A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network

Abstract

The application relates to the field of computer vision, and particularly discloses a cervical vertebra curvature recognition and detection method based on an improved HRNet network, which comprises the following steps: acquiring infrared thermal imaging and a polarized light image sequence, generating a neck multi-modal fusion feature map through an improved HRNet backbone network, extracting cervical lateral fascia tension texture and cervical posterior triangle microvessel texture, indirectly positioning a cervical vertebra spinous process and determining a sternocleidomastoid muscle tension boundary through the texture intersection points; constructing a cervical vertebra curvature continuous reconstruction model, generating a time-varying curvature curve family, and decoupling a skull base rotation compensation component and a real curvature change component, decoupling a sternocleidomastoid muscle shortening, a deep cervical flexor muscle endurance failure and an upper trapezius muscle fascia adhesion mode; extracting a microtremor time-frequency feature and inputting the time-frequency feature into a dynamic grading network after verifying and removing artifacts through time sequence coherence, and outputting a risk level and a muscle imbalance subtype. The application breaks through a blind area of bony landmark positioning, and realizes noninvasive accurate detection and muscle imbalance typing.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and computer vision technology, and in particular to a method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network. Background Technology

[0002] Upper crossed syndrome is a postural disorder characterized by muscle imbalance. Its typical features include tension in the sternocleidomastoid and upper trapezius muscles, and weakness in the deep cervical flexors, leading to forward head posture, shoulder girdle protraction, and straightening of the cervical spine's physiological curvature. Accurate identification of abnormal cervical curvature and muscle imbalance patterns is of significant clinical value for early screening and massage intervention.

[0003] Existing detection technologies mainly include imaging detection and non-invasive detection using computer vision. Imaging detection relies on X-rays to measure the Cobb angle, which has limitations such as radiation exposure, high cost, and inability to perform dynamic monitoring. Computer vision-based methods utilize networks such as HRNet to extract key points in the head, neck, and shoulders to calculate the cervical spine angle, offering advantages such as being non-invasive and low-cost.

[0004] However, direct application faces several significant bottlenecks: First, HRNet is designed for whole-body posture estimation, with the neck region accounting for a very small percentage. Deep bony landmarks are not visually visible in ordinary images, and existing methods rely solely on soft tissue landmarks such as the ears and nose, completely losing information about the cervical spine's bony structure. Second, existing methods directly equate head tilt angle with changes in cervical curvature, ignoring the clinical decoupling phenomenon caused by skull base rotation compensation, resulting in an extremely high false positive rate. Third, existing methods isolate and analyze local cervical spine angles, failing to utilize multimodal skin physical features such as cervical fascia tension texture and microvascular texture to infer muscle imbalance patterns, and lacking non-invasive decoupling capabilities for subtypes such as sternocleidomastoid muscle shortening and deep cervical flexor muscle endurance failure. Fourth, dynamic detection fails to distinguish the frequency domain differences between physiological microtremors and pathological muscle tone abnormality tremors, and the granularity of respiratory and swallowing artifact removal is insufficient.

[0005] Therefore, there is an urgent need for a method to identify and detect cervical curvature in upper crossed syndrome that can integrate multimodal skin physical features to indirectly locate bony landmarks, decouple the craniocervical angle based on viscoelastic time-varying properties, and integrate micro-tremor frequency domain features to decouple muscle imbalance patterns. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network includes the following specific steps:

[0009] S1. Acquire a dynamic head, neck, and shoulder image sequence of the object to be detected. The image sequence includes infrared thermal imaging images and polarized light images acquired simultaneously under the same field of view. Input the image sequence into the improved HRNet backbone network. The improved HRNet backbone network includes a multimodal feature fusion attention module, a cervical fascia tension texture perception branch, a posterior cervical triangle microvascular texture perception branch, and a bony landmark indirect localization module. The multimodal feature fusion attention module adaptively fuses the features of the infrared thermal imaging images and polarized light images at different resolution scales to generate a cervical multimodal fusion feature map. The cervical fascia tension texture perception branch is used to detect features from the neck... The skin texture direction features of the lateral neck region are extracted from the multimodal fusion feature map. The skin color micro-change features of the posterior triangle region are extracted from the multimodal fusion feature map of the neck through the microvascular texture perception branch of the posterior triangle region. Based on the skin texture direction features of the lateral neck region and the skin color micro-change features of the posterior triangle region, the multimodal physical feature parameters of the neck are extracted. The above features are fused through the bony landmark indirect positioning module. The intersection of fascial tension texture is used as the indirect positioning basis for the skin projection of the cervical spinous process. The boundary of the texture direction change of the lateral neck region is used as the tension boundary of the sternocleidomastoid muscle. The multimodal physical feature parameters of the neck, the indirect positioning coordinate sequence of the cervical spinous process and the tension boundary of the sternocleidomastoid muscle are output.

[0010] S2. Using the indirect positioning coordinate sequence of the cervical spinous processes as control nodes, the multimodal physical feature parameters of the neck as local tissue state constraints, and the tension boundary of the sternocleidomastoid muscle as muscle state constraints, a continuous reconstruction model of cervical curvature based on the viscoelastic time-varying characteristics of the neck is constructed. The model is used to perform viscoelastic time-varying analysis on dynamic head, neck and shoulder image sequences to generate a family of viscoelastic time-varying cervical curvature curves, and the skull base rotation compensation component is decoupled from the actual cervical curvature change component. Based on the actual cervical curvature change component, the multimodal physical feature parameters of the neck, and the tension boundary of the sternocleidomastoid muscle, the sternocleidomastoid muscle shortening mode, the deep cervical flexor muscle endurance failure mode, and the upper trapezius muscle fascia adhesion mode are decoupled, and the cervical muscle mechanical load index is calculated. The actual cervical curvature morphological feature vector, muscle imbalance mode features, and cervical muscle mechanical load index are output.

[0011] S3. Perform S1 on the dynamic head, neck, and shoulder image sequence to obtain the features of each frame, and perform S2 to obtain the sequence-level analysis results, obtaining the true cervical curvature morphological feature vector, muscle imbalance pattern features, and cervical muscle mechanical load index corresponding to each frame; extract the time-frequency features of cervical microtremor through the cervical microtremor frequency domain analysis module to distinguish between physiological microtremors and pathological abnormal muscle tone tremors; fuse the time-frequency features with the true cervical curvature morphological feature vector through the temporal coherence verification module to remove craniocervical angle decoupling residuals and respiratory and swallowing artifacts; input the verified true cervical curvature morphological feature vector, together with the muscle imbalance pattern features and cervical muscle mechanical load index, into the upper crossed syndrome dynamic grading network for fusion, and output the upper crossed syndrome dynamic risk level and muscle imbalance subtype.

[0012] The beneficial effects of this invention are as follows:

[0013] 1. Breakthrough in overcoming the bottleneck of deep bony landmarks being visually invisible: By fusing infrared thermography with polarized light multimodal imaging, the projection of the cervical spinous process is indirectly located by utilizing the deviation of the tension texture of the cervical lateral fascia and the micro-color changes of the microvessels in the posterior cervical triangle. This solves the problem that existing methods rely solely on sparse soft tissue surface landmarks, resulting in the complete loss of bony structural information, and achieves accurate indirect location of deep bony landmarks under non-invasive conditions.

[0014] 2. Achieving clinical decoupling of cervical curvature and craniocervical angle, and precise classification of muscle imbalance: Based on the time-varying characteristics of cervical viscoelasticity, a continuous curvature reconstruction model is constructed. By utilizing the difference between the rigid trajectory of skull base rotation and the viscoelastic hysteresis response of cervical curvature, the model effectively distinguishes between skull base rotation compensation and true curvature changes. Multimodal physical features are integrated to decouple three muscle imbalance modes: sternocleidomastoid muscle shortening, deep cervical flexor muscle endurance failure, and upper trapezius muscle fascia adhesion, providing an objective classification basis for precise massage intervention.

[0015] 3. Improve the stability of dynamic detection and the accuracy of artifact removal: By extracting time-frequency features through frequency domain analysis of neck microtremors, physiological microtremors and pathological tremors of abnormal muscle tone are distinguished. Combined with temporal coherence verification, respiratory and swallowing artifacts and craniocervical angle decoupling residuals are effectively removed, which significantly improves the reliability and clinical applicability of muscle imbalance pattern determination and risk classification results in dynamic sequences. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method proposed in Embodiment 1 of the present invention;

[0017] Figure 2 This is an architecture diagram of the muscle imbalance mode decoupling network in Embodiment 1 of the present invention;

[0018] Figure 3 This is a framework diagram of the system proposed in Embodiment 2 of the present invention;

[0019] Figure 4 This is a comparison diagram of the effects of the present invention and the prior art in Embodiment 3 of the present invention. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] Example 1

[0022] Please see the appendix Figure 1 A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network includes the following specific steps:

[0023] S1. Obtain dynamic head, neck, and shoulder image sequences of the object to be detected. Generate a multimodal fusion feature map of the neck using an improved HRNet backbone network. Extract the lateral cervical fascia tension texture lines and determine the intersection points of the fascia tension textures. Extract the microvascular texture of the posterior cervical triangle and determine the high-sensitivity area for subtle skin color changes. Generate multimodal physical feature parameters of the neck, indirect positioning coordinate sequences of the cervical spinous processes, and the tension boundary of the sternocleidomastoid muscle. Specifically, this includes:

[0024] S11. Synchronous acquisition of multimodal images and generation of neck multimodal fusion feature maps, specifically including:

[0025] S111. The multimodal imaging module is fixed to the back side of the target object's neck. The horizontal distance between the module and the skin at the center of the back of the neck is set to 80cm to 120cm, preferably 100cm. The multimodal imaging module consists of an infrared thermal imaging unit and a polarization imaging unit rigidly integrated on the same optical substrate.

