the fusion eigenvalue into the pre-trained muscle tension grading model to output a preliminary grade. Simultaneously, simulate the theoretical correlation curve between the muscle fiber contraction force and the vibration spectrum of the time-frequency graph at the current preliminary grade based on the Hill muscle model. If the deviation between the measured main vibration frequency amplitude and the theoretical correlation curve exceeds the preset threshold, activate S4.

[0011] S4. Reconstruct the three-dimensional muscle model based on the CT image of the target individual, calculate the attenuation coefficient of the radar wave in the fat / muscle fiber layer, use the attenuation coefficient to attenuate the main vibration frequency amplitude and the fusion eigenvalue, input the attenuation-compensated fusion eigenvalue into the muscle tension grading model again, and output the final tension level.

[0012] As a further improvement of the present technical solution, the low-frequency band weight and the mid-frequency band weight are generated according to the ratio of the main vibration frequency amplitude to the preset spasm threshold in S3, specifically including:

[0013] Establish a library of spasm thresholds associated with muscle anatomical location and body shape, and match the corresponding preset spasm thresholds according to the target individual's body mass index and muscle depth;

[0014] Calculate the dynamic deviation coefficient of the main vibration frequency amplitude relative to the preset spasm threshold, which reflects the degree of proximity between the current muscle state and the spasm critical point;

[0015] Comparing the dynamic deviation coefficient with the preset spasm risk interval, determining the spasm risk interval in which the dynamic deviation coefficient is located based on the comparison result, and performing weight distribution based on the spasm risk interval to generate low-frequency band weights and mid-frequency band weights;

[0016] The spasm risk index is constructed based on the product of the dynamic deviation coefficient and the low-frequency band weight, and is used for priority control of the fusion eigenvalue.

[0017] As a further improvement of the present technical solution, the fusion feature value is calculated in S3 according to the low-frequency band weight, the mid-frequency band weight, the low-frequency band energy proportion and the mid-frequency band energy entropy, as follows:

[0018] A motion artifact correction module is introduced. When involuntary muscle tremor is detected, the contribution of low-frequency band energy is adjusted using the spasm risk index.

[0019] Design a physiological noise suppression strategy based on mid-band energy entropy. If the current mid-band energy entropy value deviates from the historical mean by more than a preset tolerance, the sliding window mean is used to replace the outlier.

[0020] The fusion eigenvalue consists of four parts, namely, the product of the calibrated low-frequency band energy proportion and the low-frequency band weight, the product of the mid-frequency band energy entropy and the mid-frequency band weight, the entropy value credibility weighting coefficient, and the compensation term of the spasm risk index.

[0021] As a further improvement of the present technical solution, the step S3 of inputting the fused feature value into the pre-trained muscle tension grading model to output a preliminary grade further includes:

[0022] A two-tiered decision-making model was constructed. The first tier implemented spasticity risk assessment. When the spasticity risk index exceeded the warning value, a high-risk level was directly output. The second tier initiated multi-feature collaborative classification, using a random forest algorithm to analyze the fused feature values ​​and additional physiological parameters.

[0023] A time domain stability check is introduced, which modifies the preliminary grade based on the mode distribution of the last three test results.

[0024] As a further improvement of this technical solution, the S3 synchronously simulates the theoretical correlation curve based on the Hill muscle model, which is specifically implemented as follows:

[0025] The preliminary level was mapped to the muscle fiber activation rate, and a muscle contraction force calculation model was constructed. The model includes a linear conversion relationship between contraction speed and vibration frequency and a force value growth function related to muscle fiber activation rate.

[0026] A parallel elastic element force calculation model is constructed. The parallel elastic element force calculation model introduces muscle elongation. The muscle elongation is derived from the muscle bundle geometry analysis of CT images. The theoretical vibration spectrum amplitude is calculated by the vector synthesis results of the contraction element force and the elastic element force. Finally, a frequency-amplitude theoretical correlation curve is established, with the horizontal axis representing the vibration frequency and the vertical axis representing the theoretical amplitude.

[0027] As a further improvement of the present technical solution, the step of reconstructing the muscle three-dimensional model based on the CT image of the target individual in S4 specifically includes:

[0028] A deep learning segmentation network was used to identify the boundaries of the fat layer, the direction of the myofascia, and the spatial topology of the muscle fibers, and a layered dielectric property model was constructed. The fat layer was defined as a uniform dielectric layer with a fixed low dielectric constant, and the muscle fiber layer was defined as an anisotropic structure with a significantly higher dielectric constant along the fiber direction than in the perpendicular direction.

[0029] As a further improvement of the present technical solution, the radar wave attenuation coefficient is calculated in S4, and the specific steps are as follows:

[0030] Calculate the effective propagation path. The effective propagation path length is determined by the radar incident angle, the muscle fiber orientation angle, and the thickness of the fat layer. The propagation path is inversely proportional to the cosine value of the angle.

[0031] The radar wave attenuation coefficient is composed of three parts:

[0032] Fat layer attenuation term: increases exponentially with thickness;

[0033] Attenuation term parallel to muscle fibers: positively correlated with effective propagation path length and high dielectric constant;

[0034] Attenuation term in the vertical direction of muscle fibers: weakly correlated with the effective propagation path length and low dielectric constant.

[0035] As a further improvement of the present technical solution, the attenuation compensation is performed using the attenuation coefficient in S4, specifically including:

[0036] ①. Main vibration frequency amplitude compensation: Use the inverse square root of the attenuation coefficient to restore the amplitude;

[0037] ② Fusion eigenvalue compensation: Use the logarithmic function of the attenuation coefficient for nonlinear correction;

[0038] ③ Muscle cross-sectional area normalization: Normalize the compensated eigenvalues ​​based on the physiological parameters extracted from the CT model;

[0039] ④. A confidence label is generated synchronously when the final tension level is output. When the level jump before and after calibration exceeds the clinical tolerance, a manual review alert is triggered.

