Pregnancy nutrition state image evaluation and guidance system

Abstract

The invention relates to the technical field of medical health, in particular to a pregnancy nutrition state image evaluation and guidance system, which analyzes a whole body image of a pregnant woman in a standard posture through a differential geometry theory and a neural network technology to realize accurate evaluation of human body components and personalized nutrition guidance. The system comprises an image acquisition module, a differential geometric feature extraction module, a multi-scale tissue mapping module, a manifold neural network analysis module, a nutrition risk assessment module, a nutrition status prediction module and a guidance generation module. Different tissue types are distinguished by applying a multi-scale analysis technology, a network architecture capable of performing feature extraction and propagation in a non-Euclidean geometric space is constructed through manifold learning, accurate calculation of parameters such as a fat rate, a muscle rate and an edema index is realized, and finally personalized diet suggestions and exercise schemes are provided based on an evaluation result. And non-intrusive evaluation based on the image is realized.

Description

Technical Field

[0001] This invention relates to the field of medical and health technology, specifically to a system for image assessment and guidance of nutritional status during pregnancy, and more particularly to a system that uses differential geometry theory and neural network technology to analyze whole-body images of pregnant women, assess their nutritional status, and provide personalized guidance. Background Technology

[0002] Nutritional status during pregnancy is crucial for fetal development and the health of the pregnant woman. Traditional assessment methods, primarily relying on weight measurement, body mass index (BMI) calculation, and body circumference measurement, have several shortcomings. The measurement process is cumbersome, requires professional operation, has significant measurement errors, and is affected by various factors. More importantly, traditional methods cannot distinguish the proportion of fat, muscle, and water in weight gain, making remote and convenient monitoring difficult, and also hindering the provision of accurate personalized nutritional guidance.

[0003] Currently, there are some products on the market that utilize image technology for body composition analysis. However, these products are mainly targeted at the general population and do not take into account the unique physiological changes during pregnancy. Existing technologies mostly use standard human body models for comparative analysis, ignoring the specific changes in abdominal morphology during pregnancy. Conventional image analysis techniques struggle to distinguish between edema and fat accumulation during pregnancy, and existing methods are mostly based on statistical models, lacking precise modeling of human tissue distribution.

[0004] Furthermore, traditional nutritional guidance programs are mostly based on general principles and struggle to provide precise guidance tailored to individual differences. Dynamic monitoring and real-time feedback of nutritional status during pregnancy are also weaknesses of current technologies. Therefore, there is an urgent need for a system capable of accurately assessing nutritional status during pregnancy and providing personalized guidance. Summary of the Invention

[0005] The purpose of this invention is to provide a pregnancy nutrition status image assessment and guidance system. By analyzing the full-body images of pregnant women in standard postures using differential geometry theory and neural network technology, the system can accurately assess body composition and provide personalized nutrition guidance plans.

[0006] This invention proposes a pregnancy nutritional status image assessment and guidance system, including an image acquisition module, a differential geometric feature extraction module, a multi-scale tissue mapping module, a manifold neural network analysis module, a nutritional risk assessment module, a nutritional status prediction module, and a guidance generation module.

[0007] The image acquisition module is used to capture full-body images of the pregnant woman in a standard posture, both from the front and from the left side, and stores these images on a cloud server to provide standardized input data for subsequent analysis.

[0008] The differential geometry feature extraction module is connected with the image acquisition module, receives the front full-body picture and the left side full-body picture, extracts the curvature features and morphological invariants of the human body contour through surface differential feature analysis, and generates a human body morphological feature vector.

[0009] The multi-scale tissue mapping module is connected with the differential geometry feature extraction module, and performs multi-scale hierarchical decomposition on the human body morphological feature vector.

[0010] The manifold neural network analysis module is connected with the multi-scale tissue mapping module, and performs deep analysis based on the human body composition data.

[0011] The nutritional risk assessment module is connected with the manifold neural network analysis module, receives the human body composition data, inputs the current standard body weight gain and the current BMI change data into the pre-pregnancy prediction nutritional risk model, outputs the pre-pregnancy prediction nutritional risk value, and realizes quantitative evaluation of the nutritional risk during pregnancy.

[0012] The nutritional status prediction module is connected with the nutritional risk assessment module and the manifold neural network analysis module, inputs the pre-pregnancy prediction nutritional risk value, the upper arm muscle relaxation degree, the edema index and the abdominal muscle relaxation degree into the nutritional status prediction model, predicts the nutritional status during pregnancy, and obtains a comprehensive analysis result.

[0013] The guidance generation module is connected with the nutritional status prediction module, provides personalized dietary suggestions and exercise programs according to the analysis result, considers the individual characteristics of the pregnant woman such as age, height, weight, disease history, etc., and improves the pertinence and effectiveness of the guidance.

[0014] The present application has the following beneficial effects.

[0015] First, it realizes non-invasive evaluation based on images, without the need to use traditional measurement tools, reducing the discomfort of pregnant women and improving the convenience and acceptance of evaluation. Pregnant women only need to take photos in a standard posture to complete the evaluation, greatly simplifying the operation process.

[0016] Second, the accurate mapping relationship between image features and human body components is established through differential geometry theory, breaking through the limitations of traditional two-dimensional image analysis and improving the accuracy of human body component evaluation. The invention regards the human body surface as a Riemannian manifold, and uses curvature, geodesic and other differential geometric quantities to accurately represent the human body shape.

[0017] Third, multi-scale analysis technology is adopted to realize accurate differentiation of different tissue types including fat, muscle and water, solving the technical difficulty of traditional methods in differentiating different tissues. By dividing the human body into three scale regions of coarse, medium and fine, and establishing the mapping relationship between each region and the tissue distribution, the quantitative calculation of human body composition is realized.

[0018] Fourth, the neural network architecture based on manifold learning can extract and propagate features in non-Euclidean geometry space, improving the accuracy and stability of complex shape analysis. Especially in calculating the abdominal muscle relaxation degree, the manifold embedding coordinates are extracted by Laplace feature mapping, combined with manifold convolutional neural network and curvature features, to realize accurate quantification of abdominal shape.

[0019] Fifth, accurate personalized nutrition guidance is provided, considering the individual characteristics of pregnant women, improving the pertinence and effectiveness of guidance. The system generates targeted dietary recommendations and exercise programs based on the evaluation results, combined with individual characteristics such as age, height, weight, disease history, etc.

[0020] Sixth, dynamic monitoring of nutritional status during pregnancy is realized, which can timely discover potential risks and realize early intervention. Through regular evaluation and monitoring feedback, the system can dynamically adjust the guidance program to ensure nutritional health during pregnancy. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The overall structure diagram of the system of the invention.

[0022] Figure 2 The structure diagram of the differential geometry feature extraction module of the invention.

[0023] Figure 3 The structure diagram of the multi-scale tissue mapping module of the invention.

[0024] Figure 4 The structure diagram of the manifold neural network analysis module of the invention.

[0025] Figure 5This is a flowchart of the nutritional risk assessment module of the present invention.

[0026] Figure 6 This is a flowchart of the nutritional status prediction module of the present invention.

[0027] Figure 7 This is a structural block diagram of the generation module of the present invention. Detailed Implementation

[0028] Please refer to Figures 1-7 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, the image assessment and guidance system for nutritional status during pregnancy of the present invention includes an image acquisition module 1, a differential geometric feature extraction module 2, a multi-scale tissue mapping module 3, a manifold neural network analysis module 4, a nutritional risk assessment module 5, a nutritional status prediction module 6, and a guidance generation module 7.

[0030] Image acquisition module 1 is used to acquire full-body images of the pregnant woman in a standard posture, both from the front and left side, and stores these images on a cloud server. Differential geometric feature extraction module 2 is connected to image acquisition module 1, receives the full-body images from the front and left side, and extracts the curvature features and morphological invariants of the human body contour through surface differential feature analysis, generating a human body morphological feature vector. Multi-scale tissue mapping module 3 is connected to differential geometric feature extraction module 2, performs multi-scale hierarchical decomposition of the human body morphological feature vector, establishes a mapping relationship between the external contour and the internal tissue distribution, and calculates body composition data, including pre-pregnancy standard weight, pre-pregnancy standard BMI, current BMI, body fat percentage, muscle mass percentage, water percentage, and edema percentage.