[0026] The infrared thermal imaging unit uses an uncooled focal plane infrared detector with a resolution of 640 pixels × 512 pixels, a spectral response band of 8μm to 14μm, and a frame rate of 25Hz to 30Hz, preferably 25Hz.

[0027] The polarization imaging unit uses a CMOS image sensor with a resolution of 2048 pixels × 2048 pixels. The frame rate is hardware synchronized with the infrared thermal imaging unit, with a synchronization error of less than 1ms. The front end of the polarization imaging unit is equipped with a ring-shaped LED illumination array. A linear polarizer is covered in front of the LED light source. The transmission axis of the polarizer is cross-polarized at 90 degrees with the transmission axis of the polarizer in front of the CMOS sensor to suppress surface specular reflection and enhance subcutaneous scattered light signals.

[0028] Before data acquisition, the target subject should remove any obstructions from the neck and be in a bare-neck state, with the head in a neutral position, eyes looking straight ahead, and shoulders relaxed. During the acquisition process, the subject should maintain natural breathing. The duration of a single dynamic acquisition should be set to 10 to 30 seconds, preferably 15 seconds, thereby forming a dynamic head, neck, and shoulder image sequence that includes infrared thermal imaging image sequences and polarized light image sequences.

[0029] S112. The improved HRNet backbone network includes a structurally symmetrical and parameter-independent infrared input branch and polarization input branch, as well as a multimodal feature fusion attention module. The infrared input branch begins with an initial convolutional layer consisting of two cascaded 3×3 convolutional kernels. The first and second layers each have 64 output channels, with a stride of 2. Each layer is followed by a batch normalization layer and a ReLU activation function, downsampling the input infrared thermal imaging image to 1 / 4 of its original resolution and extracting primary thermal distribution features. The polarization input branch has the same structure and is used to extract primary features of subcutaneous microvessels and surface texture from polarized light images.

[0030] The outputs of the two branches are fed into a parallel multi-resolution subnetwork, which consists of four stages. The first stage contains only a 1 / 4 resolution subnetwork with 32 channels; the second stage adds a 1 / 8 resolution subnetwork with 64 channels; the third stage adds a 1 / 16 resolution subnetwork with 128 channels; and the fourth stage adds a 1 / 32 resolution subnetwork with 256 channels. Features are extracted in each stage through repeated multi-resolution convolutional blocks, each containing 3×3 depthwise separable convolutions, 1×1 pointwise convolutions, batch normalization, and ReLU activation. Adjacent subnetworks are fused across resolutions using bilinear interpolation upsampling or 3×3 convolution downsampling with a stride of 2, generating infrared thermal imaging feature maps and polarized light image feature maps at each resolution scale.

[0031] S113. A multimodal feature fusion attention module fuses infrared thermal imaging feature maps and polarized light image feature maps at various resolution scales. First, the infrared and polarized feature maps at the same resolution scale are concatenated along the channel dimension to form a dual-channel fused feature map. Then, the concatenated feature map is fed into a channel attention unit. This unit first performs global average pooling on the fused feature map, compressing the spatial dimension to 1×1. Then, it passes through two fully connected layers. The first layer compresses the number of channels to 1 / 16 of the original number, and the second layer restores the original number of channels. Finally, a channel weight vector in the range of 0 to 1 is generated using the Sigmoid function. This weight vector is multiplied channel-by-channel with the fused feature map to complete the attention weighting process. Specifically:

[0032] ,

[0033] ,

[0034] ,

[0035] In the formula: This is the fused feature map after channel stitching. These represent the feature map height, width, and total number of channels, respectively. For the first Global average pooling statistics for each channel For the first The attention weights for each channel are determined by the Sigmoid function. generate, This is the weight matrix of the fully connected layer. This is the weighted fusion feature map.

[0036] The weighted feature maps at each scale are upsampled to a uniform resolution of 1 / 4 of the original input resolution using bilinear interpolation. The upsampled feature maps from the four scales are then concatenated along the channel dimension to generate a neck multimodal fusion feature map. The compression ratio of the channel attention unit ranges from 8 to 32, with 16 being preferred. The uniform resolution is fixed at 1 / 4 of the original input resolution. The total number of channels in the concatenated neck multimodal fusion feature map is the sum of the output channels at each scale.

[0037] S114. The images in the infrared thermal imaging sequence are divided into six anatomical partitions according to a preset anatomical partition template. The template uses the midsagittal plane of the posterior neck as a reference, dividing the posterior neck into six anatomical partitions: the upper cervical midline, the middle cervical midline, the lower cervical midline, the posterior border of the left sternocleidomastoid muscle, the posterior border of the right sternocleidomastoid muscle, and the junction of the anterior borders of the left and right trapezius muscles. The partition boundaries are defined as follows: the longitudinal axis of the cervical spinous processes is used as the center line; the width of each lateral partition is set to 30mm to 50mm, preferably 40mm; the longitudinal boundaries are the interspinous spaces of the upper cervical C1 to C2, the middle cervical C3 to C4, and the lower cervical C5 to C7. The arithmetic mean of the temperature values ​​of all pixels within each partition is calculated to generate the average skin temperature of each anatomical partition, i.e.: In the formula: For the first The total number of valid pixels in each partition For the first Within the first partition Temperature value of each pixel For the first The average skin temperature of each anatomical region is measured, ranging from 30℃ to 37.5℃. If the number of pixels in a region is less than 50% of the theoretical total number of pixels in that region, the region is considered to be under-covered, and the data for that frame is discarded and re-acquired.

[0038] S12. Extraction of cervical lateral fascia tension texture and localization of texture intersection points, specifically including:

[0039] S121. The cervical multimodal fusion feature map is input into the cervical fascial tension texture perception branch. This branch first performs anatomical regional localization. Based on the anatomical landmarks of the human neck, the area between the posterior border of the sternocleidomastoid muscle and the anterior border of the trapezius muscle is identified as the lateral cervical region in the cervical multimodal fusion feature map. The specific localization method is as follows: using the midsagittal plane of the posterior neck as the longitudinal baseline, it extends horizontally to both sides. The posterior border of the sternocleidomastoid muscle is located 25mm to 40mm from the baseline, and the anterior border of the trapezius muscle is located 45mm to 65mm from the baseline. Preferably, the posterior border of the sternocleidomastoid muscle is 30mm from the baseline, and the anterior border of the trapezius muscle is 55mm from the baseline. Using this horizontal range as the lateral boundary and the longitudinal projection range of the spinous processes of cervical vertebrae C1 to C7 as the longitudinal boundary, a rectangular lateral cervical region feature area is segmented and extracted from the cervical multimodal fusion feature map. The region has a horizontal pixel width of 80 to 150 pixels and a vertical pixel height of 200 to 350 pixels in the feature map space, preferably 100 pixels horizontally and 280 pixels vertically.

[0040] S122. A multi-scale Gabor filter bank is used to calculate the filtering response of the lateral neck region. This filter bank contains 4 to 8 scales, preferably 5 scales, and each scale contains 8 to 16 directions, preferably 12 directions. The spatial wavelength of the Gabor filter kernel function is set to 4 to 16 pixels, preferably 8 pixels; the spatial aspect ratio is set to 0.3 to 0.7, preferably 0.5; the phase offset is set to 0°; and the filter kernel window size is set to 21 pixels × 21 pixels to 31 pixels × 31 pixels, preferably 25 pixels × 25 pixels. Specifically:

[0041] ,

[0042] ,

[0043] ,

[0044] In the formula: This is the kernel function of the Gabor filter. For kernel function space coordinates, For spatial wavelength, The filter direction angle, ranging from 0 to... , Let the standard deviation be the Gaussian envelope. The aspect ratio of the space. , The coordinates of the kernel function space are respectively After rotation angle Transformed vertical and horizontal local coordinates For pixels In direction The maximum response amplitude on, For the first Gabor cores of various scales Image of the characteristic region of the lateral neck area. This is the size of the kernel function's half-window.

[0045] Gabor filter kernels at each scale and in each direction are convolved with the neck region feature area in two dimensions to obtain the filter response amplitude matrix. For each pixel, the direction corresponding to the maximum value among the response amplitudes at all scales and in all directions is selected as the local texture orientation angle of that pixel. The orientation angle ranges from 0° to 180°, with an angle resolution of 15° to 30°, preferably 15°.

[0046] S123. Divide the lateral cervical region into multiple lateral cervical segments according to the longitudinal segments of the cervical vertebrae. The division is based on the longitudinal projection position of the spinous processes of the cervical vertebrae. Divide the lateral cervical region into 5 to 7 lateral cervical segments, preferably 6 segments, with a longitudinal height of 30 to 50 pixels for each segment. Perform statistical analysis on the local texture orientation angles of all pixels in each segment, calculate the histogram distribution of orientation angles, and take the orientation angle with the highest frequency as the main texture orientation angle of that segment.

[0047] Subsequently, spatial difference operations are performed on the texture orientation angles of each segment, and the Sobel operator is used to calculate the orientation angle gradients in the horizontal and vertical directions respectively. The horizontal gradient kernel of the Sobel operator is convolved with the texture orientation angle matrix of the neck region segment to obtain the horizontal gradient component, and the vertical gradient kernel is convolved with the texture orientation angle matrix to obtain the vertical gradient component. The gradient magnitude is calculated by taking the square root of the sum of the squares of the horizontal and vertical gradient components to generate the texture orientation gradient field. Specifically:

[0048] ,

[0049] In the formula: This is the segmental texture direction angle matrix for the lateral neck region. For Sobel horizontal gradient kernel, , For Sobel vertical gradient kernel, , These are the gradient components in the horizontal and vertical directions. This represents the gradient magnitude of the texture direction gradient field, ranging from 0° / pixel to 90° / pixel. This is a two-dimensional convolution operation.