[0040] A second object of the present invention is to provide a system for implementing any one of the above-mentioned muscle tension detection methods, comprising:

[0041] The signal acquisition feature generation unit is equipped with a phased array radar probe to scan the target muscle group area, capture the original micro-vibration signal, input the original signal into the short-time Fourier transform and convert it into a time-frequency graph. The main vibration frequency amplitude, low-frequency band energy ratio and mid-frequency band energy entropy are extracted and transmitted to the intelligent classification and verification unit, where:

[0042] The intelligent grading and verification unit calls the spasm threshold library, inputs the main vibration frequency amplitude, low-frequency band energy ratio and mid-frequency band energy entropy into the spasm threshold library to calculate the dynamic deviation coefficient, generates the low-frequency band weight, mid-frequency band weight and spasm risk index, calibrates the low-frequency band energy ratio based on the spasm risk index, suppresses the mid-frequency band energy entropy noise, outputs the fused eigenvalue, and screens high-risk cases according to the spasm risk index. The fused eigenvalue is input into the random forest model to output a preliminary grade, generates a theoretical correlation curve based on the preliminary grade, verifies the measured main vibration frequency amplitude deviation and generates a calibration instruction, and transmits the preliminary grade, fused eigenvalue and calibration instruction to the anatomical calibration unit;

[0043] The anatomical calibration unit uses a convolutional neural network to segment the fat layer and the muscle fiber layer, reconstruct a three-dimensional muscle model including dielectric properties, calculates the radar wave attenuation coefficient based on the muscle model, performs inverse square root compensation on the main vibration frequency amplitude, implements logarithmic correction on the fused eigenvalues, and then uses the muscle fiber cross-sectional area to standardize the compensated eigenvalues. The standardized eigenvalues ​​are input into the grading model to generate the final tension level. The level differences before and after calibration are compared to output confidence labels and clinical warnings.

[0044] A third object of the present invention is to provide a device for implementing a muscle tension detection method including any one of the above items, characterized in that it includes a phased array probe array, a signal preprocessing circuit, a data acquisition interface, a hierarchical processing module, a feature fusion processor, a two-level decision engine, a Hill model simulator, a CT image processor, an attenuation compensation calculation module and a normalized output terminal.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention uses a millimeter-wave radar array to collect micro-vibration signals on the muscle surface and generate time-frequency diagrams, extracts multi-dimensional feature parameters, combines the spasm threshold library with dynamic weight distribution to generate fused feature values, and implements preliminary classification through a pre-trained model, thus solving the problem of insufficient accuracy caused by traditional detection relying on a single feature. The Hill muscle model is introduced to simulate the theoretical correlation curve and perform deviation check on the measured signal. When the deviation exceeds the limit, the radar wave attenuation coefficient is calculated based on the three-dimensional muscle model reconstructed from the CT image, and secondary calibration is achieved through amplitude compensation and eigenvalue correction, effectively adapting to individual anatomical structure differences and reducing interference caused by different distributions of fat layers and muscle fibers. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall workflow diagram of the present invention;

[0048] Figure 2 It is a schematic diagram of the overall structure of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] See also Figure 1 As shown, one of the purposes of this embodiment is to provide a muscle tension detection method, comprising the following steps:

[0051] S1. Scan the target individual muscle group area through the millimeter wave radar array to obtain the muscle surface micro-vibration signal, and generate a time-frequency graph based on the muscle surface micro-vibration signal;

[0052] S2. Extracting features from the time-frequency graph to obtain characteristic parameters, the characteristic parameters including the amplitude of the main vibration frequency, the proportion of low-frequency band energy, and the energy entropy of the mid-frequency band;

[0053] S3. Generate low-frequency band weights and mid-frequency band weights based on the ratio of the main vibration frequency amplitude to the preset spasm threshold, and calculate a fusion eigenvalue based on the low-frequency band weight, the mid-frequency band weight, the low-frequency band energy proportion, and the mid-frequency band energy entropy. Input the fusion eigenvalue into the pre-trained muscle tension grading model to output a preliminary grade. Simultaneously, simulate the theoretical correlation curve between the muscle fiber contraction force and the vibration spectrum of the time-frequency graph at the current preliminary grade based on the Hill muscle model. If the deviation between the measured main vibration frequency amplitude and the theoretical correlation curve exceeds the preset threshold, activate S4.

[0054] To dynamically adjust the weights of characteristic parameters in different frequency bands based on muscle vibration characteristics to more accurately reflect muscle tension, S3 generates low-frequency and mid-frequency weights based on the ratio of the main vibration frequency amplitude to the preset spasm threshold. Specifically, they include:

[0055] Establish a spasm threshold library associated with muscle anatomical location and body type. Because the critical points of muscle tension in muscles at different anatomical locations and in individuals of different body types vary, thresholds need to be set specifically. The construction of the spasm threshold library requires collecting muscle tension data from different populations, classifying the muscles by anatomical location, and further subdividing each category by body mass index interval and muscle group depth range. The corresponding spasm threshold is labeled for each subdivision, and the preset spasm threshold is matched according to the body mass index and muscle group depth of the target individual. Because it is necessary to find a reference standard that best fits the physiological characteristics of the current test subject, during the test, the body mass index of the target individual is first obtained through a body fat scale, and the muscle group depth is estimated through signal analysis of the millimeter wave radar array. Then, the subdivision matching these two parameters is retrieved from the spasm threshold library, and the spasm threshold corresponding to the subdivision is extracted as the preset spasm threshold for the current test, making the deviation degree of subsequent calculations more targeted;

