Scoliosis orthosis registration method based on deep learning and biomechanical constraint

Abstract

The invention discloses a scoliosis orthosis registration method based on deep learning and biomechanical constraint, and relates to the technical field of scoliosis orthosis design, and the method comprises the steps: respectively obtaining first feature information of a three-dimensional spine model and second feature information of a three-dimensional body surface model, the second feature information comprises a geometric feature, a pressure feature and a motion feature; 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 local adaptation data with reasonable dissection; and performing joint optimization on the initial alignment model and the high-precision local adaptation data with reasonable dissection to obtain a final registration result. Geometric and motion features of a spine and a body surface are extracted through double-flow features, so that a registration network of coarse rigidity and then fine non-rigidity is carried out, and transition from geometric registration to function optimization is realized.

Description

Technical Field

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

[0002] Scoliosis is a common three-dimensional spinal deformity. It occurs when the spine deviates from the midline in the coronal (frontal) plane and curves laterally, resulting in a Cobb angle greater than 10°. Scoliosis is often accompanied by spinal rotation and changes in the sagittal (lateral) curvature. It not only affects appearance but can also impact cardiopulmonary function, the nervous system, and mental health. In severe cases, it can affect cardiopulmonary function and even lead to paralysis.

[0003] Scoliosis correction methods must be tailored based on factors such as the type and severity of the curve, as well as the patient's age and physical condition. The most commonly used brace is the Xeno brace. The Xeno brace was developed by Dr. Xeno in France in the early 1970s based on the theory of scoliosis correction with the E.G. Abbott plaster brace. The Xeno brace is a fully plastic brace with the advantages of being lightweight and simple. It utilizes the "three-point force" correction principle, providing anti-rotation, extension, and active correction through the provision of pressure pads and extension spaces, combined with growth mechanisms, respiratory training, and physical therapy. The effectiveness of the Xeno brace depends on the close fit of the brace to the trunk and the precise alignment of the pressure points with the spinal curvature.

[0004] However, current orthotic designs face the following technical challenges: Inadequate fit and comfort: Traditional plaster models require a long production cycle, and universal orthotics, due to their fixed size, cannot adapt to individual differences, resulting in poor fit and low long-term comfort. Inaccurate pressure point alignment: While existing 3D printing technology can improve fit, it lacks automated alignment methods for the spine and torso. Pressure point placement relies on manual experience, resulting in unstable correction results. Summary of the Invention

[0005] In order to solve the above technical problems, this application proposes the following technical solutions: In a first aspect, an embodiment of the present application provides a scoliosis brace registration method based on deep learning and biomechanical constraints, comprising: 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; The initial alignment model and the anatomically sound high-precision local adaptation data are jointly optimized to obtain the final registration result.

[0006] In a possible implementation, obtaining first feature 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.

[0007] In one possible implementation, 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.

[0008] In one possible implementation, the hierarchical feature extraction is then used to perform multi-scale abstraction to gradually capture the geometry, kinematics, and biomechanical properties of the spine, and output 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 , 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: The homogeneous transformation matrix representing the rigid body motion transformation from vertebra i to vertebra j is used to fully describe the position and orientation changes of vertebra j relative to vertebra i. , represents the 6 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 , through the Bi-LSTM network to fuse the temporal curvature changes With the finite element strain distribution, the multi-dimensional bending mode characteristics are output, where is the length of the spinal arc; 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.

[0009] In a possible implementation, obtaining the second feature information of the three-dimensional body surface model includes: 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. 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 to perform feature aggregation.

[0010] In a possible implementation, preprocessing the three-dimensional body surface model data includes: 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.

[0011] In one possible implementation, the dynamic graph construction using pre-processed 3D body surface model data and the layer-by-layer capture of complex geometric features of the body surface through feature space adaptive neighborhood search and hierarchical abstraction strategy include: 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 large-scale geometric features such as 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 such as bone protrusions and spinous processes.