[0050] S124. Continuous boundaries in the texture direction gradient field where the gradient magnitude exceeds a preset threshold are taken as fascial tension texture lines. The preset threshold is set to 1.2 to 2.0 times the global average gradient magnitude of the texture direction gradient field, preferably 1.5 times, i.e.: , In the formula: This represents the global average gradient magnitude of the gradient field along the texture direction. Let be the number of rows and columns of the gradient field matrix. The threshold for extracting fascial tension texture lines. It is a multiplier, ranging from 1.2 to 2.0, with 1.5 being preferred.

[0051] After thresholding the gradient field of the texture direction, continuous boundaries are extracted using 8-neighborhood connected region analysis. Short boundary noise with a length of less than 20 pixels is removed, and continuous boundaries with a length of 20 pixels or more are retained as candidate texture lines. Linear fitting is performed on the candidate texture lines, and candidate texture lines whose slope of the fitted line deviates from the principal direction angle of the corresponding cervical lateral segment texture by less than 20° are identified as fascial tension texture lines. The spatial intersection point of multiple fascial tension texture lines is defined as the intersection point. The criteria for determining the intersection point are: there are 3 or more fascial tension texture lines intersecting within a 20-pixel × 20-pixel neighborhood, and the vertical coordinate of the intersection point is located within the range of 30% to 50% of the segment height.

[0052] S13. Extraction of microvascular texture in the posterior cervical triangle and identification of highly sensitive areas for subtle changes in skin color, specifically including:

[0053] S131. The polarized light image sequence is input into the microvascular texture sensing branch of the posterior triangular neck. This branch first performs subcutaneous tissue layer signal extraction on each frame of the polarized light image sequence. Since the polarized light images are acquired under cross-polarized illumination conditions, epidermal specular reflection has been suppressed, and subcutaneous tissue scattered light dominates. For each frame of the polarized light image, median filtering is first used to remove speckle noise. The filtering window is set to 3 pixels × 3 pixels to 7 pixels × 7 pixels, preferably 5 pixels × 5 pixels. Then, a homomorphic filtering algorithm is used to separate the low-frequency illumination component and the high-frequency detail component. The high-pass cutoff frequency of the homomorphic filter is set to 0.5π rad / pixel to 0.9π rad / pixel, preferably 0.7π rad / pixel, and the gain coefficient is set to 1.5 to 3.0, preferably 2.0.

[0054] The filtered image is linearly stretched to the grayscale range of 0 to 255, i.e.: In the formula: To stretch the subcutaneous tissue layer image in pixels grayscale value at that location For the homomorphically filtered image at the pixel level grayscale value at that location The global maximum and minimum gray values ​​of the homomorphically filtered image are used to obtain the enhanced subcutaneous tissue layer image, which highlights the gray-level fluctuations caused by the superficial dermal microvascular network, providing a basis for subsequent microvascular texture analysis.

[0055] S132. Divide the subcutaneous tissue layer image according to the anatomical boundaries of the posterior cervical triangle, which is bounded by the posterior border of the sternocleidomastoid muscle, the anterior border of the trapezius muscle, and the superior border of the clavicle. Using the posterior border of the sternocleidomastoid muscle as the medial boundary, the anterior border of the trapezius muscle as the lateral boundary, the horizontal line of the superior border of the clavicle as the lower boundary, and the horizontal line of the superior nuchal line as the upper boundary, crop the subcutaneous tissue layer image of the posterior cervical triangle from the subcutaneous tissue layer image.

[0056] The spatial distribution features of pixel grayscale values ​​are extracted pixel by pixel from the image, including the grayscale mean and grayscale spatial gradient. The grayscale mean is the arithmetic mean of the grayscale values ​​of all pixels in the region. Specifically:

[0057] ,

[0058] ,

[0059] ,

[0060] In the formula: The grayscale mean of the subcutaneous tissue layer image in the posterior triangle of the neck. This represents the total number of pixels within the triangular region at the back of the neck. For the first grayscale value of each pixel. This is an image of the subcutaneous tissue layer in the posterior triangle of the neck. The horizontal and vertical gradient components of the posterior neck triangle image. This represents the grayscale gradient magnitude of the posterior neck triangle image, ranging from 0 to 255. The Sobel horizontal and vertical gradient kernels are defined in the same way as S123.

[0061] S133. Regions with gradient magnitudes exceeding a preset threshold in the grayscale gradient are marked as highly sensitive areas for subtle skin color changes. The preset threshold is set to 1.5 to 3.0 times the global mean of the grayscale gradient in the subcutaneous tissue layer image of the posterior neck triangle, preferably 2.0 times. An 8-neighborhood connected component analysis is performed on the thresholded binary image to remove isolated noise regions with an area less than 50 pixels, retaining connected components with an area greater than or equal to 50 pixels as highly sensitive areas for subtle skin color changes.

[0062] Simultaneously, the mean grayscale value of the subcutaneous tissue layer image in the posterior cervical triangle is differed from the preset normal perfusion reference value. This difference is divided by the normal perfusion reference value and then multiplied by 100. After normalization, it is used as the microvascular perfusion offset index, i.e.: In the formula: The microvascular perfusion deviation index is defined as an abnormal perfusion when the absolute value is greater than 20. The grayscale mean of the subcutaneous tissue layer image in the posterior triangle of the neck; The reference value for normal perfusion is 80 to 180, with 120 being the preferred value. The circumscribed polygons are extracted from the preserved connected components to serve as the spatial boundaries of the highly sensitive regions for subtle skin color variations.

[0063] S14. Indirect localization of cervical spinous processes and collection of multimodal physical characteristic parameters of the neck, specifically including:

[0064] S141. The two-dimensional coordinates of the intersection points of fascial tension textures in the multimodal fusion feature map of the neck are used as the candidate point set for the skin projection of the cervical spinous processes. Specifically, all segments of the lateral cervical region are traversed, and all detected intersection points of fascial tension textures within each segment are extracted. Their horizontal and vertical coordinates in the feature map coordinate system are recorded to form the candidate point set. Simultaneously, the spatial location of the skin color micro-change highly sensitive area in the multimodal fusion feature map of the neck is used as the local verification region of the candidate point set. This verification region is determined by the boundary polygon of the skin color micro-change highly sensitive area marked in the aforementioned steps. Pixels within the boundary are marked as valid verification regions, and pixels outside the boundary are marked as invalid regions.

[0065] Each intersection point in the candidate point set must be located within the valid verification area or within 10 to 30 pixels (ideally 20 pixels) of the boundary of the valid verification area to proceed to the subsequent verification process. The distance calculation uses the nearest neighbor distance from the point to the polygon. In the formula: The shortest distance from the intersection point to the boundary of the highly sensitive area for subtle changes in skin color. The coordinates of the candidate intersection point, and The boundary polygon of the high-sensitivity region for subtle changes in skin color The coordinates of the two endpoints of the strip, This represents the total number of edges of the boundary polygon. This constraint utilizes the anatomical relationship between regions of abnormal microvascular perfusion and the skin projection of spinous processes to eliminate false intersections far from high-sensitivity areas and reduce noise in the candidate point set.

[0066] S142. Perform continuity verification on the candidate point set along the longitudinal direction of the cervical spine. The continuity verification requires that the difference in longitudinal coordinates between adjacent points in the candidate point set is less than a preset segment height threshold. This threshold is set to 20 to 40 pixels, preferably 30 pixels, corresponding to the projection height of a single cervical segment in the feature map. Perform principal component analysis on the candidate point set that passes the continuity verification to fit the longitudinal axis of the cervical spine. Principal component analysis uses the two-dimensional coordinates of the candidate point set as input, extracting the first principal component direction as the longitudinal axis direction of the cervical spine, and the second principal component direction as the transverse vertical direction.

[0067] The mean of the candidate point set Let be a point on the line, and the direction vector of the first principal component be... If the direction vector of the line is , then the normal vector of the line is . constant term Thus, the equation of the longitudinal axis of the cervical spine is determined. Calculate the vertical distance from each candidate point to the fitted longitudinal axis of the cervical spine. This distance is calculated using the formula for the distance from a point to a line. In the formula: For the first The vertical distance from each candidate point to the longitudinal axis of the cervical spine is expressed in pixels. For the first The two-dimensional coordinates of each candidate point in the feature map coordinate system Equation of the longitudinal axis of the cervical spine The coefficient is calculated. Outliers with a vertical distance exceeding a preset threshold are removed. The preset threshold is set to 10 to 25 pixels, preferably 15 pixels. The remaining candidate points are sorted according to their vertical coordinates to obtain the verified sequence of spike projection points.

[0068] S143. Using each point in the verified spike projection point sequence as the center, search for the geometric center of the highly sensitive area for subtle skin color changes within a preset radius. The preset search radius is set to 15 to 35 pixels, preferably 25 pixels. For all pixels in the highly sensitive area for subtle skin color changes within the search radius, calculate the arithmetic mean of their horizontal and vertical coordinates to obtain the coordinates of the geometric center of the highly sensitive area, i.e.: In the formula: The geometric center coordinates of the highly sensitive area for subtle changes in skin color. This represents the total number of pixels in the highly sensitive area of ​​subtle skin color changes within the search radius. For the first [unit] within the search radius The coordinates of a pixel in the high-sensitivity area.

[0069] The coordinates of the geometric center and the corresponding spinous process projection point are weighted and averaged. The resulting weighted average coordinates are used as the final indirect location coordinates of the cervical spinous processes in each segment. , In the formula: The coordinates of the projected spinous process points after verification. The final indirect coordinates for locating the spinous processes of each cervical vertebrae. The weight of the spinous process projection point is set to 0.4 to 0.6, with 0.5 being preferred. The final indirect positioning coordinates of all segments are arranged in longitudinal order to generate a sequence of indirect positioning coordinates for the cervical spinous processes. This sequence contains the two-dimensional coordinates of the spinous processes of segments C1 to C7 in the feature map coordinate system, with the coordinate unit being pixels.