[0031] The manifold neural network analysis module 4 is connected to the multi-scale tissue mapping module 3. Based on body composition data, it calculates abdominal muscle laxity, abdominal edema index, leg muscle laxity, leg edema index, upper arm circumference, upper arm muscle circumference, upper arm fat circumference, and abdominal circumference. The nutritional risk assessment module 5 is connected to the manifold neural network analysis module 4. It receives body composition data, inputs current standard weight gain and current BMI changes into the pre-pregnancy nutritional risk prediction model, and outputs the pre-pregnancy predicted nutritional risk value. The nutritional status prediction module 6 is connected to the nutritional risk assessment module 5 and the manifold neural network analysis module 4. It inputs the pre-pregnancy predicted nutritional risk value, upper arm muscle laxity, edema index, and abdominal muscle laxity into the nutritional status prediction model to predict pregnancy nutritional status and obtain analysis results. The guidance generation module 7 is connected to the nutritional status prediction module 6. Based on the analysis results, it provides personalized dietary recommendations and exercise programs.

[0032] In a preferred embodiment of the present application, the image acquisition module 1 acquires a frontal full-body picture and a left lateral full-body picture of the pregnant woman in a standard posture. The specification of the standard posture ensures the consistency and accuracy of image analysis.

[0033] Specifically, the standard posture for frontal full-body image shooting requires the subject to stand with feet on the ground, hips and back against the wall, eyes level, hands naturally perpendicular to the body in front, facing the image acquisition device, eyes level, smiling, arms naturally extended and relaxed, 90° angle between left and right, palms inward, and feet naturally together. The standard posture for left lateral full-body image shooting requires the subject to stand with feet on the ground, shoulders and chest facing the wall, eyes level, hands naturally perpendicular to the body in front, facing the image acquisition device, eyes level, smiling, arms naturally extended and relaxed, 90° angle between hands, palms inward, left foot on the left side and right foot on the right side.

[0034] These standard postures take into account the principles of anthropometry, ensuring the visibility and consistency of various parts of the body in the image, providing a reliable basis for subsequent analysis. The acquired images are stored in the cloud server, facilitating remote analysis and processing by the system.

[0035] The differential geometry feature extraction module 2 is one of the core innovations of the present application, which regards the human body surface as a two-dimensional Riemannian manifold embedded in three-dimensional space, and realizes the accurate representation of human tissue distribution by analyzing the differential properties of the manifold.

[0036] As shown in Figure 2 The differential geometry feature extraction module 2 includes a three-dimensional modeling unit 21, a curvature calculation unit 22, a geodesic network construction unit 23, and a morphological invariant extraction unit 24.

[0037] The three-dimensional modeling unit 21 performs edge detection and depth estimation on the frontal full-body picture and the left lateral full-body picture, constructs a three-dimensional surface model of the human body and performs parameterized representation. In an embodiment of the present application, the three-dimensional modeling process includes four stages of edge detection, depth estimation, three-dimensional surface fusion modeling, and parameterized representation.

[0038] First, an improved Canny edge detection algorithm is used to perform edge detection on the input image. The standard Canny algorithm includes four steps of image smoothing, gradient calculation, non-maximum suppression, and double threshold detection. The present application adds an optimization process for human body contour based on the standard Canny algorithm, which includes four improvements.

[0039] The first improvement is adaptive Gaussian filtering. In the image smoothing stage, adaptive Gaussian filtering is used instead of fixed scale Gaussian filtering. The standard deviation of adaptive Gaussian filtering is dynamically adjusted according to the local texture complexity of the image, and the calculation formula is:

[0040] ,

[0041] wherein, is the adaptive standard deviation at coordinate , is the base standard deviation, taking value 1.0 pixel, is the adjustment coefficient, taking value 0.3, is the texture complexity index at the position, which is obtained by calculating the local window's gray level variance. The calculation formula of the texture complexity is as follows:

[0042] ,

[0043] wherein, is the gray level value at coordinate , is the average gray level value in the local window, is the total number of pixels in the window, taking value 49, corresponding to a 7x7 window, is the window radius, taking value 3 pixels. The adaptive mechanism makes it use larger filter scale in the texture complex area to suppress noise, and use smaller filter scale in the edge clear area to retain details.

[0044] The second improvement is the multi-directional Sobel operator. In the gradient calculation stage, the multi-directional Sobel operator is used instead of the traditional horizontal and vertical two-way Sobel operator. The present application uses 8-directional Sobel operator, and the direction angles are 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315° respectively. The gradient amplitude of each direction is calculated as follows:

[0045] ,

[0046] wherein, and are the gradient components in horizontal and vertical directions respectively. The final gradient amplitude takes the maximum value of all directions:

[0047] ,

[0048] The gradient direction is the direction angle corresponding to the maximum gradient amplitude. The multi-directional gradient calculation method can more accurately detect the edges of human body contour in all directions, especially the detection effect of oblique edges is significantly better than the traditional method.

[0049] The third improvement is adaptive threshold. In the double threshold detection stage, an adaptive threshold is used instead of a fixed threshold. The application dynamically determines the high and low thresholds according to the statistical characteristics of the global gradient amplitude of the image, and the calculation formula is:

[0050] ,

[0051] ,

[0052] wherein, is the average value of the global gradient amplitude, is the standard deviation of the global gradient amplitude, is an adjustment factor, and the value is 1.5. The adaptive threshold mechanism can adjust the detection sensitivity according to the overall quality of the image, and improve the robustness of edge detection.

[0053] The fourth improvement is edge connection. In the edge connection stage, an edge direction-based connection strategy is used. For an incomplete edge chain, the system searches for edge points near its endpoints, and if the direction of the adjacent edge point is less than the threshold, preferably 30°, from the direction of the current edge chain, it is connected to the current edge chain. The strategy can effectively repair the edge breakage caused by noise or light changes.

[0054] Through the above improvements, the edge detection algorithm of the application can accurately extract the human contour edge, providing a reliable foundation for subsequent three-dimensional modeling.

[0055] After obtaining the clear human contour edge, it is necessary to estimate the depth information of each point in the image and convert the two-dimensional image into a three-dimensional representation. The application adopts a monocular depth estimation method based on deep learning. This method learns the mapping relationship from a two-dimensional image to a depth map by training a deep neural network.

[0056] Specifically, a convolutional neural network with an encoder-decoder architecture is used. The encoder part uses ResNet-50 as the backbone network to extract multi-scale features of the image. The decoder part gradually restores the spatial resolution of the depth map through upsampling and skip connection. The input of the network is an RGB image, and the output is a corresponding depth map, and the depth value represents the distance from each point in the scene to the camera.

[0057] The training of the depth estimation network uses a large-scale labeled human depth dataset containing 10,000 pairs of human RGB images and corresponding depth maps. The depth map is obtained by laser radar or structured light scanner, and the accuracy reaches millimeter level. The training loss function uses the weighted combination of depth error and gradient error:

[0058] ,

[0059] wherein, For the depth error, the absolute error between the predicted depth and the ground truth depth is defined as:

[0060] ,

[0061] where, is the total number of image pixels, is the ground truth depth of the th pixel, is the predicted depth.

[0062] For the gradient error, which is used to preserve the edge sharpness of the depth map:

[0063] ,

[0064] where, is the gradient of the ground truth depth map at the th pixel, is the gradient of the predicted depth map. The weight coefficients and are set to 1.0 and 0.5, respectively.

[0065] The network is trained using the Adam optimizer with an initial learning rate of 0.0001 and a batch size of 8. After 100 epochs of training, the network converges. The average absolute error of the depth estimation network on the test set after training is 2.3 cm, and the relative error is 3.5%, which meets the accuracy requirements of the present application.

[0066] For the front and side images, depth estimation is performed separately to obtain two depth maps. Since the front image mainly provides depth information for the front surface of the body, and the side image provides depth information for the side surface of the body, they have complementary properties. In the subsequent three-dimensional fusion stage, this complementarity will be fully utilized to construct a complete three-dimensional surface model of the human body.

[0067] After obtaining the depth maps of the front and side, they need to be fused into a complete three-dimensional surface model. The present application uses a voxel-based fusion method to divide the three-dimensional space into a uniform voxel grid, and each voxel stores the occupancy probability at that location.

[0068] First, according to the intrinsic and extrinsic parameters of the camera, each pixel in the depth map is back-projected to the three-dimensional space to obtain a three-dimensional point cloud. For the front image, the camera is located directly in front of the body at a distance of 150 cm, and the viewing angle is perpendicular to the front of the body. For the side image, the camera is located on the left side of the body at a distance of 150 cm, and the viewing angle is perpendicular to the left side of the body.

