Scoliosis orthosis registration method based on deep learning and biomechanical constraints

Abstract

The application discloses a kind of based on deep learning and biomechanics constraint scoliosis orthosis registration method, it is related to scoliosis orthosis design technical field, including: the first feature information of three-dimensional vertebral model and the second feature information of three-dimensional body surface model are acquired respectively, the second feature information includes: geometric feature, pressure feature and motion feature;The first feature information and the second feature information are simultaneously input to the mixed configuration network pre-constructed and are rigidly registered and non-rigidly registered, the rigidly registered output initial alignment model, the non-rigidly registered output high-precision local adaptation data reasonable in anatomy;The initial alignment model and the high-precision local adaptation data reasonable in anatomy are jointly optimized to obtain final registration result. The geometric and motion characteristics of spine and body surface are extracted by double-flow feature, so as to carry out coarse rigid registration network after fine non-rigid registration, realize the leap from geometric registration to function optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of scoliosis orthosis design, in particular to a scoliosis orthosis registration method based on deep learning and biomechanical constraints. BACKGROUND

[0002] Scoliosis is a common three-dimensional spinal deformity, which refers to the deviation of the spine from the midline in the coronal plane (frontal plane), bending to the side, forming a Cobb angle greater than 10° of spinal deformity, often accompanied by rotation of the spine and changes in physiological curvature in the sagittal plane (lateral plane). It not only affects appearance, but also can affect heart and lung function, nervous system and mental health, and in severe cases, it can affect heart and lung function and even cause paralysis.

[0003] The correction method of scoliosis needs to be formulated comprehensively according to the type of scoliosis, the severity, the age and the physical condition of the patient, etc. The most commonly used correction device is the Cheneau orthosis. The Cheneau orthosis is a type of spinal lateral protrusion orthosis developed by Dr. Cheneau of France in the early 1970s based on the theory of E.G. Apothecary plaster orthosis for correcting scoliosis. The Cheneau orthosis is a full-plastic orthosis with the advantages of lightness and simplicity. It uses the "three-point force" correction principle, sets pressure pads and extension space, combines growth mechanism, breathing training and physical therapy, etc., to play the roles of anti-rotation, extension and active correction. The effectiveness of the Cheneau orthosis depends on the close fit of the orthosis to the torso and the accurate alignment of the pressure points with the scoliosis position.

[0004] However, the current orthosis design has the following technical problems: poor fit and comfort: the traditional plaster model has a long production cycle, and the general type of orthosis cannot adapt to individual differences due to fixed type, resulting in poor fit and low comfort for long-term wear. Inaccurate alignment of pressure points: although the existing three-dimensional printing technology can improve the fit, there is no automatic alignment method for the spine and torso profile, and the position of the pressure points depends on manual experience, so the correction effect is unstable. SUMMARY

[0005] In order to solve the above technical problems, the present application provides the following technical solutions:

[0006] In a first aspect, the present application provides a scoliosis orthosis registration method based on deep learning and biomechanical constraints, comprising:

[0007] Respectively acquiring first feature information of a three-dimensional spine model and second feature information of a three-dimensional body surface model, the second feature information including: geometric features, pressure features and motion features;

[0008] The first feature information and the second feature information are simultaneously input to a pre-constructed hybrid configuration network for rigid registration and non-rigid registration, the rigid registration outputs an initial alignment model, and the non-rigid registration outputs high-precision locally adapted data that is anatomically reasonable;

[0009] The initial alignment model and the high-precision locally adapted data that is anatomically reasonable are jointly optimized to obtain a final registration result.

[0010] In a possible implementation, the first feature information of the three-dimensional spine model is obtained, including:

[0011] Dynamic weight calculation and anatomical prior guidance are performed through biomechanical weighted sampling to realize feature enhancement of key regions.

[0012] Then, multi-scale abstraction is performed through hierarchical feature extraction to gradually capture the geometric, kinematic and biomechanical properties of the spine, and high-order final features are output, which take into account local geometric details and global mechanical properties.

[0013] In a possible implementation, the dynamic weight calculation and anatomical prior guidance are performed through biomechanical weighted sampling to realize feature enhancement of key regions, including:

[0014] First, the dynamic weight is calculated based on the vertebral stiffness coefficient and the distance from the main bending apex:

[0015] ;

[0016] Wherein: is the vertebral stiffness value, is the Euclidean distance from the point to the apex;

[0017] Then, improved farthest point sampling is performed, which selects the point farthest from the sampled point set and with the highest weight in each iteration, preferentially covering high-stiffness regions close to the main bending apex, and the sampling interval is encrypted in the key area and relaxed in the non-key area;

[0018] Outliers are removed in real time during sampling, and Poisson disc resampling is performed on low-density areas;

[0019] Finally, adaptive neighborhood grouping is performed, which ensures the integrity of local feature extraction during grouping, and outputs a plurality of sampling points and their neighborhood feature matrices, providing high semantic density input data for subsequent hierarchical feature aggregation.

[0020] In a possible implementation, the multi-scale abstraction is performed through hierarchical feature extraction to gradually capture the geometric, kinematic and biomechanical properties of the spine, and high-order final features are output, which take into account local geometric details and global mechanical properties, including:

[0021] The basic geometric feature extraction adopts multi-radius ball query to construct local neighborhood, calculates the eigenvalue of covariance matrix to generate curvature feature: where is the eigenvalue of the covariance matrix, and the direction-sensitive feature is extracted in the local coordinate system through the isometry convolution, and the direction attention weight is:

[0022] ;

[0023] where: The first eigenvalue of the covariance matrix, and the hyperparameter controls the sharpness of the weight distribution;

[0024] Kinematic chain modeling is introduced, and graph convolution network is used to regress and predict the relative pose parameters of adjacent vertebrae:

[0025] ;

[0026] where: denotes the homogeneous transformation matrix of the rigid motion transformation from vertebra i to vertebra j, which is used to completely describe the position and direction change of vertebra j relative to vertebra i; , denotes the 6 degrees of freedom of traversing rigid motion; denotes the multi-layer perception regression, which constrains the rotation angle of adjacent vertebrae; is the Lie algebra generator, which constitutes the complete basis of the Lie algebra;