[0056] Calculate the dynamic deviation coefficient of the main vibration frequency amplitude relative to the preset spasm threshold. The dynamic deviation coefficient reflects the degree of proximity between the current muscle state and the spasm critical point. The calculation method is to divide the main vibration frequency amplitude by the preset spasm threshold. The resulting value is the dynamic deviation coefficient. For example, if the main vibration frequency amplitude is a certain value and the preset spasm threshold is another value, if the coefficient obtained by dividing the two is 0.6, it means that the current state is still a certain distance away from the spasm critical point; if it is 0.9, it indicates that it is close to the spasm state. This coefficient intuitively presents the trend of the muscle developing towards spasm. Then, compare the dynamic deviation coefficient with the preset spasm risk interval. According to the comparison result, the spasm risk interval in which the dynamic deviation coefficient is located is determined, and weight distribution is performed according to the spasm risk interval to generate low-frequency band weights and medium-frequency band weights. Because the low-frequency band and medium-frequency band characteristics have different degrees of influence on muscle tension grading in different risk intervals, the preset spasm risk interval is divided into three levels: low risk, medium risk, and high risk. For example, a dynamic deviation coefficient less than 0.5 is low risk, 0.5 to 0.8 is medium risk, and greater than 0.8 is high risk; weight distribution The rule is: in the low-risk interval, the low-frequency band weight is set to 0.3 and the medium-frequency band weight is set to 0.7, focusing on the medium-frequency band features; in the medium-risk interval, the low-frequency band weight and the medium-frequency band weight are both set to 0.5, and the two are balanced; in the high-risk interval, the low-frequency band weight is set to 0.7 and the medium-frequency band weight is set to 0.3, focusing on the low-frequency band features, because muscles are more prone to spasm-related low-frequency vibrations at high risk, and the weight of the low-frequency band features needs to be increased. Finally, the spasm risk index is formed according to the product of the dynamic deviation coefficient and the low-frequency band weight, which is used for priority control of the fusion feature value. During calculation, the dynamic deviation coefficient is multiplied by the current low-frequency band weight, and the resulting value is the spasm risk index. For example, a dynamic deviation coefficient of 0.9 is in the high-risk range, and the low-frequency band weight is 0.7. The multiplication results in a spasm risk index of 0.63. The higher the index, the greater the risk of muscle spasm. When the subsequent fusion eigenvalue is calculated, the compensation item corresponding to the index will be given a higher priority, so that the fusion result can better reflect the potential spasm risk of the muscle, improve the sensitivity of the grading model to high-risk states, and lay the foundation for the accurate calculation of the subsequent fusion eigenvalue.

[0057] In order to integrate multi-dimensional feature parameters to accurately reflect the muscle tension state and eliminate noise and artifact interference, S3 calculates the fusion feature value based on the low-frequency band weight, mid-frequency band weight, low-frequency band energy ratio, and mid-frequency band energy entropy, as follows:

[0058] A motion artifact correction module is introduced. When involuntary muscle tremor is detected, the spasm risk index is used to adjust the contribution of the low-frequency band energy ratio. This is because involuntary tremors, such as those caused by Parkinson's disease, will produce low-frequency signals that are unrelated to muscle tension and interfere with the detection results. The motion artifact correction module analyzes the regularity of the low-frequency band signal in the time-frequency graph. If the signal is found to exhibit periodic, high-amplitude fluctuations and is inconsistent with the normal muscle contraction pattern, it is determined to be involuntary muscle tremor. At this time, the contribution of the low-frequency band energy ratio is adjusted according to the spasm risk index: the higher the spasm risk index, that is, the closer the muscle is to the spasm state, the higher the proportion of the true tension-related components in the low-frequency signal, so the adjustment amplitude is reduced, such as only reducing the contribution by 10%. The lower the spasm risk index, the higher the proportion of artifacts in the low-frequency signal may be, so the adjustment amplitude is increased, such as reducing the contribution by 30%. This dynamic adjustment reduces the interference of artifacts on low-frequency features.

[0059] A physiological noise suppression strategy for mid-band energy entropy is designed. If the current mid-band energy entropy value deviates from the historical mean by more than the preset tolerance, the sliding window mean is used to replace the outlier. This is because the mid-band energy entropy is easily affected by physiological noise such as breathing and heartbeat, resulting in abnormal fluctuations in the value. The implementation steps of the physiological noise suppression strategy are as follows: first, store the mid-band energy entropy values ​​of the target individual for the past three tests, calculate its historical mean, detect the current mid-band energy entropy value, and if the difference with the historical mean exceeds the preset tolerance, it is determined to be an outlier. At this time, the sliding window mean is used for replacement, that is, the mid-band energy entropy values ​​of two test points before and after the current moment are taken, and the average of these five points is calculated as the correction value at the current moment. The influence of instantaneous noise is eliminated through smoothing to ensure the stability of the mid-band characteristics.