[0069] The three-dimensional point clouds of the two views are represented in a unified world coordinate system, which requires coordinate transformation. Let the front camera coordinate system be , and the side camera coordinate system be , the world coordinate system is . The transformation formula of the point cloud from the camera coordinate system to the world coordinate system is:

[0070] ,

[0071] wherein, is the coordinate of the point in the world coordinate system, is the coordinate of the point in the camera coordinate system, is a rotation matrix, is a translation vector. For the front camera, the rotation matrix is a unit matrix, and the translation vector cm. For the side camera, the rotation matrix is a matrix rotating 90° around axis, and the translation vector cm.

[0072] In the world coordinate system, a three-dimensional space is divided into a voxel grid, and the voxel size is 1cm × 1cm × 1cm. For each voxel , according to the number of point clouds projected into the voxel, the occupancy probability thereof is calculated:

[0073] ,

[0074] wherein, is the number of points projected into the voxel , and is a regularization constant, which is 5, used to avoid division by zero error. The occupancy probability represents the possibility that the voxel is located on the human body surface, and the greater the value, the more likely it belongs to the human body surface.

[0075] In order to fuse information from different perspectives, the invention adopts a weighted average method. For the same voxel, the occupancy probability from the front perspective is , the occupancy probability from the side perspective is , and the fused occupancy probability is:

[0076] ,

[0077] wherein, and are weight coefficients, which are dynamically determined according to the visibility of the perspective. For the voxels on the front surface of the body, the weight of the front perspective is larger, and for the voxels on the side surface of the body, the weight of the side perspective is larger. The calculation formula of the weight is:

[0078] , ,

[0079] where, is the angle between the surface normal of a voxel and the front view direction, is the angle with the side view direction. The smaller the angle, the clearer the observation of the voxel from that view, the larger the weight.

[0080] By voxel fusion, the present application can comprehensively utilize the information of front and side images to construct a complete human three-dimensional surface model. Next, the surface needs to be extracted from the voxel model and parameterized.

[0081] After obtaining the voxel model, the isosurface is extracted by using the marching cubes algorithm, and the voxel model is converted into a triangular mesh representation. The marching cubes algorithm traverses each voxel, and according to the occupancy probability of its 8 vertices, judges the topological structure of the isosurface in the voxel, and generates the corresponding triangular patches. The threshold of the isosurface is set to 0.5, that is, the voxel with an occupancy probability greater than 0.5 is considered to be located inside the human body.

[0082] The extracted triangular mesh contains about 50000 vertices and 100000 triangular patches, which accurately describes the three-dimensional geometry of the human body. In order to facilitate subsequent differential geometry analysis, the triangular mesh needs to be parameterized, that is, to establish a one-to-one correspondence between the three-dimensional surface and the two-dimensional parameter domain.

[0083] The present application adopts the conformal parameterization method, which keeps the local angle unchanged during the parameterization process and reduces the shape distortion. Specifically, the human surface is divided into multiple topologically simple regions, such as trunk, limbs, head, etc., and each region is parameterized separately. For each region, select a boundary as the boundary of the parameter domain, such as the cross section of the upper and lower ends of the trunk region, which is mapped to the upper and lower boundaries of the two-dimensional parameter domain.

[0084] The mathematical representation of parameterization is a mapping function which maps each point on the three-dimensional surface to a point in the two-dimensional parameter domain. Conformal parameterization requires that the mapping maintains local conformality, that is, maintains the angle, which can be achieved by solving the Laplace equation:

[0085] ,

[0086] where, is the Laplace operator. On the discrete triangular mesh, the Laplace operator is discretized by cotangent weight:

[0087] ,

[0088] where, ​the function value at the vertex , the vertex set of the neighborhood of the vertex , the local area of the vertex , and the corresponding angle in the triangle on both sides of the edge .

[0089] By solving the Laplace equation combined with boundary conditions, the parametric coordinates of each vertex can be obtained. After parameterization, the human body surface has a regular representation in the two-dimensional parameter domain, which facilitates curvature calculation and feature extraction.

[0090] At this point, the three-dimensional modeling unit 21 has completed the complete conversion from two-dimensional images to three-dimensional parameterized surfaces, laying the foundation for subsequent differential geometry analysis.

[0091] The curvature calculation unit 22 calculates the mean curvature and Gaussian curvature of each point on the surface based on the human three-dimensional surface model, generating a curvature distribution map. Curvature is an important geometric quantity that describes the local bending degree of a surface, and is of great significance for understanding the morphological features of the human body.

[0092] In differential geometry, the curvature of each point on a surface is described by two principal curvatures and , which are obtained by analyzing the first and second fundamental forms of the surface. The mean curvature and Gaussian curvature are defined as:

[0093] , ,

[0094] The mean curvature describes the overall bending degree of the surface, and the Gaussian curvature reflects the intrinsic geometric properties of the surface. The region with positive Gaussian curvature is convex, such as the protruding part of the abdomen, and the region with negative Gaussian curvature is saddle-shaped, such as the side of the waist.

[0095] On a discrete triangular mesh, the curvature is calculated using a local fitting method. For each vertex , select the vertices in its one-ring neighborhood to form a local surface patch. Perform quadratic surface fitting on the local surface patch, and the fitted quadratic surface equation is:

[0096] ,

[0097] where are the unknown coefficients. These coefficients are solved by least squares method, which minimizes the deviation of the fitted surface from the vertices of the local patches. After the fitting, the principal curvatures of the quadric surface can be obtained from the eigenvalues of the Hessian matrix:

[0098] ,

[0099] Principal curvatures and are the two eigenvalues of the matrix . Then the mean curvature and Gaussian curvature are calculated:

[0100] , ,

[0101] The curvature distribution map is obtained by calculating the curvature of all vertices on the human body surface. The curvature distribution map is displayed in pseudo-color mode, different curvature values correspond to different colors, which intuitively reflects the bending characteristics of each part of the human body.

[0102] The curvature distribution map has important applications in pregnancy nutrition assessment. For example, the average curvature of the abdomen changes with gestational age, which can be used to monitor fetal development and changes in maternal body shape. The curvature characteristics of the legs and upper arms can reflect the distribution of muscle and fat, which helps to assess the nutritional status.

[0103] The geodesic network construction unit 23 defines the human body key points as geodesic network nodes, calculates the shortest geodesic distance between nodes, and forms a geodesic network. The geodesic is the shortest path between two points on a surface, and the geodesic distance is the length along the geodesic, which is an important tool to describe the intrinsic geometry of the surface.

[0104] A series of key points are defined on the human body surface, including anatomical landmark points and functional points. Anatomical landmark points include the top of the head, shoulder peak, elbow joint, wrist joint, anterior superior iliac spine, knee joint, ankle joint, etc. These points have standard definitions in anthropometry. Functional points are defined according to specific application requirements, such as the center point of the abdomen and the highest point of the hips. The present invention defines a total of 50 key points, covering the main parts of the human body.

[0105] For each pair of key points and , the geodesic distance between them is calculated. The calculation of geodesic distance uses the fast marching method, which efficiently calculates the geodesic distance from the source point to all other points on the surface by solving the Eikonal equation.

[0106] The basic idea of the fast marching method is to regard the calculation of the geodesic distance field as a wave front propagation process. From the source point The geodesic distance of the source point is 0. The wave front propagates outward at a constant speed, and the time it takes to reach other points on the surface is the geodesic distance of the point. On a discrete triangular mesh, the fast marching method iteratively updates the geodesic distance of the vertices until the distances of all vertices converge.

[0107] In a specific implementation, a priority queue is maintained to store the vertices to be updated and their current estimated geodesic distances. Initially, the geodesic distance of the source point is 0 and is added to the queue. Each time the vertex with the smallest geodesic distance is taken out of the queue, it is marked as determined, and the geodesic distances of its neighboring vertices are updated. The update formula is based on the geometry of the local triangle, ensuring that the estimate of the geodesic distance follows the Eikonal equation. The iteration process continues until all vertices are marked as determined, at which point the geodesic distance field from the source point to all vertices is obtained.

[0108] The above process is repeated for all key point pairs to obtain the geodesic distance matrix where the element represents the geodesic distance between key points and . The geodesic distance matrix constitutes a representation of the geodesic line network, which encodes the intrinsic geometry of the human body surface and is not affected by external viewing angles and posture changes.