[0027] Focus on the Cobb angle related features, project the point cloud to the Cobb angle measurement plane, calculate the projection coordinates and the distance to the main curvature line, and fuse the time series curvature change and the finite element strain distribution through the Bi-LSTM network, output the multi-dimensional bending mode features, where is the arc length of the spine;

[0028] Integrate the real-time finite element solving layer to calculate the node stress and generate the regularization gradient, and ensure the biomechanical rationality;

[0029] Finally, through trilinear interpolation upsampling and cross-level residual connection, the inverse distance weighted interpolation is used for upsampling, and the high-order final features are output, which take into account the local geometric details and global mechanical characteristics:

[0030] ;

[0031] where: is the multi-layer perception, which realizes the nonlinear fusion and dimension reduction of cross-level features; and respectively represent the first layer feature and the fourth layer feature extracted by hierarchical feature extraction on a three-dimensional spine model; represents up-sampling, represents channel splicing.

[0032] In a possible implementation, the second feature information of the three-dimensional body surface model is obtained, including:

[0033] The three-dimensional body surface model data is preprocessed;

[0034] The dynamic graph is constructed using the preprocessed three-dimensional body surface model data, and the hierarchical abstraction strategy is used to realize the layer-by-layer capture of the complex geometric features of the body surface.

[0035] The geometric features, pressure features and motion features are obtained by combining the low-level features of the first layer and the high-level features of the last layer in the dynamic graph construction.

[0036] In a possible implementation, the three-dimensional body surface model data is preprocessed, including:

[0037] First, bilateral filtering is used to suppress noise, and isolated noise points are removed by statistical outlier removal;

[0038] Then, Poisson disc sampling is performed to balance the point cloud density, and details are retained in high-curvature areas, and local encryption sampling is performed in low-density areas;

[0039] Finally, a pre-trained posture-pressure simulation model is used to generate a normalized pressure heat map, and the pressure value is fused as an additional channel to the point cloud to form a multi-dimensional matrix, providing geometric and biomechanical dual-mode input for subsequent dynamic graph convolution.

[0040] In a possible implementation, the dynamic graph is constructed using the preprocessed three-dimensional body surface model data, and the hierarchical abstraction strategy is used to realize the layer-by-layer capture of the complex geometric features of the body surface, including:

[0041] The first layer constructs a k-NN graph with original coordinates, calculates initial edge features in a neighborhood with a preset radius, and extracts large-scale geometric features such as local curvature;

[0042] The second layer searches for neighbors in the feature space, calculates dual-mode edge features of three-dimensional coordinate difference and feature difference, and focuses on medium-scale structures in the scapula and iliac crest;

[0043] The third layer further shrinks the neighborhood range in the high-dimensional feature space, combines the dynamically updated pressure-sensitive weight to filter adjacent points, and focuses on capturing the microscopic features of the millimeter-level high-stress areas such as the spinous process and the spinous process.

[0044] In one possible implementation, each layer contains edge feature calculation, which realizes high-precision feature modeling by dynamically capturing local geometric relationship and biomechanical differences, including:

[0045] For each center point and its neighborhood points obtained by dynamic k-NN search Concatenate the multi-dimensional input vector, which contains the center point coordinates , the relative coordinate difference of neighborhood points , and the feature difference , to form a joint representation that combines geometry and semantics:

[0046] ;

[0047] Then input the concatenated vector into a three-layer MLP for nonlinear transformation to extract high-order interaction features.

[0048] Finally, perform channel dimension max-pooling on the edge features of the k neighborhood points to retain the most discriminative local patterns and suppress noise interference:

[0049] ;

[0050] If the input contains pressure values, concatenate the normalized pressure difference to enhance biomechanical perception.

[0051] In one possible implementation, the feature aggregation combining the low-level features of the first layer and the high-level features of the last layer in the dynamic graph construction obtains the geometric features, pressure features, and motion features, including:

[0052] Upsample the high-level features output by the third layer to the resolution of the first layer by tri-linear interpolation, and at the same time, upgrade the low-level features of the first layer to the same dimension as the high-level features by 1x1 convolution;

[0053] The high-level features and the low-level features are concatenated and then passed through a Sigmoid gating function to generate dynamic weights , which are used to weight the high-level and low-level features;

[0054] Subsequently, apply a pre-trained pressure mask to biomechanically enhance the fused features, amplify the feature values in the high-pressure area by a first preset multiple, and attenuate the feature values in the low-pressure area to a second preset multiple;

[0055] Finally, output the geometric features , pressure features , and motion features through three independent fully connected layers.

[0056] In a possible implementation, the step of inputting the first feature information and the second feature information into a pre-constructed hybrid configuration network to perform rigid registration and non-rigid registration includes:

[0057] The second feature information is fused in a channel to form a unified representation, and a full connection layer is used to compress the dimension and reserve key information, for later finite element simulation verification.

[0058] In the rigid registration, anatomical landmark matching and global transformation solving are performed to realize initial alignment of the spine and the body surface.

[0059] In the non-rigid registration, conditional thin-plate spline deformation field prediction and biomechanical constraint optimization are performed to realize local anatomical adaptation.

[0060] In a possible implementation, the step of performing rigid registration to realize initial alignment of the spine and the body surface includes:

[0061] First, feature matching is performed based on the geometric characteristics of the spine skeleton and the geometric characteristics of the body surface Cosine similarity is used to screen out high-confidence corresponding point pairs.

[0062] Then, an optimal rigid transformation matrix is calculated by using the Kabsch algorithm to solve a rotation matrix and a translation vector , which minimizes the mean square error of the corresponding point pairs:

[0063] ;

[0064] Wherein: is the coordinates of the i-th anatomical landmark extracted from the three-dimensional spine skeleton model; is the coordinates of the j-th corresponding feature point extracted from the three-dimensional body surface model;

[0065] Subsequently, the RANSAC algorithm is used to remove abnormal matching, and the global transformation parameters are obtained after iterative optimization.