[0060] The fused feature value consists of four parts: the product of the calibrated low-frequency band energy proportion and the low-frequency band weight, the product of the mid-frequency band energy entropy and the mid-frequency band weight, the entropy value credibility weighting coefficient, and the compensation term of the spasm risk index. Multi-component fusion achieves comprehensive integration of feature information. The specific calculation process is as follows:

[0061] The product of the calibrated low-frequency band energy proportion and the low-frequency band weight is calculated by multiplying the motion-artifact-corrected low-frequency band energy proportion by the corresponding low-frequency band weight to highlight the differentiated low-frequency features across different risk intervals. The product of the mid-frequency band energy entropy and the mid-frequency band weight is calculated by multiplying the noise-suppressed mid-frequency band energy entropy by the corresponding mid-frequency band weight to reflect the contribution of mid-frequency features to the current risk level. The entropy credibility weighting coefficient is dynamically set based on the stability of the mid-frequency band energy entropy. If the entropy fluctuation of the five most recent detection points is less than a preset threshold, the coefficient is set to 1.0; otherwise, it is set to 0.8 to reduce the influence of unstable features. The compensation term for the spasticity risk index is calculated by multiplying the spasticity risk index by a fixed coefficient of 0.2 as an additional supplementary term to enhance the influence of high-risk states on the fusion results. The sum of these four values ​​yields the final fusion feature value. This value integrates the key features of the high and low frequency bands, reduces the interference of noise and artifacts through correction and compensation mechanisms, and highlights the impact of spasticity risk, providing more reliable input parameters for the subsequent muscle tension grading model.

[0062] To accurately output a preliminary level of muscle tension based on the fused feature values ​​while simultaneously meeting the requirements of rapid identification of high-risk states and comprehensive judgment of multiple features, S3 inputs the fused feature values ​​into a pre-trained muscle tension grading model to output a preliminary level, further including:

[0063] A two-tiered decision-making model was constructed, improving classification efficiency and accuracy through layered processing. The first tier implements spasm risk assessment, as high-risk conditions approaching spasm require rapid identification and priority treatment. The spasm risk index's alert value is set based on clinical data. When the spasm risk index exceeds the alert value, a high-risk level is directly output without further analysis, enabling a rapid response to emergencies. If the alert value is not exceeded, the second tier initiates multi-feature collaborative classification. The second tier uses a random forest algorithm to analyze and fuse feature values ​​with additional physiological parameters. These additional physiological parameters include the target individual's age, muscle fatigue (estimated from pre-test exercise records), and underlying medical history (such as stroke, cerebral palsy, and other conditions that may affect muscle tone). These parameters are retrieved from the system's linked health records. The random forest algorithm constructs multiple decision trees to analyze the fused feature values ​​and additional physiological parameters separately, and then combines the output results of all decision trees to determine the initial grade. Multi-feature collaborative analysis can reduce misjudgments caused by single feature deviations and improve the comprehensiveness of grading. The second level initiates multi-feature collaborative classification, using the random forest algorithm to analyze the fused feature values ​​and additional physiological parameters.

[0064] A time-domain stability check is introduced, which modifies the preliminary grade based on the mode distribution of the three most recent test results. Because a single test may be affected by transient interference, the stability of multiple results can better reflect the true muscle state. During operation, the system automatically stores the preliminary grade of the current test and the grades of the previous two tests, and calculates the mode of these three grades (that is, the grade with the most occurrences). If the current preliminary grade is consistent with the mode, the grade is maintained. If they are inconsistent, and the mode occurs two or more times, the preliminary grade is corrected to the grade corresponding to the mode. If the three grades are different, the current preliminary grade is retained but marked as "pending confirmation", indicating that more tests are needed. For example, if the three most recent test grades are medium risk, medium risk, and high risk, respectively, and the mode is medium risk, if the current preliminary grade is high risk, it will be revised to medium risk. If the three grades are low tension, normal tension, and high risk, respectively, the current grade is retained and marked as pending confirmation. This time-domain stability check can effectively filter out abnormal results caused by transient noise.

[0065] In order to verify the rationality of the measured muscle vibration signal through the theoretical model, S3 simultaneously simulates the theoretical correlation curve based on the Hill muscle model. The specific implementation is as follows:

[0066] The preliminary level is mapped to the muscle fiber activation rate, because the muscle fiber activation rate is the core parameter reflecting the strength of muscle contraction and is directly related to the muscle tension level. The mapping rules are set according to clinical research, and a muscle contraction meta-force calculation model is constructed. This model is used to simulate the force generated by active muscle contraction. The model includes a linear conversion relationship between contraction speed and vibration frequency and a force value growth function related to muscle fiber activation rate. The functional relationship is based on muscle physiology research. When the activation rate increases from 0% to 100%, the contraction meta-force gradually increases from 0 to the maximum value. Through these two parts of the relationship, the corresponding contraction meta-force and related vibration frequency characteristics can be calculated according to the current muscle fiber activation rate.

[0067] A parallel elastic element-force calculation model is constructed because muscles will also generate passive elastic force due to elastic deformation when contracting, which needs to be included in the comprehensive consideration of the model. The parallel elastic element-force calculation model introduces muscle elongation, which is obtained from the muscle bundle geometry analysis of CT images. The muscle bundle length of the target muscle group is extracted from the CT image, and the difference between the current muscle bundle length and the muscle bundle length in the resting state is calculated, and then divided by the resting length to obtain the elongation. The parallel elastic element-force increases with the increase of muscle elongation. Its calculation rules refer to the study of the elastic properties of muscle tissue. Afterwards, the theoretical vibration spectrum amplitude is calculated by the vector synthesis results of the contraction element-force and the elastic element-force. Vector synthesis comprehensively considers the direction and magnitude of the contraction element-force (active force) and the elastic element-force (passive force). The resultant force generated by the joint action of the two determines the intensity of the micro-vibration on the muscle surface. The greater the resultant force, the higher the amplitude of the vibration spectrum in theory. According to the correspondence between the magnitude of the resultant force and the vibration amplitude (summarized based on a large amount of measured data), the theoretical amplitude at different vibration frequencies is calculated, and finally a frequency-amplitude theoretical correlation curve is established. The horizontal axis is the vibration frequency, covering the common frequency range of muscle micro-vibration, and the vertical axis is the theoretical amplitude. By connecting the theoretical amplitudes of each frequency point in frequency order, a complete theoretical correlation curve is formed. This curve reflects the theoretical relationship between vibration frequency and amplitude under the muscle state corresponding to the current preliminary level, providing a standard reference for subsequent comparison of measured signal deviations and a scientific theoretical basis.