[0109] The geodesic line network plays an important role in shape analysis. Traditional Euclidean distances are affected by changes in body posture, such as raising or lowering the arms, which changes the Euclidean distance from the shoulder to the wrist. Geodesic distances measure along the surface and are independent of posture, better reflecting the intrinsic shape of the human body. By analyzing the geodesic line network, posture-independent shape features can be extracted, improving the robustness of evaluation.

[0110] The shape invariant extraction unit 24 extracts posture-independent shape features based on the surface differential invariant theory and constructs a shape feature vector. Shape invariants are geometric quantities that remain unchanged under certain transformations, such as Gaussian curvature, which remains unchanged under isometric transformation, and average curvature, which remains unchanged under rigid body transformation.

[0111] The present invention extracts the following types of shape invariants.

[0112] The first type is curvature integral invariant. For each region of the human body, such as the head, torso, limbs, etc., the curvature integral within the region is calculated:

[0113] , ,

[0114] where is the surface of the th region, and ​respectively the mean and Gaussian curvatures, is the area element. The curvature integral reflects the overall bending property of the region, and is insensitive to local small deformations.

[0115] On a discrete triangle mesh, the curvature integral is computed by weighted sum of curvatures of all triangles:

[0116] , ,

[0117] where, is the set of triangles contained in the th region, and are the mean and Gaussian curvatures of triangle , is the area of triangle.

[0118] The second category is geodesic distance statistics. Based on the geodesic network, the statistical features of geodesic distances between key points in each region are computed, such as mean, standard deviation, maximum, minimum, etc. These statistics reflect the size and shape properties of the region.

[0119] For the th region, containing key point set , the mean of geodesic distances is:

[0120] ,

[0121] where, is the number of key points. Similarly, the standard deviation , maximum and minimum are computed.

[0122] The third category is shape distribution descriptor. Shape distribution is computed by randomly sampling pairs of points on the surface, computing their geometric quantities such as geodesic distance, Euclidean distance, angle, etc., and forming a histogram of these quantities. Shape distribution can compactly describe the overall shape property of the surface, and is robust to pose and view changes.

[0123] For human body surface, 10000 pairs of points are randomly sampled, and the geodesic distance between each pair is computed to construct a histogram of geodesic distances. The histogram is divided into 50 bins, and the count of each bin is normalized to form a dimension of the shape distribution descriptor. The descriptor is a 50-dimensional vector, denoted as .

[0124] The fourth category is volume and surface area features. For each region of the human body, its volume and surface area The volume reflects the size of the region, and the ratio of the surface area to the volume reflects the compactness of the region.

[0125] The volume and surface area are calculated based on the triangle mesh. For a closed surface, the volume can be obtained by summing the contributions of all the triangular facets:

[0126]

[0127] wherein, are the coordinates of the three vertices of the triangular facet The surface area is the sum of the areas of all the triangular facets:

[0128]

[0129] The above-mentioned shape invariants are combined to form a human body shape feature vector . Assuming that the human body is divided into 10 regions, 2-dimensional curvature integral, 4-dimensional geodesic distance statistics, 2-dimensional volume and surface area are extracted for each region, a total of 8-dimensional features, plus the global shape distribution descriptor 50-dimensional, a total of shape feature vector dimension.

[0130] The shape feature vector compactly encodes the geometric and topological properties of the human body, providing rich input information for subsequent multi-scale tissue mapping. Through the differential geometry feature extraction module 2, the present application realizes the conversion from the original image to the high-level geometric features, which is the key link of the whole system.

[0131] The multi-scale tissue mapping module 3 is another core innovation point of the present application, which establishes the mapping relationship between the human body contour and the internal tissue distribution through multi-scale analysis method, and realizes the quantitative estimation of fat, muscle, water and other tissue components.

[0132] As shown in Figure 3 , the multi-scale tissue mapping module 3 includes a multi-scale decomposition unit 31, a tissue distribution density calculation unit 32 and a part-specific compensation unit 33.

[0133] The multi-scale decomposition unit 31 constructs a curvature scale space, calculates the curvature change of the human body contour at different scales, and identifies the tissue boundary surface through the extreme points in the scale space. The scale space theory is an important tool in computer vision, which can extract structure information at different levels by analyzing signals at different scales.

[0134] In the present application, the construction of the curvature scale space is based on the heat diffusion equation. The curvature field on the human surface is regarded as the initial temperature distribution, and the evolution with time satisfies the heat diffusion equation: ​​

[0135] ,

[0136] where is the Laplacian operator. The solution of the heat diffusion equation represents a smoothed version of the curvature field at different scales, with the scale parameter controlling the degree of smoothing, The larger

[0137] On a discrete triangular mesh, the heat diffusion equation is solved using an implicit Euler method:

[0138] ,

[0139] where is the identity matrix, is the Laplacian matrix, is the time step, is the curvature field at time By iterative solving, the curvature field at different scales is obtained.

[0140] The present invention computes the curvature field at five scales, 0.1s, 0.5s, 1s, 2s, and 5s, forming a curvature scale space. At each scale, the spatial distribution of curvature is analyzed, and the extreme points are identified. The extreme points correspond to the locations where the curvature changes significantly, which are often the boundaries of different tissues. For example, at the fine scale of 0.1s, the curvature field retains rich details and can identify the boundary between skin and subcutaneous fat. At the medium scale of 1s, the curvature field is smoother and can identify the boundary between fat and muscle. At the coarse scale of 5s, only the main structural features are retained, and the outlines of major organs and bones can be identified.

[0141]

[0142] By identifying the tissue boundaries at multiple scales, the present invention achieves hierarchical segmentation of the human body, dividing the human surface into three levels: coarse scale, medium scale, and fine scale.

[0143] At the coarse scale level, the human body is divided into head, torso, and limb regions. This level of division is based on the main anatomical structures of the human body and is achieved by identifying the connection points between the torso and the head, torso and limbs. Specifically, in the coarse scale of 5s in the curvature scale space, the extreme points of curvature are sought, which correspond to the joints and connection points of the body, such as the neck, shoulders, hips, etc.

[0144] ​​​​​​At the meso-scale level, the torso is further divided into chest, abdomen, waist, and hip regions. This level of division is based on anatomical partitions of the torso, achieved by identifying the changes in curvature in the horizontal direction. At the meso-scale of the curvature scale space, the average value of curvature is calculated along the longitudinal axis of the torso at 1 -second intervals, and a curvature-height curve is plotted. The extreme points on the curve correspond to the boundaries of different regions, such as the boundary between the chest and abdomen at the level of the lower edge of the sternum, and the boundary between the waist and hips at the level of the iliac crest.

[0145] At the fine-scale level, the abdomen is further divided into upper abdomen, umbilical region, and lower abdomen regions. This level of division focuses on the abdominal region of particular interest during pregnancy, achieved by identifying local curvature features. At the fine-scale of the curvature scale space, the curvature distribution of the abdominal surface is analyzed to identify the dividing lines of curvature changes. The upper abdomen corresponds to the region from the lower edge of the rib to the upper umbilical region, the umbilical region corresponds to the region around the umbilical, and the lower abdomen corresponds to the region from the lower umbilical to the pubic symphysis.

[0146] Through multi-scale decomposition, the present application divides the human body surface into regions of multiple scales, each corresponding to different levels of anatomical structure, providing fine spatial resolution for subsequent tissue mapping.

[0147] The tissue distribution density calculation unit 32 defines a tissue distribution density function, establishes a mapping relationship based on curvature features and gray scale distribution, and generates a continuous density distribution function. The tissue composition of the human body includes skin, subcutaneous fat, muscle, bone, etc., which differ in density, gray scale, and curvature features. The present application realizes quantitative estimation of tissue distribution by establishing a mapping relationship between these features and tissue types.

[0148] For each point on the human surface, a tissue distribution density function is defined, which represents the tissue density at a distance of from the surface point in the direction of the normal vector. The unit of tissue density is g / cm³, and the density ranges of different tissues are as follows: skin about 1.1 g / cm³, fat about 0.9 g / cm³, muscle about 1.06 g / cm³, and bone about 1.85 g / cm³.

[0149] The calculation of the tissue distribution density function is based on the following physical model. It is assumed that the human tissue is distributed in layers along the normal vector direction, from outside to inside, in the order of skin layer, fat layer, muscle layer, and bone layer. The thickness and density of each layer are estimated according to the curvature features and gray scale distribution.