[0066] Finally, biomechanical constraints are added, and the constraints are integrated into the loss function by using the Lagrange multiplier method to ensure that the registration result meets the anatomical rationality, and an initial alignment model is output.

[0067] In a possible implementation, the step of performing non-rigid registration to realize local anatomical adaptation includes:

[0068] The fused body surface pressure feature and the motion feature are input, and a conditional TPS network is used to predict a control point set and the corresponding weights , generating point-wise displacement field:

[0069] ;

[0070] where: is the radial basis function, denotes the Euclidean distance between control points and target points ; affine transformation matrix and translation vector are network prediction parameters, matrix handles global linear deformation, vector compensates for overall offset;

[0071] Joint finite element simulation verification, mapping the deformation field to the body surface soft tissue mesh, realizing anatomically reasonable high-precision local adaptation.

[0072] In a possible implementation, the joint optimization of the initial alignment model and the anatomically reasonable high-precision local adaptation data obtains a final registration result, comprising:

[0073] By calculating the Chamfer distance loss of non-rigid registration of the bone-surface anatomical landmark point pair, the global anatomical alignment of the bone and the surface is ensured, and the initial coordinate uniformity is established by matching the key landmark points:

[0074] ;

[0075] where: is the i-th anatomical landmark point coordinate of the spinal skeleton; is the j-th corresponding landmark point coordinate of the surface scan; is a rigid transformation that maps the surface point to the spinal coordinate; hyperparameter is the number of matched landmark point pairs; Calculate the nearest neighbor matching;

[0076] Realize local anatomical adaptation, balance the thin plate spline energy non-rigid registration loss function of deformation smoothness and high pressure area rigidity constraint, punish the second derivative of the deformation field, and ensure geometric continuity:

[0077] ;

[0078] where: is the deformation displacement vector of the i-th surface vertex; hyperparameter is the number of surface vertices participating in deformation; hyperparameter is the smoothing item weight, balancing the deformation strength and geometric continuity; is the spatial gradient of the deformation field, reflecting the local distortion degree;

[0079] The von Mises stress constraint is used to calculate the finite element physical loss so as to limit the stress of the soft tissue unit to not exceed the safety threshold:

[0080] ;

[0081] in: For the The von Mises stress of each soft tissue unit is calculated by finite element simulation to reflect the biomechanical load; hyperparameters is the total number of soft tissue units, which is determined by the volume mesh discretization; 2.0 is the upper limit threshold of the safety stress;

[0082] Finally, the total loss function is determined as:

[0083] .

[0084] In the embodiments of the present application, in order to address the core problem that the positioning of the compression points of the spinal orthosis relies on manual experience and lacks biomechanical rationality, the geometric and motion characteristics of the spine and the body surface are extracted through dual-stream features, thereby performing a coarse rigid followed by a fine non-rigid alignment network, achieving a leap from geometric alignment to functional optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 A schematic diagram of a scoliosis brace registration method based on deep learning and biomechanical constraints provided in an embodiment of the present application;

[0086] Figure 2 A schematic diagram of the feature extraction process of a three-dimensional spine model provided in an embodiment of the present application;

[0087] Figure 3 A schematic diagram of a feature extraction process for a three-dimensional body surface model provided in an embodiment of the present application;

[0088] Figure 4 This is a visualization result diagram of the registration provided in the embodiment of this application. DETAILED DESCRIPTION

[0089] The present invention will be described below with reference to the accompanying drawings and specific implementation methods.

[0090] See also Figure 1 The scoliosis brace registration method based on deep learning and biomechanical constraints provided in this embodiment includes:

[0091] S101, respectively obtaining first characteristic information of a three-dimensional spine model and second characteristic information of a three-dimensional body surface model.

[0092] In this embodiment, the scoliosis orthosis design database contains complete three-dimensional data of 1,000 patients. The body surface model of each patient is obtained through high-precision 3D scanning, and contains a triangular mesh consisting of approximately 50,000-100,000 vertices. The three-dimensional coordinates of three key compression points are specially marked ( ), these compression points are determined by clinical experts based on the "three-point force" correction principle, corresponding to the surface projection position of the top vertebra of scoliosis and the two counterforce points; at the same time, the database stores the corresponding spinal CT scan data for each patient in the DICOM standard format, including continuous cross-sectional images from C1 to the sacrum (layer thickness 1mm), and generates an enhanced spinal model containing vertebral segmentation labels (using the AOSpine standard) and biomechanical parameters (such as vertebral stiffness coefficient) through three-dimensional reconstruction. All data are strictly de-identified and spatially aligned to ensure the anatomical consistency between the surface model and the spinal CT in a unified coordinate system, with the error controlled within ±1.5mm. First, the three-dimensional spinal model data is preprocessed. After the input three-dimensional spinal X-ray is segmented and three-dimensionally reconstructed, the reconstructed three-dimensional spinal model is input into the feature extractor, containing approximately 50,000-100,000 vertices, each of which records floating-point three-dimensional coordinates ( , precision ) and motion constraint parameters defined based on the AOSpine clinical guidelines (such as rotational freedom of ±5° in the sagittal plane and ±3° in the coronal plane). Data validation was then performed: topological analysis was used to check the continuity of the vertebral sequence, and the threshold for the distance between adjacent thoracic vertebrae was set at 8-12. , lumbar vertebrae 12-15 , the fracture area triggers the hole repair based on radial basis function; at the same time, Outlier Detection ( ) automatically marks and isolates coordinate abnormal points. In the anatomical coordinate system normalization stage, the sacral promontory (the midpoint of the upper edge of S1) is used as the origin and the pelvic statistical shape template is aligned using the iterative closest point (ICP) algorithm (positioning error ), and an orthogonal coordinate system is constructed: the Z axis is determined by the B-spline fitting main curve line from the C7 spinous process vertex to the midpoint of the sacral promontory, the X axis is the direction of the line connecting the centers of the bilateral acetabulum, and the Y axis is determined by the right-hand rule ( ) is generated, and finally the original coordinates are transformed into the normalized space (target registration error) through the rotation matrix and translation vector In the biomechanical parameter fusion stage, based on the preset stiffness table (such as 、 ) assigns a baseline stiffness value to each vertebra and adds a dynamic stiffness value to each vertex through point-level refinement (taking into account the stiffness gradient from cortical bone to trabecular bone) ( ), while generating 24x3 Boolean matrix encoding vertebral degree-of-freedom constraints (e.g. thoracic vertebrae rotation locked in sagittal plane). The output data contains geometry coordinates, 24-dimensional vertebral labels and stiffness values augmented point cloud (N x 28 matrix) and motion constraint table, and quality control is completed through coordinate alignment error ( ), label mapping accuracy ( ) and stiffness gradient smoothness (slope ), providing standardized input for subsequent spinal orthopedic planning and biomechanically weighted sampling.