[0068] S4. Reconstruct the three-dimensional muscle model based on the CT image of the target individual, calculate the attenuation coefficient of the radar wave in the fat / muscle fiber layer, use the attenuation coefficient to attenuate the main vibration frequency amplitude and the fusion eigenvalue, input the attenuation-compensated fusion eigenvalue into the muscle tension grading model again, and output the final tension level.

[0069] To accurately construct a model that reflects the muscle's anatomical structure and electromagnetic properties and provides a basis for subsequent radar wave attenuation coefficient calculations, S4 reconstructs a 3D muscle model based on the target individual's CT images. This includes:

[0070] A deep learning segmentation network is used to identify the boundaries of the fat layer, the direction of the myofascia, and the spatial topology of muscle fibers. These structures are key anatomical features that affect radar wave propagation and require precise distinction to ensure the authenticity of the model. The deep learning segmentation network uses a convolutional neural network trained on a large number of human muscle CT images. After inputting the CT image of the target individual (covering continuous tomographic images of the detected muscle groups), the network automatically identifies and marks the boundaries of different tissues through multi-layer feature extraction: the boundary of the fat layer appears as the edge of the area with a low CT value (darker grayscale in the image), the direction of the myofascia is a thin layer of connective tissue that separates muscle bundles (medium CT value, linear distribution), and the spatial topology of muscle fibers is determined by identifying the direction, thickness, and arrangement of muscle bundles (cord-like, extending along the direction of muscle contraction). After identification is completed, the system generates a segmentation result map with different structural labels to provide structured anatomical data for three-dimensional reconstruction. Then, a dielectric property layer model is constructed based on the segmentation results. Among them, the fat layer is defined as a uniform dielectric layer with a fixed low dielectric constant. The muscle fiber layer is defined as an anisotropic structure with a dielectric constant significantly higher along the fiber direction than in the vertical direction. Because the dielectric properties of different tissues (i.e., their ability to conduct and store electromagnetic waves) vary significantly, they directly affect the attenuation of radar waves and require layered definition to accurately simulate electromagnetic wave propagation. Specifically, the fat layer is defined as a uniform dielectric layer with a relatively uniform internal structure. The attenuation pattern of electromagnetic waves propagating through it is stable, so its dielectric constant is fixed to a low constant (set by referring to the electromagnetic properties measurement data of human fat tissue). The muscle fiber layer is defined as an anisotropic structure. Because muscle fibers are arranged along a specific direction (for example, limb muscles are mostly along the long axis), electromagnetic waves encounter less obstruction when propagating along the fiber direction. When propagating in the perpendicular direction, the attenuation is more significant due to the dense fiber arrangement. Therefore, the dielectric constant along the fiber direction is set to be significantly higher than that in the perpendicular direction (referring to the research data on the electromagnetic properties of muscle fiber tissue. The ratio of the two is set according to the muscle type. For example, the anisotropy difference of striated muscle is greater than that of smooth muscle). Through this step, anatomical structural features are accurately extracted from CT images and then combined with the electromagnetic properties of the tissue to construct a layered model. The reconstructed three-dimensional muscle model not only reflects the actual spatial structure but also accurately simulates the propagation characteristics of radar waves in different tissues, providing a precise physical model foundation for the subsequent calculation of the radar wave attenuation coefficient.

[0071] In order to accurately calculate the energy loss of radar waves in muscle tissue and provide a basis for subsequent signal compensation, the radar wave attenuation coefficient is calculated in S4. The specific steps are as follows:

[0072] Calculate the effective propagation path. The effective propagation path length is determined by the radar incident angle, the angle of muscle fiber orientation, and the thickness of the fat layer. The propagation path is inversely proportional to the cosine value of the angle, as the radar wave's incident angle and muscle fiber orientation directly affect the actual distance it travels through tissue. The fat layer thickness is a fixed physical length along the path. Specifically, the radar incident angle is obtained from the radar probe's installation parameters; the muscle fiber orientation angle is extracted from the 3D muscle model, which is the angle between the muscle fiber arrangement and the radar incident direction; and the fat layer thickness is obtained from the fat layer boundary identification results in the CT image. The calculation of the effective propagation path length requires a comprehensive consideration of these three parameters: the smaller the cosine value of the angle (the larger the angle), the longer the propagation path; the thicker the fat layer, the longer the propagation path. For example, when the radar incident angle is 30 degrees (larger cosine value), the muscle fiber strike angle is 30 degrees (larger cosine value), and the fat layer thickness is 1 cm, the effective propagation path length is short. If the incident angle increases to 60 degrees (decreased cosine value), the strike angle increases to 60 degrees (decreased cosine value), and the fat layer thickness increases to 2 cm, the path length increases significantly, accurately reflecting the actual distance the radar wave travels through the tissue. The radar wave attenuation coefficient is then composed of the superposition of three parts. Because the radar wave is absorbed by the fat layer and the muscle fibers in the parallel and perpendicular directions during propagation, it is attenuated and needs to be calculated separately and summarized:

[0073] The fat layer attenuation term increases exponentially with thickness. The thicker the fat layer, the more radar energy is absorbed, and the more pronounced the attenuation. The calculations are based on the absorption characteristics of millimeter waves by fat tissue. For every 0.5 cm increase in thickness, the attenuation term increases at a fixed rate, reflecting this exponential growth characteristic.