[0150] ​​First, estimate the thickness of each layer. The skin layer thickness is relatively fixed, approximately 2 mm. The fat layer thickness is related to local curvature and body mass index; areas with greater curvature, such as the protruding abdomen, have thicker fat, while areas with less curvature, such as the distal extremities, have thinner fat. The muscle layer thickness is related to local grayscale and curvature; areas with well-developed muscles, such as the thighs and upper arms, have thicker muscles.

[0151] Fat layer thickness The estimation formula is:

[0152] ,

[0153] in, The basal fat layer thickness is set to 5 mm. For point The average curvature, The curvature coefficient is 10 mm / (1 / cm). The value is 2 mm / (kg / m²), where BMI is the body mass index. This formula reflects the trend of fat layer thickness increasing with curvature and BMI.

[0154] Muscle layer thickness The estimation formula is:

[0155] ,

[0156] in, The base muscle layer thickness is set at 20 mm. For point The image grayscale value, The average grayscale value. This is the grayscale coefficient, with a value of 0.3 mm / grayscale level. This formula reflects the positive correlation between muscle layer thickness and image grayscale; areas with well-developed muscles are typically darker in the image.

[0157] Based on the thickness of each layer, the tissue distribution density function Defined as a piecewise function:

[0158] ,

[0159] in, The skin layer thickness is set to 2 mm. These are the densities of skin, fat, muscle, and bone, respectively.

[0160] To avoid abrupt changes in density, a smooth transition is introduced at the boundaries between layers. The sigmoid function is used for smoothing.

[0161] ,

[0162] in, and The density of two adjacent layers, For the boundary position, The steepness coefficient controls the speed of the transition, and its value is 5 mm⁻¹.

[0163] By using the tissue distribution density function, the volume and mass of each region can be calculated, thus obtaining the proportion of each tissue. For the th Each region, fat volume The calculation formula is:

[0164] ,

[0165] in, For the first The curved surface of each region The area is in yuan. This is a depth element. In discrete computation, numerical integration is used:

[0166] ,

[0167] in, As vertices The corresponding local area. Similarly, calculate the muscle volume. and water volume Water is mainly found in muscle and adipose tissue, with muscle containing about 75% water and fat containing about 10% water.

[0168] The formulas for calculating fat percentage, muscle percentage, and water percentage are as follows:

[0169] , , ,

[0170] in, For the first The total volume of each region.

[0171] To improve the accuracy of tissue mapping, this invention employs deep learning methods to optimize the aforementioned physical model. Specifically, a convolutional neural network is constructed, with the input being the three-dimensional shape features and two-dimensional image features of each region, and the output being the fat percentage, muscle percentage, and water percentage of each region.

[0172] As described in claim 6, the convolutional neural network comprises 5 convolutional layers and 3 fully connected layers. The convolutional layers are used to extract spatial features, and the fully connected layers are used for feature fusion and regression prediction. The network input is a 130-dimensional morphological feature vector. and image features of each region. Image features are extracted by applying a pre-trained CNN network ResNet-18 to image patches of each region, resulting in a 512-dimensional feature vector.

[0173] The morphological features and the image features are concatenated to obtain dimensional input feature vector. The feature vector is input into a 5-layer convolutional neural network, followed by batch normalization and ReLU activation function after each convolution. The number of channels of the convolutional layers are 128, 256, 512, 512, 256, respectively. The size of the convolution kernel is 1x1, because the input is already a high-level feature vector, and spatial convolution is not needed.

[0174] After 5 layers of convolution, the dimension of the feature vector is 256. Next, it is input into a 3-layer fully connected network with 256, 128, 30 neurons, respectively. The 30 neurons of the last layer correspond to the 3 tissue proportion fat rates, muscle rates, and water rates of the 10 regions, with a total of 30 outputs.

[0175] The network training uses a large-scale labeled data set containing 5000 human samples, each containing a three-dimensional surface model, an image, and real tissue composition data measured by medical imaging such as CT and MRI. The training loss function is the mean square error:

[0176] ,

[0177] wherein, is the number of samples, is the th real tissue proportion of the th sample, is the predicted value.

[0178] The network training uses the Adam optimizer with an initial learning rate of 0.001 and a batch size of 32, and converges after 80 epochs of training. The average absolute error of the network after training on the test set is 2.1%, which can accurately predict the tissue composition of each region of the human body.

[0179] Through the tissue distribution density calculation unit 32, the present application establishes an accurate mapping relationship from external morphological features to internal tissue distribution, realizing non-invasive estimation of human composition.

[0180] The part-specific compensation unit 33 constructs a compensation function for pregnancy-specific regions, realizing time-varying parameter adjustment strategies for different gestational weeks. The body shape of pregnant women changes significantly, especially the abdominal region gradually increases with fetal development, and the tissue distribution of the abdomen is significantly different from that of non-pregnant women. Traditional human composition estimation models do not consider these specialities, which will lead to evaluation errors.

[0181] The present application constructs a compensation function for the abdominal region, which corrects the thickness estimates of the fat layer and muscle layer. The compensation function takes into account the influence of gestational weeks in weeks and part coordinates and is defined as

[0182] ,

[0183] ,

[0184] wherein and are the fat layer and muscle layer thicknesses estimated by the base model, and are compensation factors.

[0185] The compensation factors and are obtained by analyzing medical image data of pregnant women. For the abdominal region, the fat layer thickness changes little with increasing gestational weeks, but the muscle layer thins due to the separation of the rectus abdominis muscle, and the uterus grows to occupy more space. Therefore, is close to 0, is negative.

[0186] Specifically, and are calculated as

[0187] , ,

[0188] wherein and are the baseline compensation coefficients, taking values of 0.05 and -0.15 respectively, is a spatial weight function, reflecting the degree of influence of the part by gestational weeks, is a gestational week weight function, reflecting the influence intensity of gestational weeks on body shape.

[0189] The spatial weight function is defined as

[0190] ,

[0191] wherein is the abdominal center point, is the geodesic distance from the point to the abdominal center, is a spatial scale parameter, taking a value of 15 cm. This function indicates that the closer to the abdominal center, the stronger the compensation effect.

[0192] The gestational week weight function is defined as:

[0193] ,

[0194] The function reflects the degree of body shape change at each stage of pregnancy. The change is small in the early pregnancy stage of 12 weeks or less, and the weight is 0-0.2. The change accelerates in the middle pregnancy stage of 13-27 weeks, and the weight increases from 0.2 to 0.8. The change is most significant in the late pregnancy stage of 28 weeks or more, and the weight increases from 0.8 to 1.0.

[0195] In practical applications, in order to improve flexibility, the present application adopts a segmented function form to define and Different parameter values are used for different stages of pregnancy. The pregnancy period can be divided into three stages: early stage, 12 weeks or less; middle stage, 13-27 weeks; and late stage, 28 weeks or more. Different parameter values are set for each stage.

[0196] Through site-specific compensation, the present application can accurately handle the morphological changes specific to pregnancy and improve the accuracy of tissue distribution estimation.

[0197] The manifold neural network analysis module 4 is another core innovation point of the present application, which combines the manifold learning theory in differential geometry with deep neural networks to construct a network architecture that can perform feature extraction and propagation in a non-Euclidean geometric space.

[0198] As shown in Figure 4 , the manifold neural network analysis module 4 includes a manifold embedding layer 41, a geometric feature propagation layer 42, and a parameter prediction layer 43.

[0199] The manifold embedding layer 41 constructs a mapping function from an image to a manifold space, designs a convolution operation suitable for curved surface structure, and establishes a unified representation space for front and side images. Manifold embedding is a process of mapping high-dimensional data to a low-dimensional manifold, which helps to capture the intrinsic structure of the data.

[0200] The manifold embedding function used in the present application maps a two-dimensional image space to a manifold space . The function is implemented through a neural network, and the specific structure is a multi-layer convolutional neural network, which includes multiple convolutional layers, pooling layers, and fully connected layers. Among them, represents a two-dimensional Euclidean space, i.e., an image space, represents a manifold space, which is a non-Euclidean space, represents a mapping function from the image space to the manifold space .

[0201] To accommodate curved surface structures, this invention designs a surface convolution operation. Traditional convolution is performed in Euclidean space, while surface convolution takes into account the geometric properties of the surface and defines the convolution kernel based on geodesic distance rather than Euclidean distance.