[0093] To ensure that the orthosis pressure distribution modeling prioritizes biomechanically sensitive areas (e.g. apex vertebrae, end vertebrae), false negatives caused by stress concentration are avoided.

[0094] Referring to Figure 2 , in this embodiment, to obtain the first feature information of the three-dimensional spinal vertebra model, the following steps are performed: dynamic weight calculation and anatomical prior guidance are performed through biomechanically weighted sampling to realize feature enhancement of key areas. Then, multi-scale abstraction is performed through hierarchical feature extraction to gradually capture the geometric, kinematic and biomechanical properties of the spine, and high-order final features are output that take into account local geometric details and global mechanical properties.

[0095] Biomechanically weighted sampling realizes feature enhancement of key areas through dynamic weight calculation and anatomical prior guidance. The specific steps include: first, calculate the dynamic weight based on the vertebral stiffness coefficient ( to ) and the distance from the apex vertebra:

[0096] ;

[0097] wherein: is the vertebral stiffness value, is the Euclidean distance from the point to the apex vertebra (unit: mm), the weight of the apex vertebra area is increased , and the weight of non-key areas is reduced Improved farthest point sampling (FPS) is then performed, which selects the point with the highest weight and farthest distance to the sampled point set in each iteration, giving priority to high stiffness regions near the apex of the main curve. The sampling interval is tightened to 3 mm in critical areas and relaxed to 8 mm in non-critical areas. Outliers are removed in real time during sampling (neighborhood density < 10 points / 20 mm3), and Poisson disc resampling is performed on low-density areas (such as the pedicle) to achieve a target density of 2 points / mm2. Finally, adaptive neighborhood grouping is performed: a 10 mm radius ball query is used in the top vertebra area, and the radius is expanded to 20 mm in the lumbar-sacral area. If the number of neighborhood points is less than 32, the radius is gradually increased to 1.5 times and virtual points are added (generated by normal interpolation) to ensure the integrity of local feature extraction. The output is a 1024-point sampling point and its neighborhood feature matrix (32 neighborhood points x 64 dimensions), providing high semantic density input data for subsequent hierarchical feature aggregation.

[0098] Hierarchical feature extraction captures the geometric, kinematic, and biomechanical properties of the spine through multi-scale abstraction, divided into 4 levels:

[0099] Level 1 stage (1024 points → 64 dimensions) is basic geometric feature extraction, which uses multi-radius ball queries (10 mm / 20 mm / 30 mm) to construct local neighborhoods, calculates the eigenvalues of the covariance matrix to generate curvature features ( , where is the eigenvalue of the covariance matrix), and extracts direction-sensitive features in the local coordinate system (tangent direction determined by the spinous process connecting line) through SO(3) equivariant convolution, with direction attention weight:

[0100] ;

[0101] where: the th eigenvalue of the covariance matrix , and the hyperparameter controls the sharpness of the weight distribution.

[0102] Level 2 stage (512 points → 128 dimensions) introduces kinematic chain modeling, using a graph convolution network (GCN) to regress and predict the relative pose parameters (6 degrees of freedom) of adjacent vertebrae:

[0103] ;

[0104] where: denotes the homogeneous transformation matrix of the rigid body motion transformation from vertebra i to vertebra j, which fully describes the position and direction changes of vertebra j relative to vertebra i; denotes the 6 degrees of freedom of the rigid body motion; denotes a multi-layer perceptron regression that constrains the rotation angle of adjacent vertebrae; For the generators of Lie algebra, the complete basis of Lie algebra, any rigid body motion can be expressed as a linear combination of these bases.

[0105] Level 3 stage (256 points → 256 dimensions) focuses on Cobb angle related features, projects point cloud to Cobb angle measurement plane (determined by the center point of the end vertebra and the normal of the top vertebra), and calculates the projection coordinates Distance to the main bending curve . Through the Bi-LSTM network, the time series curvature change (S(t)) and the finite element strain distribution are fused, and the 256-dimensional bending mode features are output. , Arc length of the spine.

[0106] Level 4 stage (128 points → 512 dimensions) integrates a real-time finite element solving layer to calculate node stress (σ(t)) and generate normalized gradients to ensure biomechanical reasonableness.

[0107] Finally, through trilinear interpolation upsampling and cross-level residual connection:

[0108] ;

[0109] Among them: is a multi-layer perceptron that realizes nonlinear fusion and dimensionality reduction of cross-level features; and represent the first and fourth layer features extracted from the three-dimensional spine model respectively; represents upsampling, which increases the resolution of deep features to match the shallow layer. represents channel splicing, which splices different layer features in the channel dimension. Upsampling uses inverse distance weighted interpolation (IDW). The output is a 512-dimensional high-order final feature that balances local geometric details (error ≤0.7mm) and global mechanical properties (stress error ≤4.8%).

[0110] Referring to Figure 3 , the second feature information of the three-dimensional body surface model is obtained, including: preprocessing the three-dimensional body surface model data. Using the preprocessed three-dimensional body surface model data to construct a dynamic graph, and capturing the complex geometric features of the body surface layer by layer through feature space adaptive neighborhood search and hierarchical abstraction strategy. Combine the low-level features of the first layer and the high-level features of the last layer in the dynamic graph construction to obtain the geometric features, pressure features and motion features.