[0074] The attenuation term parallel to the muscle fibers is positively correlated with the effective propagation path length and high dielectric constant. This is because when radar waves propagate along the muscle fibers, they have a larger contact area with the fibers. Furthermore, the dielectric constant parallel to the muscle fibers is higher, resulting in greater energy absorption. The longer the path and the higher the dielectric constant, the more significant the attenuation. This attenuation term is calculated by multiplying the effective propagation path length by the dielectric constant parallel to the muscle fibers, and then multiplying it by a fixed proportional coefficient (set based on experimental data on the electromagnetic properties of muscle tissue).

[0075] The attenuation term perpendicular to muscle fibers is weakly correlated with the effective propagation path length and low dielectric constant. This is because when radar waves pass perpendicularly through muscle fibers, the cross-contact area with the fibers is small, and the dielectric constant in the perpendicular direction is low, resulting in less energy absorption and, therefore, less attenuation. The calculation method is similar to that for the parallel direction, but the proportionality coefficient is set to a smaller value (e.g., 1 / 3 of the parallel coefficient) to reflect this weak correlation. The sum of these three attenuation terms is the radar wave attenuation coefficient, which comprehensively reflects the energy loss caused by different tissues and directions during radar wave propagation and provides a precise quantitative basis for subsequent attenuation compensation.

[0076] In order to eliminate signal distortion caused by tissue absorption during radar wave propagation and make the characteristic parameters more consistent with the actual muscle state, S4 uses the attenuation coefficient for attenuation compensation, specifically including:

[0077] ①. Compensate for the amplitude of the main vibration frequency and use the inverse of the square root of the attenuation coefficient to restore the amplitude. Because the amplitude of the main vibration frequency is greatly affected by attenuation, its true level needs to be restored by inverse correlation with the degree of attenuation. In specific operations, first obtain the calculated radar wave attenuation coefficient, calculate the square root of the coefficient, and then take its inverse to obtain the amplitude recovery coefficient. Multiply the measured amplitude of the main vibration frequency by this recovery coefficient to obtain the compensated amplitude. For example, when the attenuation coefficient is large, the inverse of its square root is small, which means that the measured amplitude needs to be increased more significantly to offset the attenuation effect. In this way, the amplitude can truly reflect the intensity of muscle vibration;

[0078] ②. Perform fusion eigenvalue compensation and use the logarithmic function of the attenuation coefficient for nonlinear correction. Because the fusion eigenvalue contains multi-dimensional information, its attenuation law is nonlinear, and the logarithmic function can better adapt to this complex relationship. During operation, first take the logarithm of the attenuation coefficient (if the attenuation coefficient is 0, take 0 to avoid meaningless values), use the obtained logarithmic value as the correction factor, and add the correction factor to the fusion eigenvalue (if the logarithm is negative, it is equivalent to subtracting a positive value) to achieve nonlinear adjustment. For example, when the attenuation coefficient is small, the logarithmic correction factor is small, and the adjustment amplitude of the fusion eigenvalue is also small; when the attenuation coefficient is large, the absolute value of the correction factor increases, and the adjustment amplitude increases accordingly, so as to balance the distortion of the eigenvalues ​​under different attenuation levels;

[0079] ③. Normalization of muscle cross-sectional area. Normalize the compensated eigenvalues ​​based on the physiological parameters extracted from the CT model. Because the cross-sectional area of ​​muscles varies among individuals (e.g., athletes have thicker muscles), it will affect the absolute value of the vibration signal, and this individual difference needs to be eliminated. Extract the cross-sectional area of ​​the target muscle group from the three-dimensional muscle model reconstructed by CT (take the cross-sectional area of ​​the thickest part of the muscle), calculate the ratio of this cross-sectional area to the average cross-sectional area of ​​healthy people of the same age and gender, and obtain the normalization coefficient. Divide the compensated fusion eigenvalue by this normalization coefficient to free the eigenvalue from the influence of muscle size and achieve comparability between different individuals. For example, if the muscle cross-sectional area of ​​an individual is 1.2 times the average value, divide the compensated eigenvalue by 1.2 and standardize it to a value corresponding to the average level.

[0080] ④. A confidence label is generated simultaneously when outputting the final tension level. If the level jump before and after calibration exceeds the clinical tolerance, a manual review alert is triggered to ensure the reliability of the output results and respond to abnormal situations. The final tension level is obtained by inputting the compensated fusion eigenvalue into the muscle tension grading model. The confidence label is generated based on the stability of the compensation process: if the attenuation coefficient calculation is stable and the eigenvalue change pattern before and after compensation is as expected, it is marked as "high confidence"; if the attenuation coefficient fluctuates significantly or abnormal value replacement occurs during the compensation process, it is marked as "medium confidence" or "low confidence." At the same time, by comparing the preliminary grade with the final tension grade, if the grade jump exceeds the clinical tolerance (such as jumping directly from "normal tension" to "high-risk tension", the span is too large), the system will automatically trigger a manual review alert, and notify the clinician through a pop-up window or text message, prompting a manual review of the test result, and judging the final grade based on the patient's clinical manifestations to avoid misdiagnosis due to technical errors. Through these four steps of compensation and verification, the influence of radar wave attenuation and individual anatomical differences on characteristic parameters is eliminated, and the clinical reliability of the results is guaranteed through confidence labels and manual review mechanisms, so that the output final tension grade can more accurately reflect the true tension state of the muscle.