[0070] S144. Extract the characteristic boundary of the fascial tension texture line located in the sternocleidomastoid muscle region and extending longitudinally as the sternocleidomastoid muscle tension boundary. The sternocleidomastoid muscle region is determined by the aforementioned anatomical partition template, namely, the lateral region 25mm to 40mm from the midsagittal plane of the posterior neck. Within this region, select line segments of the fascial tension texture line whose direction angle is less than 30° with the longitudinal direction. The longitudinal direction is defined as the direction whose angle with the longitudinal axis of the cervical spine is less than 15°. Piece together the continuous line segments that meet the criteria to form the complete sternocleidomastoid muscle tension boundary.

[0071] Spatial morphological parameters of the tension boundary of the sternocleidomastoid muscle are extracted, including boundary length, average curvature, maximum deviation distance, and longitudinal penetration. The boundary length is the total number of pixels in the tension boundary; the average curvature is calculated by the reciprocal of the radius of curvature at the boundary point; the maximum deviation distance is the maximum vertical distance from the boundary point to the fitted straight line; and the longitudinal penetration is the ratio of the longitudinal span of the boundary to the total longitudinal height of the lateral neck region. The endpoint curvature is approximated using a unilateral three-point difference or excluded from the average calculation. Specifically:

[0072] ,

[0073] ,

[0074] ,

[0075] ,

[0076] In the formula: This represents the total number of pixels representing the tension boundary of the sternocleidomastoid muscle. For the first The two-dimensional coordinates of the tension boundary points of the sternocleidomastoid muscle in the feature map coordinate system. , For the first The curvature at each boundary point is calculated using the central difference between that point and its two adjacent points. The mean curvature of the tension boundary of the sternocleidomastoid muscle. This represents the maximum deviation distance from the tension boundary of the sternocleidomastoid muscle, expressed in pixels. Fitting a linear equation to the tension boundary of the sternocleidomastoid muscle The coefficients are determined by fitting the boundary point set using principal component analysis. Vertical penetration, dimensionless. The maximum and minimum values ​​of the ordinate of the boundary point of the sternocleidomastoid muscle tension are given. This represents the total vertical height of the feature region on the side of the neck, expressed in pixels.

[0077] The microvascular perfusion offset index, the mean skin temperature of each anatomical region, and the spatial morphological parameters of the sternocleidomastoid muscle tension boundary were collected as multimodal physical characteristic parameters of the neck.

[0078] S2. Construct a continuous reconstruction model of cervical curvature and generate a family of viscoelastic time-varying cervical curvature curves; perform principal component analysis on the curve family to decouple the skull base rotation compensation component, generate a true cervical curvature morphological feature vector, decouple muscle imbalance pattern features, and calculate the cervical muscle biomechanical load index; specifically including:

[0079] S21. Viscoelastic time-varying cervical curvature curve family reconstruction, specifically including:

[0080] S211. The continuous reconstruction model of cervical curvature consists of three layers: a control node layer, a viscoelastic constraint layer, and a boundary constraint layer.

[0081] The control node layer receives the indirect positioning coordinate sequence of the spinous processes of the cervical vertebrae and maps the coordinates of the spinous processes of each segment from C1 to C7 in the sequence to control nodes of cubic B-spline curves. There are a total of 7 nodes, and the curve segments between adjacent nodes are described by cubic polynomial parametric equations. The node vectors are evenly distributed.

[0082] The viscoelastic constraint layer receives the microvascular perfusion offset index and the mean skin temperature of each anatomical region. After normalization, the two are weighted and fused into a local tissue stiffness regulating factor, namely: In the formula: For the first Local tissue stiffness adjustment factor at each control node The weighting factor for the average skin temperature is 0.4 to 0.6, with 0.5 being preferred. The weight of the microvascular perfusion offset index is set at 0.4 to 0.6, preferably 0.5. For the normalized first Mean skin temperature of each anatomical region The normalized microvascular perfusion deviation index is obtained by using min-max linear normalization, i.e. , ,in These are the minimum and maximum mean skin temperatures of each anatomical region in the training set or the current sequence. The minimum and maximum values ​​of the microvascular perfusion deviation index are given. The local tissue stiffness adjustment factor is also considered. Converted to equivalent bending stiffness coefficient via linear mapping , This refers to the stiffness coefficient, which ranges from 0.5 to 2.0. When the value is greater than 1.0, the curve tends to be rigid, and when it is less than 1.0, the curve tends to be flexible.

[0083] The boundary constraint layer receives the tension boundary of the sternocleidomastoid muscle, uses the longitudinal penetration of its spatial morphological parameters as the effective longitudinal span constraint of the curve, and uses the lateral position of the boundary as a hard boundary restriction for the lateral offset of the curve. Curve segments that exceed the boundary are forcibly retracted to the inside of the boundary.

[0084] After the three-layer structure is cascaded, the output of the control node layer is used as the initial value. The stiffness of the viscoelastic constraint layer is modulated and the amplitude is limited by the boundary constraint layer to construct a continuous reconstruction model of cervical curvature based on the time-varying characteristics of cervical viscoelasticity.

[0085] S212. The dynamic head, neck and shoulder image sequence is processed frame by frame using the cervical curvature continuous reconstruction model. For each frame, the indirect positioning coordinate sequence of the cervical spinous process corresponding to that frame is extracted as the control node coordinates at the current moment. The coordinates of each node are input into the control node layer. At the same time, the average skin temperature of each anatomical region corresponding to that frame and the microvascular perfusion offset index are input into the viscoelastic constraint layer. The tension boundary of the sternocleidomastoid muscle in that frame is input into the boundary constraint layer.

[0086] The model uses cubic B-spline basis functions to interpolate and fit seven control nodes, generating a smooth cervical curvature curve that passes through the first and last nodes and approximates the middle nodes. The curve sampling point density is set to 10 to 30 sampling points per segment, preferably 20. The parametric equation of the cubic B-spline curve is as follows: In the formula: For the first The coordinates of each control node are uniformly distributed. , For curve parameter variables, , This is the parametric equation for the cervical curvature curve reconstructed from a single frame image using a continuous cervical curvature reconstruction model. The basis functions are cubic B-spline functions, determined by the Cox-de Boor recursive formula.

[0087] The above reconstruction operation is performed on each frame of the dynamic head, neck, and shoulder image sequence. Each frame generates an independent cervical curvature curve. The cervical curvature curves of all frames are arranged in chronological order to form a family of viscoelastic time-varying cervical curvature curves. The total number of frames in the curve family is the same as the total number of frames in the dynamic head, neck, and shoulder image sequence. The frame rate is set to 25Hz to 30Hz, and the time span covers the complete duration of a single dynamic acquisition.

[0088] S213. The time-varying properties of viscoelasticity are characterized by introducing time-dependent stress relaxation and creep functions. The stress relaxation function adopts an exponential decay form: In the formula: The initial elastic modulus ranges from 1.0 MPa to 5.0 MPa, with 2.5 MPa being preferred. This is the stress relaxation time constant, whose value is based on the rate of change of the position of the cervical spinous process indirect positioning coordinate sequence between adjacent frames. Dynamic adjustment, when Pixels / Frames The time interval is selected from 0.5 s to 1.0 s, preferably 0.8 s. Pixels / Frames The time interval is selected from 1.5s to 3.0s, with 2.0s being preferred.

[0089] The creep function adopts the Kelvin-Voigt model form: In the formula: The Kelvin spring stiffness ranges from 0.5 MPa to 2.0 MPa. The creep time constant is the first frame of the sequence. The initial value is uniformly set to 7.0 s; starting from the second frame, it is dynamically adjusted according to the sign and magnitude of the difference between the local tissue stiffness adjustment factor of the current frame and the previous frame. When the difference is greater than 0.1 and the current frame... If the value is greater than the previous frame, it is considered an upward trend. Use a time interval of 2.0 s to 4.0 s, otherwise use a time interval of 5.0 s to 10.0 s. The above time-varying parameters are updated frame by frame to ensure that the family of viscoelastic time-varying cervical curvature curves accurately reflects the dynamic response of the neck.

[0090] The stress relaxation function and creep function mentioned above are coupled by adjusting the equivalent bending stiffness weight of the control nodes of the cubic B-spline curve: the equivalent bending stiffness coefficient obtained from S211 is... As a control node Substituting the weight coefficients into the continuity constraint equations for the first and second derivatives of the B-spline curve, we correct the basis function weight allocation so that the curve... As the radius of curvature increases, it tends to become rigid. As the radius of curvature decreases, it tends to become softer.

[0091] S22. Decoupling of real curvature morphology and muscle imbalance mode, and calculation of mechanical load, specifically including:

[0092] S221. Principal component analysis is performed on the family of viscoelastic time-varying cervical curvature curves. First, the cervical curvature curves of each frame in the curve family are discretized and sampled into a sequence of equally spaced points. The number of sampling points is set to 50 to 100, preferably 70. Each sampling point contains two-dimensional coordinates. The two-dimensional coordinates of each sampling point of a single frame curve are then sorted according to... Sequential concatenation flattens into a one-dimensional vector, dimension ,in The number of sampling points per frame. The flattened vectors of all frames are stacked row-wise to form a data matrix. The number of rows in the matrix represents the total number of frames in the dynamic head, neck, and shoulder image sequence, and the number of columns represents the dimension of the flattened curve of a single frame. The data matrix is ​​then zero-mean normalized, and the covariance matrix is ​​calculated, i.e.: In the formula: Let covariance matrix be the variance matrix. The data matrix after zero-mean normalization has 100 rows. The number of columns is , The dimension of the flattened cervical curvature curve in a single frame is the one-dimensional vector dimension. This represents the total number of frames in the dynamic head, neck, and shoulder image sequence. The covariance matrix is ​​then decomposed into eigenvalues, i.e.: In the formula: For the first 1 eigenvalue, For the first The eigenvector, i.e. the eigenvector of the ? One principal component direction; extract the top 3 to 5 principal component directions arranged from largest to smallest feature value, preferably 3.