[0111] Unlike the three-dimensional spine model, the input data for body surface point cloud feature extraction is a matrix (X) ​Coordinate), initial density about 1-2 points / mm2, and auxiliary labeling of pressure-sensitive area label map (key areas such as iliac crest, scapula, sternum, etc.), format is binary mask (0 / 1). In the preprocessing stage, first, bilateral filtering is used:

[0112] ;

[0113] (spatial domain , normal domain ) to suppress noise, and isolated noise points are eliminated by statistical outlier removal (radius 5 , number of points in the neighborhood <10); then, Poisson disc sampling (target density 1 point / 2) is performed to balance the point cloud density, and details are preserved in high-curvature areas (curvature >0.1), and local dense sampling (4 points / 2) is performed in low-density areas (such as the abdomen); finally, a pre-trained posture-pressure simulation model is used to generate a normalized pressure heat map (0-1 interval), and the pressure value is fused as an additional channel to the point cloud, forming a matrix , pressure value), providing geometric and biomechanical dual-modal input for subsequent dynamic graph convolution.

[0114] DGCNN solves the problem of static neighborhood limitation and uniform feature response in traditional methods for processing body surface point clouds through dynamic graph construction. The dynamic graph construction realizes the layer-by-layer capture of complex geometric features on the body surface through feature space adaptive neighborhood search and hierarchical abstraction strategy:

[0115] The first layer constructs a k-NN graph (k=20) with the original coordinates, calculates the initial edge features within a neighborhood with a radius of about 30 mm, and extracts large-scale geometric features (64 dimensions) such as local curvature through the EdgeConv module; the second layer searches for neighbors in the feature space (64 dimensions) (k=15), calculates the dual-modal edge features of the three-dimensional coordinate difference and the feature difference, and focuses on medium-scale structures such as the scapula and iliac crest (output 128 dimensions); the third layer further shrinks the neighborhood range (k=10) in the 128-dimensional feature space, combines the dynamically updated pressure-sensitive weight to filter the adjacent points, and focuses on capturing the microscopic features of the millimeter-level high-stress areas such as the spinous process and the bone spur (output 256 dimensions). Each layer uses the FAISS library to accelerate k-NN search, and dynamically reconstructs the neighborhood relationship in the feature space (non-fixed coordinate space), so that the network can adapt to the geometric complexity of different anatomical regions - such as using sparse connections in flat areas of the torso, and automatically encrypting connections in high-curvature bone spur areas, which improves the feature coverage rate of key areas by 35% and reduces the redundant calculation amount by 42% compared with traditional fixed graph construction methods.

[0116] Each level contains edge feature computation, which achieves high-precision feature modeling by dynamically capturing local geometric relationships and biomechanical differences: first, for each center point and its neighborhood points obtained by dynamic k-NN search (based on the current feature space rather than a fixed coordinate space), a multi-dimensional input vector is spliced, containing the center point coordinates , neighborhood point relative coordinate differences , and feature differences (the total dimension of the input is 70 when the feature dimension is 70), forming a joint representation that integrates geometry and semantics:

[0117] ;

[0118] Then, the spliced vector is input into a three-layer MLP (Multi-Layer Perceptron) for nonlinear transformation to extract high-order interaction features; subsequently, the edge features of the k neighborhood points are subjected to channel dimension maximum pooling:

[0119] ;

[0120] Among them: indicates that the most discriminative features are selected through a competitive mechanism to improve the sensitivity of the model to key anatomical areas, is the joint representation of the integrated geometry and semantics obtained above.

[0121] The most discriminative local patterns (such as steep curvature or pressure sudden change areas) are retained, and noise interference is suppressed; if the input contains pressure values, additional normalized pressure differences are spliced to enhance biomechanical perception. This process is iteratively updated layer by layer, allowing the network to adapt to the neighborhood relationships at different abstraction levels - the bottom layer captures macro body posture (such as trunk bending), the middle layer identifies mesoscale structures (such as scapula concave-convex), and the high layer focuses on microscopic features (such as bone protrusion curvature), finally outputting 256 high-level features.

[0122] The final feature aggregation is achieved through cross-level gated fusion and biomechanical weighting: first, the 256-dimensional high-level features output by the third layer are upsampled to the first layer resolution (1024 points) through tri-linear interpolation, while the 64-dimensional low-level features of the first layer are upgraded to 256 dimensions through 1x1 convolution. After splicing, the Sigmoid gating function generates dynamic weights , which are used to weight the high and low level features (retaining local details and global semantics); then, a pre-trained pressure mask is applied to biomechanically enhance the fused features, with high pressure areas (such as iliac crest) having feature values amplified by 1.5 times and low pressure areas (such as abdomen) attenuated to 0.5 times; finally, three independent fully connected layers are used to decouple the output - geometric features (256 dimensions, including curvature / symmetry), pressure features (128 dimensions, encoding pressure gradient), motion features (64 dimensions, predicting deformation trend), while ensuring anatomical fidelity (error <0.2mm), the discriminative ability of pressure-sensitive regions is improved by 2.3 times, providing multi-dimensional biomechanical characterization for downstream tasks.

[0123] S102, input the first feature information and the second feature information into the pre-constructed hybrid configuration network for rigid registration and non-rigid registration, the rigid registration outputs an initial alignment model, and the non-rigid registration outputs high-precision local adaptation data with reasonable anatomy.

[0124] The registration process in this embodiment is specifically divided into rigid registration and non-rigid registration:

[0125] Before entering the registration stage, first, three body surface features are fused into a unified representation in a channel:

[0126] ;

[0127] And compressed to 256 dimensions through a fully connected layer, retaining key information for later finite element simulation verification.