[0202] For manifolds Points on Points in its neighborhood Surface convolution is defined as:

[0203] ,

[0204] in, Indicates input features, Indicates output features, Point The neighborhood is A group of points near the point, Point and The geodesic distance between them, that is, along the manifold surface from arrive The shortest path length, This represents a weighting function based on geodesic distance; the closer the distance, the greater the weight. Point Input features at the location, Point Output characteristics at that location This represents the summation operation, for All points within the neighborhood of a point Perform a weighted summation.

[0205] In practical implementations, approximate methods are used to calculate surface convolutions for computational efficiency, such as the spectral domain convolution method based on the graph Laplacian operator.

[0206] To calculate abdominal muscle relaxation, this invention employs the following steps in a manifold neural network analysis module, which define the technical solution for the calculation.

[0207] The first step is to extract local geometric information of the abdominal region from the 3D human body surface model. The abdominal region is defined as follows: within the torso, its upper boundary is at the level of the lower edge of the sternum, its lower boundary is at the level of the anterior superior iliac spine, and its lateral boundaries are at the mid-axillary line. The 3D surface of this region is represented using a manifold. This means that each point on the manifold represents... Having three-dimensional coordinates and local tangent space.

[0208] The second step involves using Laplacian eigenmaps to extract the manifold embedding coordinates of the ventral surface. For the ventral surface... One sampling point, preferred = 2000, construct the adjacency graph where is the set of vertices, is the set of edges. If the Euclidean distance between two vertices and is less than a threshold , which is set to 3 cm, an edge is connected between them.

[0209] The weight of an edge is calculated using a Gaussian kernel function:

[0210] ,

[0211] where is the Euclidean distance between two points, is the kernel width, which is set to 1.5 cm.

[0212] Construct the Laplacian matrix :

[0213] ,

[0214] where is the weight matrix, is the matrix element, is the degree matrix, which is a diagonal matrix with diagonal elements .

[0215] Solve the generalized eigenvalue problem:

[0216] ,

[0217] Get the eigenvalues and the corresponding eigenvectors . Select the eigenvectors corresponding to the first smallest non-zero eigenvalues as the low-dimensional embedding coordinates of the manifold, where is set to 10. The low-dimensional embedding coordinate matrix , the th row of the th column is the -dimensional embedding coordinate of the th sample point.

[0218] Step 3, construct the manifold convolutional neural network based on the manifold embedding coordinates. This network is specially designed to process manifold structure data, and the network architecture includes manifold convolutional layer, manifold pooling layer and fully connected layer.

[0219] The manifold convolutional layer performs convolution operation on the local neighborhood of the manifold, and the convolution kernel function is:

[0220] ,

[0221] in, For the first Layer Each convolutional kernel at point The output at that location, For point The set of neighborhood points, For edge weights, For learnable convolution kernel parameters, The ReLU activation function is used. The manifold neural network contains four manifold convolutional layers, with 32, 64, 128, and 256 kernels in each layer, respectively.

[0222] Manifold pooling layers perform downsampling operations on the manifold, reducing computational complexity. The pooling operation selects key points on the manifold and uses the farthest point sampling method to ensure that the sampling points are uniformly distributed on the manifold.

[0223] The fully connected layer maps manifold features to a muscle relaxation index. The fully connected layer consists of two layers with 512 and 1 neurons respectively, and the last layer outputs the neural network relaxation component.

[0224] The fourth step involves calculating abdominal muscle relaxation based on a fusion of the output of a manifold neural network and the curvature features of the abdominal surface. Relaxation index. The calculation formula is:

[0225] ,

[0226] in, For the relaxation component based on curvature, The relaxation component is the output of the manifold neural network. For the relaxation component based on skin texture, , , These are weighting coefficients, with values ​​of 0.3, 0.5, and 0.2 respectively, to ensure... .

[0227] Relaxation component based on curvature The calculation formula is:

[0228] ,

[0229] in, For the abdominal curved surface The average curvature of each sampling point This represents the local area weight at that point. Total number of sampling points. Mean curvature. Defined as principal curvature and Arithmetic mean:

[0230] ,

[0231] Principal curvatures are obtained by fitting a quadric surface to the local surface. The more relaxed the abdominal muscles are, the flatter the surface is, and the smaller the average curvature is. Conversely, when the muscles are tight, the surface curvature is larger.

[0232] Relaxation component output by the manifold neural network Obtained by training, the training data contains 1000 three-dimensional models of abdominal regions of pregnant women and corresponding professional medical assessment relaxation scores, and the score range is 1-10. The training loss function is mean square error:

[0233] ,

[0234] Wherein, is the number of training samples, is the network predicted relaxation of the th sample, is the true relaxation score of the sample. The network training adopts the Adam optimizer, the learning rate is 0.001, the batch size is 16, and the training converges after 50 epochs.

[0235] Relaxation component based on skin texture According to the texture feature calculation of the abdominal region image, the texture feature is extracted by using local binary pattern, and the texture complexity The calculation formula is:

[0236] ,

[0237] Wherein, is the total number of LBP patterns, which is 256, is the occurrence probability of the th pattern. The skin texture complexity is positively correlated with the muscle relaxation, and the more complex the texture is, such as the appearance of relaxation lines and stretch marks, the higher the muscle relaxation is. Relaxation component The relaxation component is obtained by normalizing the texture complexity:

[0238] ,

[0239] Wherein, and are the minimum and maximum values of the texture complexity in the training data set, which are 3.5 and 8.2 respectively.

[0240] The final calculated abdominal muscle relaxation The value range is 0-1, wherein 0 represents that the muscle is completely tight, and 1 represents that the muscle is completely relaxed. In the application of the present application, the relaxation greater than 0.7 is considered to be significantly relaxed, which needs to be paid attention to and guided by exercise.

[0241] Through the detailed steps of the above-described manifold neural network analysis and multi-feature fusion calculation, this invention clarifies the method for calculating abdominal muscle relaxation: extracting manifold embedding coordinates through Laplacian feature mapping, extracting manifold features based on manifold convolutional neural network, and combining abdominal surface curvature features and skin texture features for weighted fusion, thus solving the technical problem of how to calculate abdominal muscle relaxation in the claims.

[0242] The geometric feature propagation layer 42 is designed with a geodesic distance-based feature propagation mechanism to achieve non-uniform diffusion of features on the manifold. Geometric feature propagation is the process of transferring and fusing features on the manifold, taking into account the geometric structure of the manifold, and can better preserve boundary information.

[0243] The feature propagation mechanism used in this invention is based on the diffusion equation:

[0244] ,

[0245] in, The representation is defined on the manifold The feature function on the manifold maps each point on the manifold to a feature vector. This represents a time parameter that controls the degree of diffusion. Representing a manifold The Laplace-Beltrammian operator on surfaces is a generalization of the Laplace operator to surfaces, used to describe the diffusion process on surfaces. Characteristic function Regarding time The partial derivatives of .

[0246] The discretized form of the diffusion equation is:

[0247] ,

[0248] in, It's the step size parameter, which controls the diffusion rate. Indicates time The characteristic function of time, Indicates time The characteristic function of time, Represents the characteristic function Apply the Laplace-Beltrami operator.

[0249] To achieve non-uniform diffusion, this invention introduces an adaptive step-size mechanism, which adjusts the step size according to local geometric characteristics. The value of . In areas of high curvature, such as the turning points of the human body contour, a smaller value is used. Values ​​that preserve detail, and in low curvature regions, larger values ​​are used. This value accelerates the diffusion process.

[0250] The parameter prediction layer 43 constructs a multi-task learning structure, simultaneously predicting body composition parameters such as fat percentage, muscle percentage, water percentage, and edema percentage. Multi-task learning enables the sharing of low-level feature representations, improving the accuracy and robustness of predictions.

[0251] The multi-task learning framework employed in this invention comprises a shared feature extraction network and multiple task-specific networks. The shared network extracts general feature representations, while the task-specific networks predict different parameters based on these features.

[0252] For the task Its loss function is defined as:

[0253] ,

[0254] in, It is the sample size. Indicates sample For the task The true value, Indicates sample For the task The predicted value, Indicates task loss function, This indicates that the loss is averaged over all samples. Indicates the sample From 1 to Summation, Indicates sample For the task The mean square error is the squared difference between the predicted value and the actual value.

[0255] The total loss function is the weighted sum of the losses from each task:

[0256] ,

[0257] in, It's the number of tasks. It is a task The weight, Represents the total loss function. Indicates the task From 1 to Summation, Indicates task The weighted loss.