[0081] See also Figure 2 As shown, the second object of the present invention is to provide a system for implementing a muscle tension detection method including any one of the above items, comprising:

[0082] The signal acquisition feature generation unit is equipped with a phased array radar probe to scan the target muscle group area, capture the original micro-vibration signal, input the original signal into the short-time Fourier transform and convert it into a time-frequency graph. The main vibration frequency amplitude, low-frequency band energy ratio and mid-frequency band energy entropy are extracted and transmitted to the intelligent classification and verification unit, where:

[0083] The intelligent grading and verification unit calls the spasm threshold library, inputs the main vibration frequency amplitude, low-frequency band energy ratio and mid-frequency band energy entropy into the spasm threshold library to calculate the dynamic deviation coefficient, generates the low-frequency band weight, mid-frequency band weight and spasm risk index, calibrates the low-frequency band energy ratio based on the spasm risk index, suppresses the mid-frequency band energy entropy noise, outputs the fused eigenvalue, and screens high-risk cases according to the spasm risk index. The fused eigenvalue is input into the random forest model to output a preliminary grade, generates a theoretical correlation curve based on the preliminary grade, verifies the measured main vibration frequency amplitude deviation and generates a calibration instruction, and transmits the preliminary grade, fused eigenvalue and calibration instruction to the anatomical calibration unit;

[0084] The anatomical calibration unit uses a convolutional neural network to segment the fat layer and the muscle fiber layer, reconstruct a three-dimensional muscle model including dielectric properties, calculates the radar wave attenuation coefficient based on the muscle model, performs inverse square root compensation on the main vibration frequency amplitude, implements logarithmic correction on the fused eigenvalues, and then uses the muscle fiber cross-sectional area to standardize the compensated eigenvalues. The standardized eigenvalues ​​are input into the grading model to generate the final tension level. The level differences before and after calibration are compared to output confidence labels and clinical warnings.

[0085] The third object of the present invention is to provide a device for implementing a muscle tension detection method including any one of the above items, including a phased array probe array, a signal preprocessing circuit, a data acquisition interface, a hierarchical processing module, a feature fusion processor, a two-level decision engine, a Hill model simulator, a CT image processor, an attenuation compensation calculation module and a normalized output terminal.

[0086] The millimeter-wave radar array is used to acquire micro-vibration signals on the muscle surface and generate a time-frequency diagram. Feature parameters such as the main vibration frequency amplitude are extracted, and the eigenvalues ​​are fused and output through a pre-trained model to obtain a preliminary grade. The deviation is verified using the Hill muscle model. If the deviation exceeds the limit, a three-dimensional muscle model is reconstructed based on the CT image, and the radar wave attenuation coefficient is calculated for compensation. The final grade is output to solve the problems of insufficient accuracy and poor adaptability of traditional detection to individual differences, improve the accuracy and personalization of muscle tension detection, and provide a reliable basis for clinical diagnosis.

[0087] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A muscle tension detection method, characterized in that: The following steps are involved: S1. Scan the target individual muscle group area through the millimeter wave radar array to obtain the muscle surface micro-vibration signal, and generate a time-frequency graph based on the muscle surface micro-vibration signal; S2. Extracting features from the time-frequency graph to obtain characteristic parameters, the characteristic parameters including the amplitude of the main vibration frequency, the proportion of low-frequency band energy, and the energy entropy of the mid-frequency band; S3. Generate low-frequency band weights and mid-frequency band weights based on the ratio of the main vibration frequency amplitude to the preset spasm threshold, and calculate a fusion eigenvalue based on the low-frequency band weight, the mid-frequency band weight, the low-frequency band energy proportion, and the mid-frequency band energy entropy. Input the fusion eigenvalue into the pre-trained muscle tension grading model to output a preliminary grade. Simultaneously, simulate the theoretical correlation curve between the muscle fiber contraction force and the vibration spectrum of the time-frequency graph at the current preliminary grade based on the Hill muscle model. If the deviation between the measured main vibration frequency amplitude and the theoretical correlation curve exceeds the preset threshold, activate S4. S4. Reconstruct the three-dimensional muscle model based on the CT image of the target individual, calculate the attenuation coefficient of the radar wave in the fat / muscle fiber layer, use the attenuation coefficient to attenuate the main vibration frequency amplitude and the fusion eigenvalue, input the attenuation-compensated fusion eigenvalue into the muscle tension grading model again, and output the final tension level.

2. A muscle tension detection method according to claim 1, characterized in that: In S3, the low-frequency band weight and the mid-frequency band weight are generated according to the ratio of the main vibration frequency amplitude to the preset spasm threshold, specifically including: Establish a library of spasm thresholds associated with muscle anatomical location and body shape, and match the corresponding preset spasm thresholds according to the target individual's body mass index and muscle depth; Calculate the dynamic deviation coefficient of the main vibration frequency amplitude relative to the preset spasm threshold, which reflects the degree of proximity between the current muscle state and the spasm critical point; Comparing the dynamic deviation coefficient with the preset spasm risk interval, determining the spasm risk interval in which the dynamic deviation coefficient is located based on the comparison result, and performing weight distribution based on the spasm risk interval to generate low-frequency band weights and mid-frequency band weights; The spasm risk index is constructed based on the product of the dynamic deviation coefficient and the low-frequency band weight, and is used for priority control of the fusion eigenvalue.

3. A muscle tension detection method according to claim 2, characterized in that: In S3, the fusion feature value is calculated according to the low-frequency band weight, the mid-frequency band weight, the low-frequency band energy proportion, and the mid-frequency band energy entropy, as follows: A motion artifact correction module is introduced. When involuntary muscle tremor is detected, the contribution of low-frequency band energy is adjusted using the spasm risk index. Design a physiological noise suppression strategy based on mid-band energy entropy. If the current mid-band energy entropy value deviates from the historical mean by more than a preset tolerance, the sliding window mean is used to replace the outlier. The fusion eigenvalue consists of four parts, namely, the product of the calibrated low-frequency band energy proportion and the low-frequency band weight, the product of the mid-frequency band energy entropy and the mid-frequency band weight, the entropy value credibility weighting coefficient, and the compensation term of the spasm risk index.