[0128] The first step is rigid registration, which realizes the initial alignment of the spine and the body surface through anatomical landmark point matching and global transformation solution. First, based on the geometric features of the spine (such as C7 spinous process, sacral promontory) and the geometric features of the body surface (such as trunk midline, scapula outline), feature matching is performed, and high-confidence corresponding point pairs (such as the projection point of the sacral promontory on the body surface) are selected using cosine similarity; then the optimal rigid transformation matrix is calculated through Kabsch algorithm - solve the rotation matrix and translation vector , minimizing the mean square error of the corresponding point pairs:

[0129] ;

[0130] Among them: is the coordinates of the i-th anatomical landmark point extracted from the three-dimensional spine model; is the coordinates of the j-th corresponding feature point extracted from the three-dimensional body surface model.

[0131] Subsequently, the RANSAC algorithm is used to remove abnormal matches (such as mis-matching points in soft tissue regions), and the global transformation parameters are obtained after iterative optimization; finally, biomechanical constraints (such as the limitation of the rotation degree of freedom of the thoracic vertebrae ± 5°) are added, and the constraints are integrated into the loss function through the Lagrange multiplier method to ensure that the registration result meets the anatomical rationality (such as the distance error between the sacral promontory projection point and the pelvic landmark point on the body surface ≤1.5mm). The initial alignment model is output in this step, which provides a high-precision starting point for subsequent non-rigid registration, and the key indicators include the target registration error (TRE≤1.2mm) and the Dice coefficient of anatomical landmarks (≥0.95).

[0132] Non-rigid registration achieves local anatomical adaptation through conditional thin plate spline (TPS) deformation field prediction and biomechanical constraint optimization. The specific process is as follows: first, the fused body surface pressure features (128D) and motion features (64D) are input, and the control point set and the corresponding weight are predicted by the conditional TPS network to generate a point-by-point displacement field:

[0133] ;

[0134] wherein: is a radial basis function, represents the Euclidean distance between the control point and the target point ; the affine transformation matrix and the translation vector are network prediction parameters, the matrix handles global linear deformation, and the vector compensates for the overall offset. Finally, joint finite element simulation verification is performed to map the deformation field to the body surface soft tissue mesh, achieving anatomically reasonable high-precision local adaptation.

[0135] S103, the initial alignment model and the anatomically reasonable high-precision local adaptation data are jointly optimized to obtain the final registration result.

[0136] The final joint optimization is completed through the synergy of the multi-objective loss function so as to achieve the final three-dimensional registration effect. The core is the mixed loss function that integrates rigid registration, non-rigid deformation, and biomechanical constraints. Since the proposed model is trained with an end-to-end learning strategy, the transmission module and the deformation module are updated together.

[0137] First, the loss of non-rigid registration by calculating the Chamfer distance between the bone-body surface anatomical landmark point pairs is calculated to ensure the global anatomical alignment of the bone and the body surface, and the initial coordinate uniformity is established by matching the key landmark points:

[0138] ;

[0139] in: is the coordinate of the i-th anatomical landmark point of the spine; is the coordinate of the jth corresponding landmark point scanned on the body surface; is a rigid transformation that maps body surface points to spinal coordinates; hyperparameters is the number of matching landmark pairs; Compute nearest neighbor matching.

[0140] Secondly, in order to achieve local anatomical adaptation, a thin plate spline energy non-rigid registration loss function is used to balance deformation smoothness (avoiding sudden changes) and rigid constraints in high-pressure areas (reducing orthosis compression deformation). The second-order derivative of the deformation field is penalized to ensure geometric continuity:

[0141] ;

[0142] in: The deformation displacement vector of the vertex of the i-th body table; hyperparameter is the number of surface vertices involved in deformation; hyperparameter is the weight of the smoothness term, balancing deformation strength and geometric continuity; is the spatial gradient of the deformation field, reflecting the degree of local distortion.

[0143] Then, to ensure the physical rationality of the registration results and avoid soft tissue stress exceeding the limit or bone-interface force discontinuity, the von Mises stress constraint calculation is used to limit the soft tissue unit stress to a safety threshold (2MPa) and prevent finite element physical losses in the finite element simulation:

[0144] ;

[0145] in: For the The von Mises stress of each soft tissue unit is calculated by finite element simulation to reflect the biomechanical load; hyperparameters is the total number of soft tissue units, which is determined by the volume mesh discretization; 2.0 is the upper limit threshold of the safety stress;

[0146] That is, the total loss function is as follows:

[0147] .

[0148] In order to verify the method proposed in this embodiment, the following experiments were also conducted in this embodiment:

[0149] Table 1 Comparison of registration model performance

[0150]

[0151] As shown in Table 1, this experiment compared the performance of seven mainstream registration models on the same dataset. The evaluation indicators include three core dimensions: anatomical structure alignment accuracy (Dice similarity coefficient ( ), 95% Hausdorff distance ), topological protection capability (the proportion of voxels with Jacobian determinant of deformation field ≤ 0 ( ), as well as computational efficiency (parameter size and single-pass registration time / s). To ensure a fair and thorough evaluation of our model (Ours), we selected a series of baseline methods for comparison: VoxelMorph, a purely CNN-based approach, and OTMorph and Fau-Net, the latest models for medical image registration published in IEEE Transactions on Medical Imaging (TMI) in 2025.

[0152] Table 2 Registration visualization results

[0153]

[0154] As shown in Table 2, our registration model takes the 3D body surface model as input With 3D spine model The resulting displacement field ,use 3D spine model Perform spatial transformation and align with the body surface model. The results are displayed in front view and side view. Set it to semi-transparent to better distinguish the degree of alignment of the observed spine embedded in the body surface.

[0155] See also Figure 4 , Figure 4 (a) is a three-dimensional surface model , Figure 4 (b) is a three-dimensional spine model , Figure 4 Middle (c) is the front view of the registration result, which shows the degree of matching between the spinal midline and the spinal groove on the body surface. Figure 4 (d) shows the side view of the registration result, demonstrating how the spinal curvature (such as lumbar lordosis and thoracic kyphosis) aligns with the body contour. This verifies that the registered spinal depth is appropriate (e.g., no penetration or abnormal deviation).