[0258] During training, the total loss function is minimized. Optimize network parameters. To balance the learning difficulty of different tasks, this invention employs a dynamic weight adjustment strategy, automatically adjusting weights based on the learning progress of each task. .

[0259] Based on the manifold neural network analysis, the present application can accurately calculate the abdominal muscle relaxation degree, abdominal edema index, leg muscle relaxation degree, leg edema index, upper arm circumference, upper arm muscle circumference, upper arm fat circumference and abdominal circumference, etc. parameters, providing reliable basis for nutritional status evaluation.

[0260] The nutritional risk assessment module 5 receives the body composition data, inputs the current standard weight gain and current BMI change data into the pre-pregnancy prediction nutritional risk model, and outputs the pre-pregnancy prediction nutritional risk value.

[0261] As shown in Figure 5 , the workflow of the nutritional risk assessment module 5 includes obtaining the body composition data of the pregnant woman, extracting the current standard weight gain and current BMI change, inputting into the pre-pregnancy prediction nutritional risk model, calculating the nutritional risk probability value, and classifying the risk according to the threshold value.

[0262] The pre-pregnancy prediction nutritional risk model is obtained by the following steps. First, obtain the pre-pregnancy and pregnancy data of pregnant women with known nutritional risk assessment values. Then, calculate the corresponding pre-pregnancy standard weight gain and current BMI change according to these data. Finally, each set of pre-pregnancy standard weight gain and current BMI change data is used as a training sample for model training.

[0263] In an embodiment of the present application, the pre-pregnancy prediction nutritional risk model is implemented using a support vector machine algorithm. The model input is a two-dimensional feature vector , wherein represents the pre-pregnancy standard weight gain, in kg, represents the current BMI change, in kg / m². The model output is the nutritional risk probability value , the larger the value, the higher the risk.

[0264] According to the nutritional risk probability value predicted by the model, the present application classifies the risk into four levels. 0 to 0.15 is the first class, which is a good nutritional status and does not increase the nutritional risk. Greater than 0.15 to 0.23 is the second class, which is slightly poor in nutrition and may increase the nutritional risk. Greater than 0.23 to 0.35 is the third class, which is a higher nutritional risk and needs to consult a nutritionist. Greater than 0.35 to 1 is the fourth class, which is a very high nutritional risk.

[0265] The nutritional status prediction module 6 is connected with the nutritional risk assessment module 5 and the manifold neural network analysis module 4, inputs the pre-pregnancy prediction nutritional risk value, upper arm muscle relaxation degree, edema index and abdominal muscle relaxation degree into the nutritional status prediction model, and performs pregnancy nutritional status prediction to obtain the analysis result.

[0266] As shown in Figure 6As shown, the workflow of the nutritional status prediction module 6 includes receiving the pre-pregnancy predicted nutritional risk value and the human body parameters, inputting into the nutritional status prediction model, performing nutritional status prediction, and generating the analysis results.

[0267] The nutritional status prediction model is obtained by the following steps. First, pregnant women samples with different nutritional status evaluation values are set. Then, the pre-pregnancy standard weight gain, the pre-pregnancy predicted nutritional risk, the edema index, the upper arm muscle relaxation degree and the abdominal muscle relaxation degree of each pregnant women sample are measured. Next, the abdominal muscle relaxation degree, the edema index and the upper arm muscle relaxation degree data of each group of pregnant women samples are taken as training samples. Finally, the abdominal muscle relaxation degree and the upper arm muscle relaxation degree data of each training sample are taken as input parameters, and the edema index, the pre-pregnancy standard weight gain and the pre-pregnancy predicted nutritional risk evaluation value of each training sample are taken as output parameters for training.

[0268] In an embodiment of the present application, the nutritional status prediction model is implemented by using a random forest algorithm. The model input is a four-dimensional feature vector , wherein represents the pre-pregnancy predicted nutritional risk value, and the value range is , represents the upper arm muscle relaxation degree, and the value range is , and the greater the value, the higher the relaxation degree, represents the edema index, and the value range is , and the greater the value, the more serious the edema degree, represents the abdominal muscle relaxation degree, and the value range is , and the greater the value, the higher the relaxation degree. The model output is a multi-dimensional analysis result, including nutritional status score, various nutrient deficiency risk, weight gain prediction, etc.

[0269] Based on the analysis results of the nutritional status prediction model, the present application can provide comprehensive nutritional status evaluation for pregnant women, including standard weight gain curve, standard BMI growth curve, body shape development change graph, nutritional risk value, edema index and muscle relaxation degree, etc. These results provide a scientific basis for subsequent personalized guidance.

[0270] The guidance generation module 7 is connected with the nutritional status prediction module 6, and provides personalized dietary suggestions and exercise programs according to the analysis results.

[0271] As shown in Figure 7 , the guidance generation module 7 includes a dietary suggestion unit 71, an exercise program unit 72 and a monitoring feedback unit 73.

[0272] The diet recommendation unit 71 generates personalized diet recommendations based on the edema index and muscle relaxation degree in the analysis results, combined with the individual characteristics of the pregnant woman. The individual characteristics of the pregnant woman considered by the present application include age, height, weight, obesity, diabetes, hypertension, and high cholesterol, etc.

[0273] In generating the diet recommendations, the present application adopts a method combining rule base and case base. The rule base contains a series of nutrition guidance rules based on professional knowledge, such as limiting sodium intake when the edema index is higher than 0.6, and increasing protein intake when the muscle relaxation degree is higher than 0.5, etc. The case base stores a large number of diet plans of typical cases, and the system can recommend suitable plans for new users through similar case retrieval.

[0274] The content of the diet recommendations includes daily total calorie recommendation, nutrient ratio recommendation, recommended food list, food combination plan, and food taboo tips, etc. For pregnant women with high edema index, the system will recommend limiting sodium intake and increasing potassium intake, and provide specific low-sodium high-potassium food list and recipes.

[0275] The exercise plan unit 72 generates suitable exercise plans based on the analysis results and the individual characteristics of the pregnant woman. During pregnancy, appropriate exercise helps to control weight gain, improve muscle relaxation, and reduce the risk of edema.

[0276] The generation of exercise plans takes into account the following factors: gestational age stage, nutrition status evaluation results, pre-pregnancy exercise habits, physical condition limitations, etc. According to these factors, the system recommends suitable exercise types, intensity, frequency, and duration.

[0277] For pregnant women in the first trimester, less than or equal to 12 weeks, if the nutrition status is good and there are no special physical condition limitations, the system may recommend moderate-intensity aerobic exercise, such as walking, swimming, etc., 3-5 times a week, each time for 30 minutes. For pregnant women with high leg edema index, the system will specially recommend exercises that help promote lower limb blood circulation, such as gentle ankle joint activities, leg lifting exercises, etc.

[0278] The monitoring feedback unit 73 is used to monitor the nutrition and health status of the pregnant woman and adjust the diet recommendations and exercise plans. The present application emphasizes the importance of continuous monitoring and dynamic adjustment, and adjusts the guidance plans in a timely manner by regularly assessing the changes in the nutrition status of the pregnant woman.

[0279] The implementation of monitoring feedback includes regular image acquisition and analysis, diet diary recording, exercise log recording, etc. The system evaluates the effect of the guidance plans according to these data and makes necessary adjustments. If it is found that the edema index continues to rise, the system will further strengthen the recommendation to control sodium intake and increase the recommendation of exercises that promote circulation.

[0280] Through the guidance generation module 7, the present application realizes closed-loop management from evaluation to guidance, providing comprehensive and personalized nutritional health support for pregnant women.

[0281] The complete workflow of the system of the present application is as follows.

[0282] Image acquisition stage: At 18 weeks and 24 weeks of pregnancy, the pregnant woman is required to take a front full-body picture and a left side full-body picture in a standard posture, which are uploaded to the cloud server.

[0283] Image analysis stage: The differential geometry feature extraction module analyzes the image and extracts the curvature features and morphological invariants of the human body contour to generate a human body morphological feature vector.

[0284] Tissue mapping stage: The multi-scale tissue mapping module performs multi-scale hierarchical decomposition on the human body morphological feature vector, establishes a mapping relationship between the external contour and the internal tissue distribution, and calculates the human body composition data.

[0285] Parameter calculation stage: The manifold neural network analysis module calculates the abdominal muscle relaxation degree, abdominal edema index and other parameters based on the human body composition data.

[0286] Risk assessment stage: The nutritional risk assessment module inputs the current standard weight gain and the current BMI change data into the pre-pregnancy predicted nutritional risk model to output the pre-pregnancy predicted nutritional risk value.