4. A muscle tension detection method according to claim 3, characterized in that: The step S3 inputs the fused feature value into a pre-trained muscle tension grading model to output a preliminary grade, further comprising: A two-tiered decision-making model was constructed. The first tier implemented spasticity risk assessment. When the spasticity risk index exceeded the warning value, a high-risk level was directly output. The second tier initiated multi-feature collaborative classification, using a random forest algorithm to analyze the fused feature values ​​and additional physiological parameters. A time domain stability check is introduced, which modifies the preliminary grade based on the mode distribution of the last three test results.

5. A muscle tension detection method according to claim 4, characterized in that: The synchronization in S3 is based on the simulation of the theoretical correlation curve based on the Hill muscle model, which is specifically implemented as follows: The preliminary level was mapped to the muscle fiber activation rate, and a muscle contraction force calculation model was constructed. The model includes a linear conversion relationship between contraction speed and vibration frequency and a force value growth function related to muscle fiber activation rate. A parallel elastic element force calculation model is constructed. The parallel elastic element force calculation model introduces muscle elongation. The muscle elongation is derived from the muscle bundle geometry analysis of CT images. The theoretical vibration spectrum amplitude is calculated by the vector synthesis results of the contraction element force and the elastic element force. Finally, a frequency-amplitude theoretical correlation curve is established, with the horizontal axis representing the vibration frequency and the vertical axis representing the theoretical amplitude.

6. A muscle tension detection method according to claim 5, characterized in that: The step S4 of reconstructing a muscle three-dimensional model based on the CT image of the target individual specifically includes: A deep learning segmentation network is used to identify the boundaries of the fat layer, the direction of the myofascia, and the spatial topology of the muscle fibers, and a dielectric property layer model is constructed, in which the fat layer is defined as a uniform dielectric layer and the muscle fiber layer is defined as an anisotropic structure.

7. A muscle tension detection method according to claim 6, characterized in that: The radar wave attenuation coefficient is calculated in S4, and the specific steps are: Calculate the effective propagation path. The effective propagation path length is determined by the radar incident angle, the muscle fiber orientation angle, and the thickness of the fat layer. The propagation path is inversely proportional to the cosine value of the angle. The radar wave attenuation coefficient is composed of three parts: Fat layer attenuation term: increases exponentially with thickness; Attenuation term parallel to muscle fibers: positively correlated with effective propagation path length and high dielectric constant; Attenuation term in the vertical direction of muscle fibers: weakly correlated with the effective propagation path length and low dielectric constant.

8. A muscle tension detection method according to claim 7, characterized in that: The attenuation compensation is performed using the attenuation coefficient in S4, specifically including: ①. Main vibration frequency amplitude compensation: Use the inverse square root of the attenuation coefficient to restore the amplitude; ② Fusion eigenvalue compensation: Use the logarithmic function of the attenuation coefficient for nonlinear correction; ③ Muscle cross-sectional area normalization: Normalize the compensated eigenvalues ​​based on the physiological parameters extracted from the CT model; ④. A confidence label is generated synchronously when the final tension level is output. When the level jump before and after calibration exceeds the clinical tolerance, a manual review alert is triggered.

9. A system for implementing a muscle tension detection method according to any one of claims 1 to 8, characterized in that: include: The signal acquisition feature generation unit is equipped with a phased array radar probe to scan the target muscle group area, capture the original micro-vibration signal, input the original signal into the short-time Fourier transform and convert it into a time-frequency graph. The main vibration frequency amplitude, low-frequency band energy ratio and mid-frequency band energy entropy are extracted and transmitted to the intelligent classification and verification unit, where: The intelligent grading and verification unit calls the spasm threshold library, inputs the main vibration frequency amplitude, low-frequency band energy ratio and mid-frequency band energy entropy into the spasm threshold library to calculate the dynamic deviation coefficient, generates the low-frequency band weight, mid-frequency band weight and spasm risk index, calibrates the low-frequency band energy ratio based on the spasm risk index, suppresses the mid-frequency band energy entropy noise, outputs the fused eigenvalue, and screens high-risk cases according to the spasm risk index. The fused eigenvalue is input into the random forest model to output a preliminary grade, generates a theoretical correlation curve based on the preliminary grade, verifies the measured main vibration frequency amplitude deviation and generates a calibration instruction, and transmits the preliminary grade, fused eigenvalue and calibration instruction to the anatomical calibration unit; The anatomical calibration unit uses a convolutional neural network to segment the fat layer and the muscle fiber layer, reconstruct a three-dimensional muscle model including dielectric properties, calculates the radar wave attenuation coefficient based on the muscle model, performs inverse square root compensation on the main vibration frequency amplitude, implements logarithmic correction on the fused eigenvalues, and then uses the muscle fiber cross-sectional area to standardize the compensated eigenvalues. The standardized eigenvalues ​​are input into the grading model to generate the final tension level. The level differences before and after calibration are compared to output confidence labels and clinical warnings.

10. A device for implementing a muscle tension detection method according to any one of claims 1 to 8, characterized in that: It includes a phased array probe array, a signal preprocessing circuit, a data acquisition interface, a hierarchical processing module, a feature fusion processor, a two-level decision engine, a Hill model simulator, a CT image processor, an attenuation compensation calculation module and a normalized output terminal.

Images