[0093] Analyzing the characteristics of the eigenvectors corresponding to each principal component direction in terms of spatial frequency and amplitude distribution, principal component directions that simultaneously meet the following two conditions are identified as skull base rotation compensation components: Condition 1, the eigenvalue corresponding to this principal component is greater than 2.0 times the mean of all eigenvalues; Condition 2, the proportion of components with a spatial frequency below 0.5Hz in this principal component direction exceeds 60%, and this component usually corresponds to the first principal component. This component is removed from the data matrix projection, and the remaining principal component direction components are retained as the true cervical curvature change components. The projection coefficients of the cervical curvature curves of each frame onto the true cervical curvature change components are calculated, i.e.: In the formula: For the first The true cervical curvature morphological feature vector of the frame. It is a projection matrix composed of the eigenvectors corresponding to the retained true cervical curvature change components. For the first The flattened curve vector after zero-mean transformation of the frame; generate the real cervical curvature morphological feature vector corresponding to each frame, with the vector dimension consistent with the number of retained principal components.

[0094] S222. Using the actual cervical curvature change component, the mean skin temperature of each anatomical region, the microvascular perfusion offset index, and four spatial morphological parameters—boundary length, mean curvature, maximum deviation distance, and longitudinal penetration—of the sternocleidomastoid muscle tension boundary as input features, a muscle imbalance pattern decoupling network is constructed. This network adopts a multi-task learning architecture, consisting of a shared feature encoding layer and three parallel binary classification branches:

[0095] The shared feature encoding layer consists of two cascaded fully connected layers. The first layer has 128 neurons with the ReLU activation function, and the second layer has 64 neurons with the ReLU activation function. It is used to map the input features to the shared latent space.

[0096] The three parallel binary classification branches correspond to the sternocleidomastoid muscle shortening pattern discrimination, the deep neck flexor muscle endurance failure pattern discrimination, and the upper trapezius muscle fascial adhesion pattern discrimination, respectively. Each branch consists of a fully connected layer and a sigmoid activation function. The number of neurons in the fully connected layer is 16, and the output is a single confidence score in the range of 0 to 1.

[0097] The sternocleidomastoid muscle shortening status label, deep neck flexor muscle endurance status label, and upper trapezius muscle fascia status label were used as supervision signals. These labels were assigned by clinicians using a standardized postural assessment protocol, with values ​​of 0 or 1. During training, a binary cross-entropy loss function was used, with the three branch losses summed in a 1:1:1 ratio with equal weights. Parameters were updated using the Adam optimizer, with a learning rate set to 0.0001 to 0.001 (preferably 0.0005), a training iteration count set to 100 to 300 (preferably 200), and a batch size set to 16 to 64 (preferably 32).

[0098] S223. Input the input features into the trained muscle imbalance pattern decoupled network. The shared feature encoding layer performs a nonlinear transformation on the input features, extracts cross-pattern shared representations, and outputs a 64-dimensional shared latent vector. Three parallel binary classification branches receive this shared latent vector: the sternocleidomastoid shortening pattern discrimination branch outputs the confidence score of the sternocleidomastoid shortening pattern, the deep neck flexor endurance failure pattern discrimination branch outputs the confidence score of the deep neck flexor endurance failure pattern, and the upper trapezius fascia adhesion pattern discrimination branch outputs the confidence score of the upper trapezius fascia adhesion pattern. The confidence score of each branch ranges from 0 to 1, with a score closer to 1 indicating a higher probability of the existence of the muscle imbalance pattern. The discrimination results of the three branches are combined with their respective confidence scores to form the muscle imbalance pattern features. When the confidence score of a branch is greater than or equal to 0.5, the discrimination result of that pattern is recorded as 1, indicating the existence of the pattern; when the confidence score is less than 0.5, the discrimination result is recorded as 0, indicating the non-existence of the pattern. This muscle imbalance pattern characteristic also serves as the weighting basis for calculating the cervical muscle mechanical load index in S224.

[0099] like Figure 2 As shown, the muscle imbalance pattern decoupling network includes:

[0100] The input layer is used to receive the actual cervical curvature change component, the average skin temperature of each anatomical region, the microvascular perfusion offset index, and the spatial morphological parameters of the sternocleidomastoid muscle tension boundary.

[0101] Fully connected layer 1 is used to perform preliminary nonlinear mapping and dimensional transformation on the input features;

[0102] Fully connected layer 2 is used to extract deep shared feature representations step by step;

[0103] Branch 1 is used to identify the shortening pattern of the sternocleidomastoid muscle and output the discrimination result and confidence score;

[0104] Branch 2 is used to identify deep neck flexor muscle endurance failure patterns and output the identification results and confidence scores;

[0105] Branch 3 is used to identify the adhesion pattern of the upper trapezius fascia and output the identification result and confidence score;

[0106] The output node is used to aggregate the discrimination results and confidence scores of each branch respectively;

[0107] The output layer is used to aggregate the results from the three output nodes and generate muscle imbalance pattern features.

[0108] S224. Calculate the cervical muscle biomechanical load index. First, calculate the deviation between the mean skin temperature of each anatomical region and the normal reference value. The deviation is calculated by subtracting the normal reference value from the mean skin temperature, taking the absolute value, and then dividing by the normal reference value. The result is expressed as a percentage, i.e.: In the formula: For the first Skin temperature deviation in each anatomical region For the first Mean skin temperature of each anatomical region, in °C. The reference value for normal perfusion is 80 to 180, with 120 being preferred. The deviation of each anatomical region is weighted and summed with the microvascular perfusion deviation index according to preset weights, i.e.: In the formula: For the first The basic mechanical load index of each anatomical region. The weight for skin temperature deviation is set at 0.3 to 0.7, with 0.5 being preferred. The weight of the microvascular perfusion offset index is 0.3 to 0.7, preferably 0.5.

[0109] Then, a weighted adjustment is made based on the anatomical correspondence between each anatomical region and the muscle imbalance pattern. The posterior border of the sternocleidomastoid muscle corresponds to the sternocleidomastoid shortening pattern, the anterior border junction of the trapezius muscle corresponds to the upper trapezius fascial adhesion pattern, and the midline cervical region corresponds to the deep cervical flexor muscle endurance failure pattern. When the confidence score of a certain region's corresponding pattern is greater than 0.5, the basic mechanical load index of that region is multiplied by an amplification factor, i.e.: In the formula: For the first The mechanical load index adjusted for each anatomical region. This is the magnification factor, ranging from 1.2 to 1.5, preferably 1.3. For the first The confidence scores for the muscle imbalance patterns corresponding to each anatomical region are used. The cervical muscle mechanical load index is obtained by arithmetically averaging the adjusted mechanical load indices of all anatomical regions. In the formula: The cervical muscle mechanical load index ranges from 0 to 100. An index greater than 30 is considered an abnormal cervical muscle mechanical load. This represents the total number of anatomical regions.

[0110] S3. Extract time-frequency features of microtremors to differentiate between physiological and pathological tremors caused by abnormal muscle tone, and remove artifacts through temporal coherence verification; input the verified true cervical curvature morphology feature vector, muscle imbalance pattern features, and cervical muscle biomechanical load index into the upper crossed syndrome dynamic grading network, and output the dynamic risk level of upper crossed syndrome and muscle imbalance subtype; specifically including:

[0111] S31. Extraction of time-frequency features of microtremors and discrimination of physiological and pathological tremors, specifically including:

[0112] S311. Optical flow field calculation is performed on the posterior neck triangle region, corresponding to the highly sensitive area of ​​subtle skin color changes, in each frame of the dynamic head, neck, and shoulder image sequence. The Farneback dense optical flow algorithm is used, taking two adjacent frames of the posterior neck triangle region as input, to calculate the displacement vector of each pixel in the horizontal and vertical directions. The algorithm parameters are set as follows: neighborhood window size is 15 pixels × 15 pixels to 25 pixels × 25 pixels, preferably 21 pixels × 21 pixels; the number of pyramid layers is 3 to 5 layers, preferably 4 layers; the number of iterations per layer is 5 to 15 times, preferably 10 times; and the multi-scale scaling factor is 0.5 to 0.8, preferably 0.5.

[0113] From the calculated optical flow vector field, only the pixel displacement within the highly sensitive area of ​​subtle skin color changes is retained. The displacement of all pixels in this region over time is extracted to obtain the micro-displacement signal of the posterior neck triangle. In the formula: The total number of pixels within the highly sensitive area for subtle changes in skin tone. For the first Each pixel at time Horizontal optical flow displacement component, For the first Each pixel at time The vertical optical flow displacement component, For a moment The micro-displacement signal is measured in pixels, and the sampling frequency is consistent with the image acquisition frame rate, ranging from 25Hz to 30Hz.

[0114] S312. The micro-displacement signal is subjected to bandpass filtering. A fourth-order Butterworth bandpass filter is used, with the lower limit of the passband frequency set to 0.5Hz to 2Hz, preferably 1Hz, to filter out the slow drift component caused by breathing and slow body position adjustments; the upper limit of the passband frequency is set to 15Hz to 25Hz, preferably 20Hz, to filter out high-frequency electronic noise and environmental vibration noise. Bandpass filtering is performed on the micro-displacement signal to remove the slow drift component and high-frequency noise component, resulting in the filtered micro-displacement signal.

[0115] A Fast Fourier Transform (FFT) is performed on the filtered micro-displacement signal. The number of transform points is set between 1024 and 4096, preferably 2048. If the sequence length is insufficient, zeros are padded to the end to reach the set length. The power values ​​of each frequency component are calculated to obtain the power spectral density distribution of the micro-displacement signal, i.e.: In the formula: For the first The power spectral density values ​​for each frequency component, in pixels squared per Hz. The number of points for the Fast Fourier Transform is between 1024 and 4096, with 2048 being the preferred value. For the filtered micro-displacement signal at time... The sampled values; the frequency resolution is the sampling frequency divided by the number of transform points, i.e.: In the formula: This is the frequency resolution, measured in Hz. This is the sampling frequency, measured in Hz, which is consistent with the image acquisition frame rate.