[0012] In one possible implementation, each layer includes edge feature calculations to achieve high-precision feature modeling by dynamically capturing local geometric relationships and biomechanical differences, 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; Finally, the channel dimension maximum pooling is performed on the edge features of the k neighboring points to retain the most discriminative local patterns and suppress noise interference: ; If the input contains pressure value, the normalized pressure difference is additionally spliced To enhance biomechanical perception.

[0013] In a possible implementation, the step of performing feature aggregation by combining 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 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 .

[0014] In a possible implementation, the simultaneously inputting the first feature information and the second feature information into a pre-built hybrid configuration network for rigid registration and non-rigid registration includes: 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 constrained optimization.

[0015] In one possible implementation, the rigid registration is performed by matching anatomical landmarks and solving a global transformation to achieve initial alignment of the spine with the body surface, 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: is the coordinate of the i-th anatomical landmark point extracted from the 3D spinal bone model; is the coordinate of the jth corresponding feature point extracted from the 3D body surface model; 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.

[0016] In one possible implementation, 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 and motion characteristics As input, the control point set is predicted by 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.

[0017] In one possible implementation, the joint optimization of the initial alignment model and the anatomically sound 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: 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; 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; Finally, the total loss function is determined as: .

[0018] 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

[0019] 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; Figure 2 A schematic diagram of the feature extraction process of a three-dimensional spine model provided in an embodiment of the present application; 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; Figure 4 This is a visualization result diagram of the registration provided in the embodiment of this application. DETAILED DESCRIPTION

[0020] The present invention will be described below with reference to the accompanying drawings and specific implementation methods.

[0021] See also Figure 1 The scoliosis brace registration method based on deep learning and biomechanical constraints provided in this embodiment includes: S101, respectively obtaining first characteristic information of a three-dimensional spine model and second characteristic information of a three-dimensional body surface model.

[0022] 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) ( ), and generates a 24×3 Boolean matrix encoding the vertebral freedom constraints (such as thoracic sagittal rotation lock). The output data includes geometric coordinates, 24-dimensional vertebral labels and enhanced point clouds (N×28 matrix) and motion constraint tables, and is calculated by coordinate alignment error ( ), label mapping accuracy ( ) and stiffness gradient smoothness (slope ) and other indicators to complete quality control and provide standardized input for subsequent spinal correction planning and biomechanical weighted sampling.

[0023] To ensure that the pressure distribution of the orthosis is modeled, priority should be given to biomechanically sensitive areas (such as the apical vertebra and the distal vertebra) to avoid false negatives caused by stress concentration.

[0024] See also Figure 2 In this embodiment, the first feature information of the 3D spine model is obtained by dynamically calculating weights and guiding anatomical priors through biomechanical weighted sampling to enhance key regional features. Hierarchical feature extraction is then performed to gradually capture the geometric, kinematic, and biomechanical properties of the spine through multi-scale abstraction, outputting high-level final features that balance local geometric details with global mechanical properties.

[0025] Biomechanical weighted sampling achieves feature enhancement of key areas through dynamic weight calculation and anatomical prior guidance. The specific steps include: first, based on the vertebral stiffness coefficient ( to ) and the distance from the main curve apex to calculate the dynamic weight: ; in: is the vertebral body stiffness value, is the Euclidean distance from the point to the apex (unit: mm), and the weight of the apex area is increased , non-critical areas reduce . Then, improved farthest point sampling (FPS) is performed. In each iteration, the point farthest from the sampled point set and with the highest weight is selected, prioritizing coverage of areas with high stiffness and close to the apex of the main bend. The sampling interval is increased to 3mm in critical areas and relaxed to 8mm in non-critical areas. During the sampling process, outliers are removed in real time (neighborhood density <10 points / 20mm³), and low-density areas (such as pedicles) are resampled using Poisson disks (target density 2 points / mm²). Finally, adaptive neighborhood grouping is performed: the apical vertebral area is queried using a sphere with a radius of 10mm, and the lumbar sacral area is expanded to 20mm. 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 interpolation along the normal) to ensure the integrity of local feature extraction. The output is 1024 sampling points and their neighborhood feature matrix (32 neighborhood points × 64 dimensions), providing high-semantic density input data for subsequent hierarchical feature aggregation.