[0156] In the embodiments of the present application, "at least one" refers to one or more, and "multiple" refers to two or two or more. The "and / or" describes the association relationship of the associated objects, which means that there can be three kinds of relationships, for example, A and / or B, which can represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. Wherein A, B can be singular or plural. The character " / " generally represents that the associated objects before and after are in an "or" relationship. "At least one of the following" and the like means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b and c can represent: a, b, c, a-b, a-c, b-c, or a-b-c, where a, b, c can be single or multiple.

[0157] The above description is merely a specific implementation of the present application. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. The protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A scoliosis brace registration method based on deep learning and biomechanical constraints, characterized by: include: respectively acquiring first characteristic information of the three-dimensional spine model and second characteristic information of the three-dimensional body surface model, wherein the second characteristic information includes: geometric characteristics, pressure characteristics, and motion characteristics; Inputting the first feature information and the second feature information into a pre-built hybrid configuration network simultaneously for rigid registration and non-rigid registration, wherein the rigid registration outputs an initial alignment model, and the non-rigid registration outputs anatomically reasonable high-precision local adaptation data; Jointly optimizing the initial alignment model and the anatomically reasonable high-precision local adaptation data to obtain a final registration result; The method of jointly optimizing the initial alignment model and the anatomically reasonable high-precision local adaptation data to obtain a final registration result includes: The loss of non-rigid registration is achieved by calculating the Chamfer distance between the bone and body surface anatomical landmarks to ensure global anatomical alignment between the skeleton and the body surface, and the initial coordinate uniformity is established by matching key landmarks: ； in: is the coordinate of the i-th anatomical landmark point of the spine; is the coordinate of the jth corresponding landmark point scanned on the body surface; is a rigid transformation that maps body surface points to spinal coordinates; hyperparameters is the number of matching landmark pairs; Compute nearest neighbor matching; A thin plate spline energy non-rigid registration loss function that achieves local anatomical adaptation and balances deformation smoothness with rigid constraints in high-pressure areas, penalizing the second-order derivatives of the deformation field to ensure geometric continuity: ； in: No. Deformation displacement vector of individual table vertices; hyperparameters is the number of surface vertices involved in deformation; hyperparameter is the weight of the smoothness term, balancing deformation strength and geometric continuity; is the spatial gradient of the deformation field, reflecting the degree of local distortion; The von Mises stress constraint is used to calculate the finite element physical loss so as to limit the stress of the soft tissue unit to not exceed the safety threshold: ； in: For the The von Mises stress of each soft tissue unit is calculated by finite element simulation to reflect the biomechanical load; hyperparameters is the total number of soft tissue units, which is determined by the volume mesh discretization; 2.0 is the upper limit threshold of the safety stress; The final total loss function is determined as: .

2. The scoliosis brace registration method based on deep learning and biomechanical constraints according to claim 1, characterized in that: The obtaining of first characteristic information of the three-dimensional spine model includes: Dynamic weight calculation and anatomical prior guidance are performed through biomechanical weighted sampling to achieve feature enhancement in key areas; Then, multi-scale abstraction is performed through hierarchical feature extraction to gradually capture the geometric, kinematic and biomechanical characteristics of the spine, and output high-order final features that take into account both local geometric details and global mechanical properties.

3. The scoliosis brace registration method based on deep learning and biomechanical constraints according to claim 2, characterized in that: The dynamic weight calculation and anatomical prior guidance through biomechanical weighted sampling are used to enhance the features of key areas, including: First, the dynamic weight is calculated based on the vertebral stiffness coefficient and the distance between the main curve and the top vertebra: ； in: is the vertebral body stiffness value, is the Euclidean distance from the point to the vertex; Then, an improved farthest point sampling is performed. In each iteration, the point farthest from the sampled point set and with the highest weight is selected, giving priority to covering areas with high stiffness and close to the main bend apex. The sampling interval is denser in critical areas and wider in non-critical areas. During the sampling process, outliers are removed in real time, and Poisson disk resampling is performed on low-density areas; Finally, adaptive neighborhood grouping is performed to ensure the integrity of local feature extraction during grouping. The output is a matrix of multiple sampling points and their neighborhood features, providing high-semantic-density input data for subsequent hierarchical feature aggregation.

4. The scoliosis brace registration method based on deep learning and biomechanical constraints according to claim 2 or 3, characterized in that: The method then uses hierarchical feature extraction to perform multi-scale abstraction to gradually capture the geometric, kinematic, and biomechanical properties of the spine, outputting high-order final features that take into account both local geometric details and global mechanical properties, including: Basic geometric feature extraction uses multi-radius sphere query to construct a local neighborhood and calculates the eigenvalues ​​of the covariance matrix to generate curvature features: ,in is the eigenvalue of the covariance matrix, and the direction-sensitive features are extracted in the local coordinate system through equivariant convolution. The direction attention weight is: ； in: The covariance matrix features, among which , hyperparameters Control the sharpness of weight distribution; Kinematic chain modeling is introduced, and graph convolutional network regression is used to predict the relative posture parameters of adjacent vertebrae: ； in: From the vertebral body to the vertebral body The homogeneous transformation matrix of the rigid body motion transformation is used to fully describe the vertebral body Relative to the vertebral body changes in position and orientation; , represents the six degrees of freedom of traversal rigid body motion; Represents multi-layer perceptron regression, constraining the rotation angles of adjacent vertebrae; are Lie algebra generators, forming a complete basis of Lie algebra; Focus on the Cobb angle related features, project the point cloud to the Cobb angle measurement plane, and calculate the projection coordinates Distance to the main curve ,By fusing the temporal curvature change and the finite element strain distribution through the Bi-LSTM network,,the multi-dimensional bending mode features are output; Integrate a real-time finite element solution layer to calculate nodal stresses and generate regularized gradients to ensure biomechanical rationality; Finally, through trilinear interpolation upsampling and cross-level residual connections, upsampling uses inverse distance weighted interpolation to output high-order final features that take into account both local geometric details and global mechanical properties: ； in: A multi-layer perceptron that realizes nonlinear fusion and dimensionality reduction of cross-level features; and They represent the first-layer features and the fourth-layer features extracted from the hierarchical features of the three-dimensional spine model; represents upsampling, Indicates channel splicing.