[0116] S313. Extract the first and second power integral values ​​of the power spectral density distribution within a preset frequency band. Perform numerical integration on the power spectral density values ​​within the two preset frequency bands using the trapezoidal rule to obtain the first and second power integral values, expressed in pixels squared. , In the formula: The first power integral value corresponds to the preset frequency band. The frequency range is 2Hz to 8Hz, corresponding to the dominant frequency band of physiological microtremors. This is the second power integral value, corresponding to a preset frequency band. The frequency range is 8Hz to 15Hz, corresponding to the dominant frequency band of pathological dystonia tremor.

[0117] The ratio of the first power integral value to the second power integral value is calculated as a frequency domain feature of the microflutter, i.e.: In the formula: This represents the frequency domain characteristics of micro-flutter. Second power integral value. Less than the preset minimum value ε=10 -6 At that time, the frequency domain characteristics of micro-tremors were directly applied. The threshold is set to 100 and is defined as physiological microtremor. When the frequency domain characteristics of microtremor are lower than the preset threshold, it is defined as pathological dystonia. The preset threshold is set to 1.5 to 3.0, preferably 2.0. When the frequency domain characteristics of microtremor are higher than or equal to the preset threshold, it is defined as physiological microtremor.

[0118] S32. Temporal coherence verification and artifact removal, specifically including:

[0119] S321. Obtain the true cervical curvature morphological feature vector and microtremor frequency domain features corresponding to each frame: The true cervical curvature morphological feature vector originates from S221, with dimensions consistent with the number of retained principal components, ranging from 3 to 5 dimensions; the microtremor frequency domain features originate from S313, and are single-dimensional scalars. Arrange the true cervical curvature morphological feature vectors in the order of acquisition time along the frame number to form a curvature feature time sequence, with the sequence length equal to the total number of frames in the dynamic head, neck, and shoulder image sequence. Arrange the microtremor frequency domain features in the same time order to form a tremor feature time sequence. The sampling interval for both sequences is the reciprocal of the image acquisition frame rate, i.e., 0.033 seconds to 0.04 seconds.

[0120] To eliminate dimensional differences, each component of the curvature feature time series is standardized to have a mean of 0 and a standard deviation of 1, i.e.: In the formula: For the standardized first Frame number The true cervical curvature morphological feature vector components. For the original number Frame number The true cervical curvature morphological feature vector components. For the first The mean of the dimensional components over time. For the first The standard deviation of the dimensional components over time. The time series sequences of tremor characteristics are also standardized to ensure that the two sequences are on the same numerical scale, i.e.: In the formula: For the standardized first The frequency domain characteristics of frame micro-flutter. For the original number The frequency domain characteristics of frame micro-flutter. This represents the temporal mean of the frequency domain characteristics of microtremors. This represents the standard deviation of the frequency domain characteristics of microtremors over time.

[0121] S322. Calculate the cross-correlation function of the tremor feature time series and the curvature feature time series in the time dimension. Since the curvature feature time series is a multi-dimensional vector and the tremor feature time series is a single-dimensional scalar, a dimension-by-dimensional cross-correlation calculation method is adopted. For each component of the curvature feature time series, calculate the cross-correlation function with the tremor feature time series, with the delay range set to -10 frames to +10 frames. The cross-correlation function at each delay... The value at that location is ,in Extracting the coherence coefficient at zero delay is... , which serves as the coherence coefficient for this dimension component, i.e.: In the formula: For the first The correlation coefficient between the dimensional components and the frequency domain characteristics of microtremors at zero delay. This represents the total number of frames in the dynamic head, neck, and shoulder image sequence.

[0122] The final coherence coefficient is obtained by taking the arithmetic mean of the coherence coefficients of all dimensional components, i.e.: In the formula: The dimension of the feature vector representing the true cervical curvature morphology is the number of principal components retained. The final coherence coefficient ranges from -1 to +1. The closer the absolute value is to 1, the stronger the temporal synchronization between the two sequences; the closer it is to 0, the weaker the synchronization.

[0123] S323. Compare the coherence coefficient with a preset threshold, which is set to 0.3 to 0.6, preferably 0.45. When the coherence coefficient is lower than the preset threshold, artifacts are determined to exist within that time period. Artifact types include craniocervical angle decoupling residuals and respiratory / swallowing artifacts. The time period where artifacts are determined to exist is marked as an abnormal interval, and the interval boundary is extended forward and backward by 2 to 5 frames, preferably 3 frames, as an interpolation correction window. If adjacent abnormal intervals overlap after expansion, the overlapping intervals are merged into a single continuous interpolation correction window.

[0124] If the abnormal interval is located at the beginning of the sequence, quadratic polynomial extrapolation interpolation is used, with the normal feature vector of the 5th frame and the normal feature vector at the boundary of the abnormal interval as nodes. If it is located at the end of the sequence, extrapolation interpolation is used similarly, with the normal feature vector at the boundary of the abnormal interval and the normal feature vector of the 5th frame from the end as nodes. If the abnormal interval is located in the middle of the sequence, standard cubic spline interpolation is used, with the normal feature vectors at both boundaries as nodes. Cubic spline interpolation correction is applied to the true cervical curvature morphology feature vector within the window. A cubic spline curve is constructed using the normal feature vector at the window boundary as nodes. The feature vectors of each frame within the window are recalculated, replacing artifact-contaminated outliers. The cubic spline interpolation basis function is... ,in The coefficients represent the sampling times corresponding to normal frames at the window boundaries. The cervical spine curvature morphology feature vector is determined by simultaneously solving the eigenvector values ​​at the window boundary and the continuity condition of the first derivative. After interpolation correction, a verified time-series sequence of the cervical spine curvature morphology feature vector is generated.

[0125] S33. Dynamic hierarchical classification and result output of multi-source fusion features, specifically including:

[0126] S331. The true cervical curvature morphological feature vector, muscle imbalance pattern features, and cervical muscle biomechanical load index, verified by temporal coherence, are concatenated at the feature level. The true cervical curvature morphological feature vector has 3 to 5 dimensions, determined by the number of principal components retained in S221. The muscle imbalance pattern features include the discrimination results and confidence scores for the sternocleidomastoid shortening pattern, the deep cervical flexor endurance failure pattern, and the upper trapezius fascia adhesion pattern, totaling 6 dimensions. The cervical muscle biomechanical load index is a 1-dimensional scalar.

[0127] Before splicing, all the above feature components are standardized to have a mean of 0 and a standard deviation of 1, i.e.: In the formula: For the standardized first Each feature component For the original number Each feature component For the first The mean of each feature component on the training set or the current sequence. For the first The standard deviation of each feature component on the training set or the current sequence. The standardized feature components are arranged in a fixed order, i.e., the true cervical curvature morphology feature vector first, the muscle imbalance pattern feature in the middle, and the cervical muscle mechanical load index last, to form a multi-source fusion feature vector with a total dimension of 10 to 12.

[0128] S332. The multi-source fusion feature vector is input into the dynamic hierarchical network for crossed syndrome. This network consists of three parts: a shared bottom-level feature extraction layer, a dynamic risk level classification branch, and a muscle imbalance subtype classification branch. The shared bottom-level feature extraction layer contains two cascaded fully connected layers. The first fully connected layer has 128 neurons with ReLU activation, and the second fully connected layer has 64 neurons with ReLU activation. This layer maps the multi-source fusion feature vector to the shared deep feature space. To suppress overfitting, a Dropout layer is inserted between the two fully connected layers, with a dropout rate set to 0.3 to 0.5, preferably 0.4. The shared deep feature dimension is 64-dimensional and serves as the common input for both classification branches.

[0129] S333. Input the shared deep features into the dynamic risk level classification branch and the muscle imbalance subtype classification branch respectively: The dynamic risk level classification branch contains a fully connected layer and a softmax activation function. The fully connected layer has 4 neurons, corresponding to the four dynamic risk levels of upper crossed syndrome: low risk, low-to-medium risk, medium-to-high risk, and high risk. The softmax function outputs the probability distribution of each level, i.e.: In the formula: The first in the dynamic risk level classification branch of upper crossed syndrome The output probability of each risk level For dynamic risk level classification branch fully connected layer The output value of each neuron This represents the total number of dynamic risk levels. The level corresponding to the highest probability is taken as the final dynamic risk level of upper crossed syndrome.

[0130] The muscle imbalance subtype classification branch contains a fully connected layer and a Softmax activation function. The fully connected layer has 8 neurons, corresponding to eight muscle imbalance subtypes: normal, isolated sternocleidomastoid shortening, isolated deep neck flexor endurance failure, isolated upper trapezius fascial adhesion, mixed type A, mixed type B, mixed type C, and undetermined type. The Softmax function outputs the probability distribution of each subtype, i.e.: In the formula: The first subtype in the classification of muscle imbalance The output probability of each subtype The fully connected layer of the subtype classification branch for muscle imbalance The output value of each neuron The total number of muscle imbalance subtypes, The subtype corresponding to the highest probability is taken as the final muscle imbalance subtype.

[0131] The combined mapping relationship between the eight subtypes and the three binary classification results is as follows: 000 corresponds to the normal type, 001 corresponds to the simple upper trapezius fascia adhesion type, 010 corresponds to the simple deep neck flexor muscle endurance failure type, 011 corresponds to the mixed type A, 100 corresponds to the simple sternocleidomastoid muscle shortening type, 101 corresponds to the mixed type B, 110 corresponds to the mixed type C, and 111 corresponds to the undetermined type.

[0132] Example 2

[0133] Please see the appendix Figure 2 A cervical curvature recognition and detection system for upper crossed syndrome based on an improved HRNet network, used to implement a method for recognizing and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network, comprising:

[0134] The multimodal image acquisition module is used to acquire dynamic head, neck and shoulder image sequences of the object to be detected. The image sequences include infrared thermal imaging image sequences and polarized light image sequences acquired synchronously under the same field of view.