[0026] Hierarchical feature extraction gradually captures the geometric, kinematic, and biomechanical properties of the spine through multi-scale abstraction and is divided into four layers: Level 1 (1024 points → 64 dimensions) is used to extract basic geometric features. Multi-radius sphere query (10mm / 20mm / 30mm) is used to construct local neighborhoods and calculate 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 (the tangent direction is determined by the derivative of the spine connection) through SO(3) equivariant convolution. The direction attention weight is: ; in: The covariance matrix Features , hyperparameters Controls the sharpness of the weight distribution.

[0027] Level 2 (512 points → 128 dimensions) introduces kinematic chain modeling and uses graph convolutional networks (GCNs) to regress and predict the relative pose parameters of adjacent vertebrae (6 degrees of freedom): ; in: The homogeneous transformation matrix representing the rigid body motion transformation from vertebra i to vertebra j is used to fully describe the position and orientation changes of vertebra j relative to vertebra i. , represents the 6 degrees of freedom of traversal rigid body motion; Represents multi-layer perceptron regression, constraining the rotation angles of adjacent vertebrae; They are Lie algebra generators and constitute the complete basis of Lie algebra. Any rigid body motion can be expressed as a linear combination of these bases.

[0028] Level 3 (256 points → 256 dimensions) focuses on Cobb angle-related features, projects the point cloud onto the Cobb angle measurement plane (determined by the end vertebra center point and the top vertebra normal), and calculates the projection coordinates. Distance to the main curve . Fusion of temporal curvature changes through Bi-LSTM network ( , is the spinal arc length) and finite element strain distribution, and outputs 256-dimensional bending mode features.

[0029] Level 4 (128 points → 512 dimensions) integrates a real-time finite element solution layer to calculate node stresses ( ) and generates regularized gradients to ensure biomechanical plausibility.

[0030] Finally, trilinear interpolation upsampling and cross-level residual connection are performed: ; 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; It represents upsampling, which increases the resolution of deep features to match that of shallow layers. "" represents channel concatenation, concatenating features from different layers along 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 ≤ 0.7 mm) with global mechanical properties (stress error ≤ 4.8%).

[0031] See also Figure 3 Obtaining second feature information of the three-dimensional body surface model includes: preprocessing the three-dimensional body surface model data. Using the preprocessed three-dimensional body surface model data to construct a dynamic graph, and capturing complex geometric features of the body surface layer by layer through feature space adaptive neighborhood search and a 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, the geometric features, pressure features, and motion features are obtained.

[0032] Different from the 3D spine model, the input data for surface point cloud feature extraction is obtained through 3D scanning. matrix( Coordinates), with an initial density of about 1-2 points / mm², and auxiliary labeling of pressure-sensitive area labels (key areas such as the iliac crest, scapula, and sternum) in the form of a binary mask (0 / 1). The preprocessing stage first uses bilateral filtering: ; (Spatial Domain , Normal domain ) suppresses noise while removing statistical outliers (radius 5 The number of points in the neighborhood is less than 10) to eliminate isolated noise points; then Poisson disk sampling is performed (target density 1 point / ²) Balance the point cloud density and retain details in high curvature areas (curvature > 0.1), and perform local dense sampling (4 points / ²); Finally, a normalized pressure heat map (0-1 interval) 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 matrix( , pressure value), providing geometric and biomechanical dual-modal input for subsequent dynamic graph convolution.