[0287] State prediction stage: The nutritional status prediction module inputs the pre-pregnancy predicted nutritional risk value, upper arm muscle relaxation degree, edema index and abdominal muscle relaxation degree into the nutritional status prediction model to predict the nutritional status during pregnancy and obtain the analysis result.

[0288] Guidance generation stage: The guidance generation module provides personalized dietary recommendations and exercise programs according to the analysis result and the individual characteristics of the pregnant woman.

[0289] Dynamic monitoring stage: By regularly repeating the above process, the dynamic monitoring of the nutritional status during pregnancy and the timely adjustment of the guidance program are realized.

[0290] The above-described embodiments only express the specific implementation of the present application, which is described in detail and specifically, but it should not be understood as a limitation on the scope of the present patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

Claims

1. A system for image-based assessment and guidance of nutritional status during pregnancy, characterized in that, include: The image acquisition module is used to acquire full-body images of the pregnant woman in a standard posture, both from the front and from the left side, and to store the full-body images from the front and from the left side to a cloud server. The differential geometric feature extraction module, connected to the image acquisition module, is used to receive the frontal full-body image and the left side full-body image, and extract the curvature features and morphological invariants of the human body contour through surface differential feature analysis to generate a human body morphological feature vector. The multi-scale tissue mapping module, connected to the differential geometric feature extraction module, is used to perform multi-scale hierarchical decomposition on the human body morphology feature vector, dividing the human body surface into multiple scale regions. Based on the three-dimensional shape features and two-dimensional image features of each region, a mapping relationship between the external contour and the internal tissue distribution is established through a deep learning model. Human body composition data is calculated based on the volume and tissue density ratio of each region. The human body composition data includes pre-pregnancy standard weight, pre-pregnancy standard BMI, current BMI, fat percentage, muscle percentage, water percentage, and edema percentage. The manifold neural network analysis module, connected to the multi-scale tissue mapping module, is used to extract the manifold embedding coordinates of the abdominal surface based on the human body composition data through Laplacian feature mapping, and to calculate the abdominal muscle relaxation, abdominal edema index, leg muscle relaxation, leg edema index, upper arm circumference, upper arm muscle circumference, upper arm fat circumference and abdominal circumference based on the fusion of manifold convolutional neural network and abdominal surface curvature features. The nutrition risk assessment module, connected to the manifold neural network analysis module, is used to receive the body composition data, input the current standard weight gain and current BMI change data into the preconception prediction nutrition risk model, and output the preconception prediction nutrition risk value. The nutrition status prediction module is connected to the nutrition risk assessment module and the manifold neural network analysis module. It is used to input the pre-pregnancy predicted nutrition risk value, upper arm muscle relaxation, edema index and abdominal muscle relaxation into the nutrition status prediction model to predict the nutrition status during pregnancy and obtain the analysis results. The guidance generation module, connected to the nutritional status prediction module, is used to provide personalized dietary recommendations and exercise plans based on the analysis results.

2. The system according to claim 1, characterized in that, The differential geometric feature extraction module includes: A 3D modeling unit is used to perform edge detection and depth estimation on the frontal full-body image and the left side full-body image, construct a 3D human body surface model and perform parameterized representation; The curvature calculation unit, connected to the three-dimensional modeling unit, is used to calculate the average curvature and Gaussian curvature of each point on the surface based on the three-dimensional human body surface model, and generate a curvature distribution map. A geodesic network construction unit, connected to the curvature calculation unit, is used to define key points of the human body as nodes of the geodesic network, calculate the shortest geodesic distance between nodes, and form a geodesic network. The morphological invariant extraction unit, connected to the geodesic network construction unit, is used to extract morphological features independent of shooting angle and pose based on the theory of surface differential invariants, and construct a morphological feature vector.

3. The system according to claim 1, characterized in that, The multi-scale tissue mapping module includes: Multi-scale decomposition units are used to construct curvature scale space, calculate the curvature changes of human body contours at different scales, and identify tissue interfaces through extreme points in scale space. The tissue distribution density calculation unit, connected to the multi-scale decomposition unit, is used to define the tissue distribution density function, establish a mapping relationship based on curvature features and gray-level distribution, and generate a continuous density distribution function. A site-specific compensation unit, connected to the tissue distribution density calculation unit, is used to construct compensation functions for pregnancy-specific regions, thereby implementing time-varying parameter adjustment strategies for different gestational weeks.

4. The system according to claim 1, characterized in that, The manifold neural network analysis module includes: Manifold embedding layers are used to construct mapping functions from images to manifold spaces, design convolution operations that adapt to curved surface structures, and establish a unified representation space for frontal and side images; A geometric feature propagation layer, connected to the manifold embedding layer, is used to design a feature propagation mechanism based on geodesic distance to achieve non-uniform diffusion of features on the manifold. The parameter prediction layer, connected to the geometric feature propagation layer, is used to construct a multi-task learning structure to simultaneously predict body composition parameters such as fat percentage, muscle percentage, water percentage, and edema percentage.

5. The system according to claim 1, characterized in that, The multi-scale tissue mapping module divides the human body surface into three levels: coarse scale, meso scale, and fine scale. The coarse scale is divided into the head, trunk, and limb regions. The meso scale is further divided into the chest, abdomen, waist, and buttock regions within the trunk. The fine scale is further divided into the upper abdomen, periumbilical region, and lower abdomen regions within the abdomen.

6. The system according to claim 1, characterized in that, The multi-scale tissue mapping module uses a convolutional neural network to establish mapping relationships. The convolutional neural network includes 5 convolutional layers and 3 fully connected layers. By extracting the three-dimensional shape features and two-dimensional image features of each region, it outputs the fat percentage, muscle percentage and water percentage of each region.

7. The system according to claim 1, characterized in that, The manifold neural network analysis module calculates abdominal muscle relaxation through the following steps: The three-dimensional surface of the abdominal region is extracted, an adjacency graph is constructed from the sampling points on the abdominal surface, the weight of the edge is calculated using the Gaussian kernel function, the Laplacian matrix is ​​constructed and the generalized eigenvalue problem is solved to obtain the low-dimensional embedding coordinates of the manifold. A manifold convolutional neural network is constructed based on manifold embedded coordinates, including manifold convolutional layers, manifold pooling layers, and fully connected layers, to extract manifold features of the ventral surface; The average curvature of the abdominal surface is calculated as the curvature relaxation component. The neural network relaxation component is output through a manifold neural network. The texture features of the abdominal image are extracted as the texture relaxation component. The three components are weighted and fused to obtain the abdominal muscle relaxation.

8. The system according to claim 1, characterized in that, The nutrition risk assessment module obtains the preconception predictive nutrition risk model through the following steps: Obtain data on pregnant women with known nutritional risk assessment values ​​before and during pregnancy; Calculate the corresponding pre-pregnancy standard weight gain and current BMI change based on the data; The model was trained using the pre-pregnancy standard weight gain and current BMI change data for each group.

9. The system according to claim 1, characterized in that, The standard shooting posture specified in the image acquisition module is: Full-body frontal image: The subject is standing with feet on the ground, buttocks and back against the wall, eyes looking straight ahead, hands naturally perpendicular to the body in front of the body, facing the image acquisition device, eyes looking straight ahead, smiling, arms naturally extended and relaxed, at a 90° angle to the left and right, palms facing inward, and feet naturally together. Full-body image from the left side: The subject is standing with their feet on the ground, shoulders and chest facing the wall, eyes looking straight ahead, arms naturally hanging down in front of them, facing the image acquisition device, eyes looking straight ahead, smiling, arms naturally extended and relaxed, hands at a 90° angle with palms facing inward, left and right feet on the left and right sides respectively.

10. The system according to claim 1, characterized in that, The guidance generation module includes: The dietary recommendation unit is used to generate personalized dietary recommendations based on the edema index and muscle relaxation in the analysis results, combined with the individual characteristics of the pregnant woman, wherein the individual characteristics of the pregnant woman include age, height, weight, obesity, diabetes, hypertension and high cholesterol. The exercise program unit, connected to the diet advice unit, is used to generate a suitable exercise program based on the analysis results and the individual characteristics of the pregnant woman. The monitoring and feedback unit, connected to the diet advice unit and the exercise program unit, is used to monitor the nutritional health status of pregnant women and adjust the diet advice and exercise program accordingly.

Images