5. The scoliosis brace registration method based on deep learning and biomechanical constraints according to claim 1, characterized in that: Obtaining second characteristic information of the three-dimensional body surface model, including: Preprocessing the three-dimensional body surface model data; The pre-processed 3D body surface model data is used to construct a dynamic graph, and the complex geometric features of the body surface are captured layer by layer through feature space adaptive neighborhood search and hierarchical abstraction strategy. Combining the low-level features of the first layer and the high-level features of the last layer in the dynamic graph construction to perform feature aggregation to obtain the geometric features, pressure features and motion features; Preprocessing of 3D body surface model data, including: Firstly, bilateral filtering is used to suppress noise, and isolated noise points are eliminated by statistical outlier removal. Then, Poisson disk sampling is performed to balance the point cloud density, retain details in high curvature areas, and perform local dense sampling in low-density areas; Finally, a normalized pressure heat map is generated through the pre-trained posture-pressure simulation model, and the pressure value is fused into the point cloud as an additional channel to form a multi-dimensional matrix to provide geometric and biomechanical dual-modal input for subsequent dynamic graph convolution.

6. The scoliosis brace registration method based on deep learning and biomechanical constraints according to claim 5, characterized in that: The method of constructing a dynamic graph using pre-processed 3D body surface model data and capturing layer-by-layer complex geometric features of the body surface through feature space adaptive neighborhood search and hierarchical abstraction strategy includes: The first layer constructs a k-NN graph with the original coordinates, calculates the initial edge features within the neighborhood of the preset radius, and extracts the large-scale geometric features of the local curvature; The second layer re-searches the nearest neighbors in the feature space, calculates the bimodal edge features of the three-dimensional coordinate difference and the feature difference, and focuses on the medium-scale structures of the scapula and iliac crest; The third layer further shrinks the neighborhood range in the high-dimensional feature space, and combines the dynamically updated pressure-sensitive weights to screen adjacent points, focusing on capturing the microscopic features of millimeter-level high-stress areas of bony protrusions and spinous processes.

7. The scoliosis brace registration method based on deep learning and biomechanical constraints according to claim 6, characterized in that: Each layer includes edge feature calculations, dynamically capturing local geometric relationships and biomechanical differences to achieve high-precision feature modeling, including: For each center point and the neighborhood points obtained by dynamic k-NN search Splice multidimensional input vectors, the multidimensional input vectors contain center point coordinates , relative coordinate difference of neighboring points and characteristic difference , forming a joint representation that integrates geometry and semantics: ； The concatenated vector is then fed into a three-layer MLP for nonlinear transformation to extract high-order interaction features; The last pair The edge features of the neighborhood points are pooled in the channel dimension to retain the most discriminative local patterns and suppress noise interference: ； in: It means that the most discriminative features are selected through the competition mechanism to improve the sensitivity of the model to key anatomical areas. is the joint representation of fusion geometry and semantics obtained above; If the input contains pressure value, the normalized pressure difference is additionally spliced To enhance biomechanical perception.

8. The scoliosis brace registration method based on deep learning and biomechanical constraints according to claim 6, characterized in that: The step of combining the low-level features of the first layer and the high-level features of the last layer in the dynamic graph construction to perform feature aggregation to obtain the geometric features, pressure features, and motion features includes: The high-level features output by the third layer are upsampled to the resolution of the first layer through trilinear interpolation, and the low-level features of the first layer are upscaled to the same dimension as the high-level features through 1×1 convolution; The high-level features and the low-level features are concatenated and dynamic weights are generated through the Sigmoid gating function. ,according to Weighted fusion of high- and low-level features; Then, a pre-trained pressure mask is applied to biomechanically enhance the fused features, where the feature value of the high-pressure area is amplified by a first preset multiple and the feature value of the low-pressure area is attenuated to a second preset multiple; Finally, the geometric features are decoupled and output through three independent fully connected layers , pressure characteristics and motion characteristics ; The first feature information and the second feature information are simultaneously input into a pre-built hybrid configuration network for rigid registration and non-rigid registration, including: The second feature information is fused into a unified representation by channel and the dimension is compressed through a fully connected layer to retain key information for subsequent finite element simulation verification; During rigid registration, the initial alignment of the spine and the body surface is achieved through anatomical landmark matching and global transformation. Non-rigid registration achieves local anatomical adaptation through conditional thin plate spline deformation field prediction and biomechanical constraint optimization; During rigid registration, the spine is initially aligned with the body surface through anatomical landmark matching and global transformation, including: First, based on the geometric characteristics of the spine and surface geometric characteristics Perform feature matching and use cosine similarity to filter out high-confidence corresponding point pairs; Then the optimal rigid body transformation matrix is ​​calculated by the Kabsch algorithm to solve the rotation matrix With translation vector , minimize the mean square error of corresponding point pairs: ； in: The first Coordinates of anatomical landmarks; The first The coordinates of the corresponding feature points; Then the RANSAC algorithm is used to eliminate abnormal matches, and accurate global transformation parameters are obtained after iterative optimization; Finally, biomechanical constraints are added and incorporated into the loss function through the Lagrange multiplier method to ensure that the registration results are anatomically reasonable and output the initial alignment model.

9. The scoliosis brace registration method based on deep learning and biomechanical constraints according to claim 8, characterized in that: The non-rigid registration achieves local anatomical adaptation through conditional thin plate spline deformation field prediction and biomechanical constraint optimization, including: The surface pressure characteristics after fusion With motion features as input, the control point set is predicted through the conditional TPS network and corresponding weights , generating a point-by-point displacement field: ； in: is the radial basis function, Represents the control points in the deformation field With the target point Euclidean distance between; affine transformation matrix With translation vector is the network prediction parameter, matrix Processing global linear deformation, vector Compensate for overall offset; Combined with finite element simulation verification, the deformation field is mapped to the surface soft tissue mesh to achieve anatomically reasonable and high-precision local adaptation.

Images