[0033] DGCNN solves the problems of static neighborhood limitations and uniform feature response in traditional methods of surface point cloud processing by constructing dynamic graphs. Dynamic graph construction achieves layer-by-layer capture of complex surface geometric features through adaptive neighborhood search in feature space and hierarchical abstraction strategy: The first layer constructs a k-NN graph (k=20) using the original coordinates, calculates the initial edge features within a neighborhood with a radius of approximately 30 mm, and extracts large-scale geometric features such as local curvature (64 dimensions) through the EdgeConv module; the second layer re-searches the nearest neighbors (k=15) in the feature space (64 dimensions), calculates the bimodal edge features of three-dimensional coordinate differences and feature differences, 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 dynamically updated pressure-sensitive weights to screen adjacent points, and focuses on capturing the microscopic features of millimeter-level high-stress areas such as bony prominences and spinous processes (output 256 dimensions). Each layer accelerates k-NN search through the FAISS library and dynamically reconstructs neighborhood relationships in the feature space (non-fixed coordinate space), allowing the network to adapt to the geometric complexity of different anatomical regions. For example, sparse connections are used in flat areas of the trunk, while high-curvature bony protrusion areas are automatically connected with encrypted connections. Compared with traditional fixed graph construction methods, this method increases the feature coverage of key areas by 35% and reduces redundant calculations by 42%.

[0034] Each level includes edge feature calculation, which achieves high-precision feature modeling by dynamically capturing local geometric relationships and biomechanical differences: First, for each center point and the neighborhood points obtained by dynamic k-NN search (Based on the current feature space rather than a fixed coordinate space), splicing multi-dimensional input vectors - including the center point coordinates , relative coordinate difference of neighboring points and characteristic difference (When the feature dimension is 70, the total input dimension is 70), forming a joint representation that integrates geometry and semantics: ; Next, the concatenated vector is input into a three-layer MLP (multi-layer perceptron) for nonlinear transformation to extract high-order interaction features. Subsequently, the channel dimension maximum pooling is performed on the edge features of the k neighboring points: ; 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 fused geometry and semantics obtained above.

[0035] Retain the most discriminative local patterns (such as steep curvature or pressure mutation areas) and suppress noise interference; if the input contains pressure values, additionally splice the normalized pressure difference This process is iteratively updated layer by layer, allowing the network to adapt to neighborhood relationships at different levels of abstraction. The bottom layer captures macroscopic posture (such as torso curvature), the middle layer identifies mesoscale structures (such as scapular concavity), and the upper layer focuses on microscopic features (such as the curvature of bony prominences), ultimately outputting 256 high-level features.

[0036] 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 trilinear interpolation, and at the same time, the 64-dimensional low-level features of the first layer are upsampled to 256 dimensions through 1×1 convolution. After the two are spliced, dynamic weights are generated through the Sigmoid gating function. ,according to The high- and low-level features are weighted and fused (retaining local details and global semantics). A pre-trained pressure mask is then applied to biomechanically enhance the fused features, amplifying the feature values ​​of high-pressure areas (such as the iliac crest) by 1.5 times and attenuating them to 0.5 times in low-pressure areas (such as the abdomen). Finally, the output is decoupled through three independent fully connected layers - the geometric features. (256 dimensions, including curvature / symmetry), pressure characteristics (128 dimensions, encoding pressure gradient), motion features (64 dimensions, predicting deformation trends), while ensuring anatomical fidelity (error < 0.2mm), the feature discrimination of pressure-sensitive areas is increased by 2.3 times, providing multi-dimensional biomechanical representation for downstream tasks.

[0037] S102, simultaneously inputting the first feature information and the second feature information into a pre-built hybrid configuration network 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.

[0038] In this embodiment, the registration process is specifically divided into rigid registration and non-rigid registration: Before entering the registration stage, the three surface features are first fused into a unified representation by channel: ; It is then compressed to 256 dimensions through a fully connected layer, retaining key information for subsequent finite element simulation verification.