[0135] An improved HRNet backbone network module is connected to a multimodal image acquisition module. This module processes each frame of infrared thermal imaging and polarized light images from a dynamic head, neck, and shoulder image sequence through an infrared input branch and a polarization input branch, respectively. The resulting multimodal feature fusion attention module generates a neck multimodal fusion feature map.

[0136] The bony landmark indirect localization module, connected to the improved HRNet backbone network module, is used to extract the cervical lateral fascial tension texture lines from the cervical multimodal fusion feature map and determine the intersection point of the fascial tension texture, extract the microvascular texture of the posterior cervical triangle from the polarized light image sequence and determine the high-sensitivity area of ​​skin color micro-change, and generate cervical multimodal physical feature parameters, cervical spinous process indirect localization coordinate sequence and sternocleidomastoid muscle tension boundary.

[0137] The continuous reconstruction and decoupling module for cervical curvature is connected to the indirect positioning module of bony landmarks. It is used to construct a continuous reconstruction model of cervical curvature based on the viscoelastic time-varying characteristics of the neck, using the indirect positioning coordinate sequence of the cervical spinous processes as control nodes, the multimodal physical characteristic parameters of the neck as local tissue stiffness adjustment factors, and the tension boundary of the sternocleidomastoid muscle as boundary constraint. It generates a family of viscoelastic time-varying cervical curvature curves, decouples the skull base rotation compensation component from the actual cervical curvature change component, generates the actual cervical curvature morphological feature vector, decouples the muscle imbalance pattern features, and calculates the mechanical load index of the cervical muscles.

[0138] The microtremor frequency domain verification module is connected to the cervical curvature continuous reconstruction and decoupling module. It is used to extract the time-frequency features of microtremors from dynamic head, neck and shoulder image sequences to distinguish between physiological and pathological abnormal muscle tone tremors. After time-series coherence verification and artifact removal, it generates the verified true cervical curvature morphological feature vector.

[0139] The dynamic grading output module, connected to the microtremor frequency domain verification module and the cervical curvature continuous reconstruction and decoupling module, is used to input the verified real cervical curvature morphological feature vector, muscle imbalance pattern features and cervical muscle mechanical load index into the upper crossed syndrome dynamic grading network, and output the dynamic risk level of upper crossed syndrome and muscle imbalance subtype.

[0140] Example 3

[0141] To verify the technical effectiveness of the cervical curvature recognition and detection method for upper crossed syndrome based on the improved HRNet network described in this invention, the following experimental scheme was adopted for verification:

[0142] I. Experimental Subjects and Data Collection

[0143] One hundred and twenty patients with upper crossed syndrome and sixty healthy controls, aged 18 to 55 years, were recruited. A multimodal imaging module was used to simultaneously acquire infrared thermal imaging and polarized light image sequences at a frame rate of 30 frames per second for 10 seconds. The cervical Cobb angle was measured using lateral X-ray radiographs as the gold standard, and two senior massage therapists determined muscle imbalance pattern labels through clinical musculoskeletal ultrasound and manual palpation.

[0144] II. Comparison Method Settings

[0145] Existing technology group: The standard HRNet network is used, the input is a single frame RGB color image, key points on the surface of soft tissues such as nose, ears and shoulders are extracted, the cervical curvature is approximated by the head tilt angle, and a single frame static classifier is used for risk assessment.

[0146] This invention group adopts the complete technical solution described in claims 1 to 10 of this application, including improved HRNet multimodal fusion, indirect localization of bony landmarks, viscoelastic time-varying curvature reconstruction, muscle imbalance mode decoupling, microtremor frequency domain verification, and dynamic grading network.

[0147] III. Evaluation Indicators

[0148] Spinous process localization accuracy: the percentage of samples where the deviation between the indirect localization coordinates and the X-ray spinous process projection position is less than 5mm;

[0149] Curvature reconstruction accuracy rate: The percentage of samples where the reconstructed curvature deviates from the X-ray Cobb angle by less than 3°;

[0150] Correct decoupling rate of skull base rotation compensation: the percentage of samples that correctly distinguish between skull base rotation compensation and true curvature changes;

[0151] Accuracy of muscle imbalance pattern discrimination: Percentage of samples that correctly identified the three muscle imbalance patterns;

[0152] Risk stratification accuracy: The percentage of samples whose dynamic risk level is consistent with the comprehensive assessment of clinicians;

[0153] Dynamic artifact removal success rate: the percentage of frames that correctly identify and remove breathing and swallowing artifacts;

[0154] Overall clinical testing effectiveness rate: Based on the above indicators, the output results can directly guide the sample proportion of clinical massage intervention.

[0155] IV. Experimental Results

[0156] Independent samples t-test and chi-square test showed significant differences between the two groups in all indicators (P<0.01). Specific data are shown in Table 1 below. Figure 4 As shown:

[0157] Table 1: Comparison of Results

[0158]

[0159] V. Results Analysis

[0160] Experimental results show that this invention, through the fusion of infrared thermography and polarized light multimodal imaging, indirectly locates the cervical spinous processes using fascial tension texture and microvascular texture, improving the accuracy of spinous process localization by 47 percentage points compared to existing technologies. Based on the viscoelastic time-varying characteristics, decoupling skull base rotation compensation improves the curvature reconstruction accuracy by 37 percentage points, effectively avoiding false positives caused by the head tilt angle approximation method. By decoupling the network through muscle imbalance patterns, the discrimination accuracy of three muscle imbalance patterns is improved by 31 percentage points. Combined with microtremor frequency domain analysis and coherence verification, the success rate of dynamic artifact removal is improved by 24 percentage points, significantly enhancing the stability of dynamic detection. The overall clinical detection effectiveness rate has increased from 55% to 88%, demonstrating that this invention can meet the actual clinical needs of cervical curvature identification and detection in upper crossed syndrome.

[0161] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0162] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this specification. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network, characterized in that, include: The dynamic head, neck and shoulder image sequence of the object to be detected is obtained, and a multimodal fusion feature map of the neck is generated by the improved HRNet backbone network. The fascial tension texture line of the neck side is extracted and the intersection point of the fascial tension texture is determined. Extract the microvascular texture of the posterior cervical triangle and identify the high-sensitivity area for subtle changes in skin color, and generate multimodal physical feature parameters of the neck, indirect positioning coordinate sequence of cervical spinous processes, and tension boundary of the sternocleidomastoid muscle; A continuous reconstruction model of cervical curvature was constructed to generate a family of viscoelastic time-varying cervical curvature curves. Principal component analysis was performed on the family of curves to decouple the skull base rotation compensation component, generate the true cervical curvature morphological feature vector, decouple the muscle imbalance pattern features, and calculate the cervical muscle mechanical load index. Extracting the time-frequency features of microtremors to distinguish between physiological and pathological dystonia, and eliminating artifacts through temporal coherence verification; The verified real cervical curvature morphological feature vector, muscle imbalance pattern features, and cervical muscle mechanical load index are input into the upper crossed syndrome dynamic classification network to output the dynamic risk level of upper crossed syndrome and muscle imbalance subtype.

2. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network according to claim 1, characterized in that, The process involves acquiring dynamic head, neck, and shoulder image sequences of the target object, and generating a multimodal fusion feature map of the neck using an improved HRNet backbone network. Specifically, this includes: The infrared thermal imaging unit and the polarization imaging unit are set on the back side of the target object's neck. Under the condition of bare neck and head in a neutral position, infrared thermal imaging image sequence and polarization light image sequence are acquired simultaneously to form a dynamic head, neck and shoulder image sequence. Each frame of infrared thermal imaging image and polarized light image is input into the infrared input branch and polarized input branch of the improved HRNet backbone network, respectively. After processing by the initial convolutional layer and parallel multi-resolution subnet, infrared thermal imaging image feature map and polarized light image feature map are generated at each resolution scale. The multimodal feature fusion attention module concatenates the feature maps of infrared thermal imaging and polarized light images at the same resolution scale along the channel dimension and then applies attention weighting. The weighted fusion feature maps at each scale are then upsampled to a uniform resolution and cascaded to generate a multimodal fusion feature map of the neck. The infrared thermal imaging image sequence is divided into partitions according to a preset anatomical partition template. The average temperature of pixels within each partition is calculated to generate the average skin temperature of each anatomical partition. In the formula: For the first The total number of valid pixels in each partition For the first Within the first partition Temperature value of each pixel.

3. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network according to claim 2, characterized in that, Extracting the fascial tension texture lines on the cervical side and determining the intersection points of the fascial tension textures, specifically including: The multimodal fusion feature map of the neck is input into the neck fascial tension texture perception branch, and the feature region of the lateral neck area is segmented and extracted according to the anatomical position of the posterior border of the sternocleidomastoid muscle and the anterior border of the trapezius muscle. A multi-scale Gabor filter bank was used to calculate the filtering response of the feature region of the neck side area, and the maximum response direction of each pixel at each directional scale was extracted as the local texture direction angle. The lateral cervical region is divided into multiple lateral cervical segments according to the longitudinal segments of the cervical spine. The highest frequency value of the local texture direction angle in each segment is counted as the principal texture direction angle. Spatial difference operation is performed on the texture direction angles of each segment to generate a texture direction gradient field. Specifically: , In the formula: This is the segmental texture direction angle matrix for the lateral neck region. For Sobel horizontal gradient kernel, For Sobel vertical gradient kernel, These are the gradient components in the horizontal and vertical directions. The gradient magnitude is the gradient field along the texture direction. This is a two-dimensional convolution operation; The continuous boundary where the gradient magnitude in the texture direction gradient field exceeds a preset threshold is taken as the fascial tension texture line, and the intersection point of multiple fascial tension texture lines in space is taken as the fascial tension texture intersection point.

4. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network according to claim 3, characterized in that, Extracting the microvascular texture of the posterior triangular area of ​​the neck and identifying highly sensitive areas for subtle changes in skin color, specifically including: The polarized light image sequence is input into the microvascular texture sensing branch of the posterior triangular neck, and the subcutaneous tissue layer signal is extracted from the polarized light image sequence to obtain the subcutaneous tissue layer image; The subcutaneous tissue layer image was segmented according to the anatomical boundaries of the posterior cervical triangle, and the spatial distribution features of pixel gray values ​​were extracted, including the gray mean and gray space gradient. Specifically: , , In the formula: The grayscale mean of the subcutaneous tissue layer image in the posterior triangle of the neck. This represents the total number of pixels within the triangular region at the back of the neck. For the first grayscale value of each pixel. This is an image of the subcutaneous tissue layer in the posterior triangle of the neck. The horizontal and vertical gradient components of the posterior neck triangle image. The grayscale gradient magnitude of the posterior triangle region of the neck image; Regions in the grayscale gradient with a rate of change exceeding a preset threshold are marked as high-sensitivity areas for subtle skin color changes. The difference between the mean grayscale value and the preset normal perfusion reference value is normalized and used as the microvascular perfusion offset index. In the formula: The microvascular perfusion deviation index. The grayscale mean of the subcutaneous tissue layer image in the posterior triangle of the neck. This is a reference value for normal infusion.

5. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network according to claim 4, characterized in that, Generate multimodal physical feature parameters of the neck, indirect positioning coordinate sequences of cervical spinous processes, and tension boundaries of the sternocleidomastoid muscle, specifically including: The two-dimensional coordinates of the intersection of fascial tension textures are used as the candidate point set for the skin projection of the cervical spinous process, and the location of the highly sensitive area of ​​skin color micro-change is used as the local verification area of ​​the candidate point set. The continuity of the candidate point set along the longitudinal direction of the cervical spine is verified. The longitudinal axis of the cervical spine is fitted by principal component analysis. The vertical distance between each candidate point and the longitudinal axis of the cervical spine is calculated. Outliers exceeding the preset threshold are removed to obtain the verified spinous process projection point sequence. Using each point in the validated spinous process projection point sequence as the center, the geometric center of the high-sensitivity zone for subtle skin color changes is searched within a preset radius. The weighted average of the geometric center and the corresponding spinous process projection point is used as the indirect positioning coordinates of the cervical spinous processes of each segment, i.e.: , In the formula: The coordinates of the projected spinous process points after verification. The final indirect coordinates for locating the spinous processes of each cervical vertebrae. Weight the projection points of the spinous processes; generate the indirect positioning coordinate sequence of the cervical spinous processes; The characteristic boundary of the fascial tension texture line located in the sternocleidomastoid muscle region and extending longitudinally was extracted as the tension boundary of the sternocleidomastoid muscle, and its spatial morphological parameters were extracted. The microvascular perfusion offset index, the mean skin temperature of each anatomical region, and the spatial morphological parameters were collected into the multimodal physical characteristic parameters of the neck.

6. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network according to claim 5, characterized in that, A continuous reconstruction model of cervical curvature was constructed, generating a family of viscoelastic time-varying cervical curvature curves, specifically including: Using the indirect positioning coordinate sequence of the cervical spinous processes as curve control nodes, and the microvascular perfusion offset index and the mean skin temperature of each anatomical region as local tissue stiffness adjustment factors. In the formula: Weighted by mean skin temperature, The weights of the microvascular perfusion offset index are... For the normalized first Mean skin temperature of each anatomical region The normalized microvascular perfusion deviation index; the local tissue stiffness modulatory factor. Converted to equivalent bending stiffness coefficient via linear mapping Using the tension boundary of the sternocleidomastoid muscle as the boundary constraint condition of the curve, a continuous reconstruction model of cervical curvature based on the time-varying viscoelastic properties of the neck was constructed. The cervical curvature curves are reconstructed frame by frame from a dynamic head, neck and shoulder image sequence using a continuous cervical curvature reconstruction model. The cervical curvature curves of all frames constitute a family of viscoelastic time-varying cervical curvature curves. The time-varying properties of viscoelasticity are characterized by stress relaxation function and creep function. The parameters of stress relaxation function are dynamically adjusted according to the rate of change of position of adjacent frames in the indirect positioning coordinate sequence of cervical spinous processes, and the parameters of creep function are dynamically adjusted according to the time series change trend of local tissue stiffness adjustment factor.

7. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network according to claim 6, characterized in that, Principal component analysis is performed on the family of curves to decouple the skull base rotation compensation component, generate a true cervical curvature morphological feature vector, decouple muscle imbalance pattern features, and calculate the cervical muscle biomechanical load index, specifically including: Principal component analysis was performed on the family of viscoelastic time-varying cervical curvature curves. The components corresponding to the rotational degrees of freedom of the skull base-axis joint were removed as skull base rotation compensation components, and the remaining components were used as the true cervical curvature change components. The projection coefficients of each frame were calculated to generate the true cervical curvature morphological feature vectors corresponding to each frame. Using the actual cervical curvature change component, the mean skin temperature of each anatomical region, the microvascular perfusion offset index, and the spatial morphological parameters of the sternocleidomastoid muscle tension boundary as input features, and three muscle state labels as supervision signals, a muscle imbalance pattern decoupling network with three parallel binary classification branches is trained. The input features are fed into the trained muscle imbalance pattern decoupled network. Three parallel binary classification branches distinguish the three muscle imbalance patterns and generate confidence scores. The discrimination results and confidence scores constitute the muscle imbalance pattern features. The mean skin temperature of each anatomical region is weighted and summed according to preset weights based on the deviation between the mean and normal reference values ​​and the microvascular perfusion deviation index. In the formula: For the first The basic mechanical load index of each anatomical region. Weighting for skin temperature deviation. The weights of the microvascular perfusion offset index are... For the first Skin temperature deviation in each anatomical region , For the first Mean skin temperature of each anatomical region The values ​​are the reference values ​​for normal perfusion; then, they are weighted and adjusted according to the confidence scores of the muscle imbalance patterns corresponding to each anatomical region, i.e.: In the formula: For the first The mechanical load index adjusted for each anatomical region. This is the magnification factor. For the first Confidence scores for muscle imbalance patterns corresponding to each anatomical region; calculation of the cervical muscle mechanical load index. In the formula: The cervical muscle mechanical load index ranges from 0 to 100. An index greater than 30 is considered an abnormal cervical muscle mechanical load. This represents the total number of anatomical regions.

8. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network according to claim 1, characterized in that, Extracting the time-frequency features of microtremors to differentiate between physiological and pathological dystonia, specifically including: Optical flow field calculations were performed on the posterior triangle of the neck, which corresponds to the high-sensitivity area of ​​skin color micro-changes in each frame of the dynamic head, neck and shoulder image sequence. The displacement of the surface pixels of the posterior triangle of the neck in the time series was extracted to obtain the micro-displacement signal. The micro-displacement signal is bandpass filtered to remove slow drift components and high-frequency noise components. The filtered micro-displacement signal is then subjected to fast Fourier transform to obtain the power spectral density distribution. Extract the first and second power integral values ​​of the power spectral density distribution within a preset frequency band, i.e.: , In the formula: The first power integral value corresponds to the preset frequency band. The frequency range is 2Hz to 8Hz, corresponding to the dominant frequency band of physiological microtremors. This is the second power integral value, corresponding to a preset frequency band. The frequency range is 8Hz to 15Hz, corresponding to the dominant frequency band of pathological dystonia tremor; the ratio of these two frequencies is calculated as the frequency domain characteristic of microtremor. ,Right now: When the frequency domain characteristics of microtremor are below a preset threshold, it is determined to be pathological dystonia; otherwise, it is determined to be physiological microtremor.

9. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network according to claim 8, characterized in that, Artifacts were removed through temporal coherence verification, specifically including: Obtain the real cervical curvature morphological feature vector and microtremor frequency domain features corresponding to each frame. Arrange the real cervical curvature morphological feature vector in time order to form a curvature feature time sequence, and arrange the microtremor frequency domain features in time order to form a tremor feature time sequence. Calculate the cross-correlation function of the tremor feature time series and the curvature feature time series in the time dimension, and extract the coherence coefficient of the cross-correlation function at zero delay, i.e.: In the formula: For the first The correlation coefficient between the dimensional components and the frequency domain characteristics of microtremors at zero delay. This represents the total number of frames in the dynamic head, neck, and shoulder image sequence. For the standardized first Frame number The true cervical curvature morphological feature vector components. For the standardized first Frequency domain characteristics of frame micro-flutter; When the coherence coefficient is lower than the preset threshold, it is determined that there are artifacts in the time period. The true cervical curvature morphology feature vector in the time period with artifacts is corrected by cubic spline interpolation, outliers are removed, and the verified true cervical curvature morphology feature vector is generated.

10. A method for identifying and detecting cervical curvature in upper crossed syndrome based on an improved HRNet network according to claim 9, characterized in that, The validated cervical curvature morphological feature vector, muscle imbalance pattern features, and cervical muscle biomechanical load index are input into the upper crossed syndrome dynamic grading network to output the dynamic risk level of upper crossed syndrome and muscle imbalance subtype, specifically including: The real cervical curvature morphology feature vector, muscle imbalance pattern feature, and cervical muscle mechanical load index, which have been verified by time-series coherence, are spliced ​​together at the feature level to form a multi-source fusion feature vector. The multi-source fusion feature vector is input into the shared bottom feature extraction layer of the dynamic hierarchical network for cross syndrome. The shared bottom feature extraction layer consists of two cascaded fully connected layers to generate shared deep features. The shared deep features are input into the dynamic risk level classification branch and the muscle imbalance subtype classification branch, respectively. Both classification branches include a fully connected layer and a Softmax activation function, which generate the dynamic risk level of upper crossed syndrome and the muscle imbalance subtype, respectively.

Images