[0039] The first step is rigid registration, which achieves the initial alignment of the spine and the body surface by matching anatomical landmarks and solving global transformation. First, based on the geometric characteristics of the spine bones (such as C7 spinous process, apex of sacral promontory) and body surface geometric features (such as the trunk midline and shoulder blade contour) for feature matching, and use cosine similarity to filter out high-confidence corresponding point pairs (such as the projection points of the sacral promontory on the body surface); then use the Kabsch algorithm to calculate the optimal rigid body transformation matrix - solve the rotation matrix With translation vector , minimize the mean square error of corresponding point pairs: ; in: is the coordinate of the i-th anatomical landmark point extracted from the 3D spinal bone model; is the coordinate of the jth corresponding feature point extracted from the 3D body surface model.

[0040] The RANSAC algorithm is then used to eliminate anomalous matches (such as mismatched points in soft tissue areas). Iterative optimization results in accurate global transformation parameters. Finally, biomechanical constraints (such as a ±5° rotational freedom limit for the thoracic spine) are added and incorporated into the loss function using the Lagrange multiplier method to ensure that the registration results are anatomically sound (for example, the distance error between the sacral promontory projection and the surface pelvic landmarks is ≤1.5mm). This step outputs an initial alignment model, providing a high-precision starting point for subsequent non-rigid registration. Key metrics include the target registration error (TRE ≤1.2mm) and the Dice coefficient of anatomical landmarks (≥0.95).

[0041] 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 surface pressure features are used to (128D) and motion characteristics (64D) is input and the control point set is predicted by 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. Finally, combined with finite element simulation verification, the deformation field is mapped to the body surface soft tissue mesh to achieve anatomically reasonable and high-precision local adaptation.

[0042] S103, jointly optimizing the initial alignment model and the anatomically reasonable high-precision local adaptation data to obtain a final registration result.

[0043] The final joint optimization is accomplished through the collaboration of multiple objective loss functions to achieve the final 3D registration effect. The core is a hybrid loss function that integrates rigid registration, non-rigid deformation, and biomechanical constraints. Since our proposed model is trained with an end-to-end learning strategy, the transfer module and the deformation module can be updated jointly.

[0044] The first is the non-rigid registration loss by calculating the Chamfer distance of the bone-body anatomical landmark pairs to ensure the global anatomical alignment of the skeleton and the body surface, and to establish the initial coordinate uniformity 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.

[0045] 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: ; 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.

[0046] 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: ; 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; That is, the total loss function is as follows: .

[0047] In order to verify the method proposed in this embodiment, the following experiments were also conducted in this embodiment: Table 1 Comparison of registration model performance

[0048] 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.

[0049] Table 2 Registration visualization results

[0050] 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.

[0051] 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).

[0052] In the embodiments of the present application, "at least one" refers to one or more, and "more" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. Among them, A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b and c can be represented by: a, b, c, ab, ac, bc, or abc, where a, b, c can be single or multiple.

[0053] The above description is merely a specific embodiment of the present application. Any person skilled in the art may easily conceive of variations or substitutions within the technical scope disclosed in this application, and such variations or substitutions shall be within the scope of protection of this application. The scope of protection of this application shall be subject to the scope of protection 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; The initial alignment model and the anatomically sound high-precision local adaptation data are jointly optimized to obtain the final registration result.

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 , 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: The homogeneous transformation matrix representing the rigid body motion transformation from vertebra i to vertebra j is used to fully describe the position and orientation changes of vertebra j relative to vertebra i. , represents the 6 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 , through the Bi-LSTM network to fuse the temporal curvature changes With the finite element strain distribution, the multi-dimensional bending mode characteristics are output, where is the length of the spinal arc; 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: The obtaining of second characteristic information of the three-dimensional body surface model includes: 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 large-scale geometric features such as 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 such as bone 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; Finally, the channel dimension maximum pooling is performed on the edge features of the k neighboring points 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: is the coordinate of the i-th anatomical landmark point extracted from the 3D spinal bone model; is the coordinate of the jth corresponding feature point extracted from the 3D body surface model; 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 and motion characteristics As input, the control point set is predicted by 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.

10. The scoliosis brace registration method based on deep learning and biomechanical constraints according to claim 1, characterized in that: 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: 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; 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; Finally, the total loss function is determined as: 。

Images