Flexible scoliosis brace personalized adaptation system based on artificial intelligence analysis

Abstract

The invention discloses a flexible scoliosis brace personalized adaptation system based on artificial intelligence analysis. The flexible scoliosis brace personalized adaptation system comprises a dimensional scanning module, a feature extraction and curve fitting module, a machine learning prediction module and an adaptation verification and closed loop module. The three-dimensional scanning module is used for adopting multi-frequency phase structured light projection and dual-band laser to assist calibration; and the feature extraction and curve fitting module is used for accurately matching the three-dimensional image of the spine of the patient with the reference anatomical model Mref through an improved I-ICP algorithm. The invention belongs to the technical field of medical auxiliary data processing, and aims to solve the problems that in the prior art, the measurement precision is limited, quantitative evaluation standards are lacked, individual differences are difficult to adapt, and a dynamic tracking and rapid iteration mechanism is lacked. The method has the technical effects that the problems that the measurement precision is limited, a quantitative evaluation standard is lacked, individual differences are difficult to adapt, and a dynamic tracking and rapid iteration mechanism is lacked are effectively solved.

Description

Technical Field

[0001] The present application belongs to the field of medical auxiliary data processing technology, and specifically relates to a personalized adaptation system for a flexible scoliosis brace based on artificial intelligence analysis. Background Art

[0002] Scoliosis is a complex three-dimensional spinal deformity characterized by coronal lateral curvature, sagittal anomalies, and axial vertebral rotation. Brace therapy, the primary non-surgical intervention for adolescent idiopathic scoliosis (AIS), relies on controlling the progression of spinal deformity through external correction. Currently, the efficacy of brace therapy is highly dependent on the precise matching of the brace to the patient's anatomy. Therefore, a personalized fitting system for flexible scoliosis braces based on artificial intelligence analysis plays a crucial role in precisely matching the brace to the patient.

[0003] However, most current personalized fitting systems for flexible scoliosis braces have significant technical flaws: contact measurement tools cannot accurately capture the three-dimensional deformity characteristics of the spine, two-dimensional X-rays make it difficult to fully assess the surface morphology, and although conventional CAD / CAM systems can achieve partial digitization, they are not intelligent enough and still require a lot of manual intervention. They also rely on manual measurements by therapists, resulting in limited accuracy (measurement error can reach ±5mm); the lack of quantitative evaluation standards results in significant differences in fitting results among different institutions; the limited size classification of prefabricated braces makes it difficult to adapt to individual differences; patients in the growth and development period need to frequently adjust the brace but lack dynamic tracking and rapid iteration mechanisms. Summary of the Invention

[0004] The present application provides a personalized adaptation system for a flexible scoliosis brace based on artificial intelligence analysis, which aims to solve the problems in the existing technology such as limited measurement accuracy, lack of quantitative evaluation standards, difficulty in adapting to individual differences, and lack of dynamic tracking and rapid iteration mechanisms.

[0005] A personalized fitting system for flexible scoliosis braces based on artificial intelligence analysis, including a 3D scanning module, a feature extraction and curve fitting module, a machine learning prediction module, and an adaptation verification and closed-loop module;

[0006] The 3D scanning module is used to achieve point cloud capture with 0.1mm accuracy using multi-frequency phase structured light projection and dual-band laser assisted calibration;

[0007] The feature extraction and curve fitting module is used to accurately match the patient's spine 3D image with the reference anatomical model Mref using an improved I-ICP algorithm, automatically detect 32 bony landmarks and 10 soft tissue contours of scoliosis patients, extract the 3D morphological features of the spine, and calculate 18 geometric quantities in three planes;

[0008] The machine learning prediction module is based on a hierarchical stacking ensemble, integrating three types of base learners: gradient boosting tree, extreme random forest, and radial basis kernel SVM. It optimizes hyperparameters through tree-structured Bayesian search and outputs the brace model, shell thickness function, and 16×16 pressure matrix.

[0009] The adaptation verification and closed-loop module collects pressure distribution data in real time through a 16×16 flexible resistive sensor array, and uses partitioned model predictive control (MPC) in collaboration with variable structure PID to achieve a 0.42-second step response. At the same time, it can encrypt and upload real-time data to the cloud, triggering incremental learning to optimize model parameters.

[0010] Furthermore, the three-dimensional scanning module also includes denoising and pre-alignment functions;

[0011] The denoising process is used to remove outliers and retain anatomical features;

[0012] The pre-alignment performs preliminary registration on the original point cloud based on the improved iterative closest point (I-ICP) algorithm to eliminate the posture offset error.

[0013] Furthermore, the three-dimensional scanning module supports smartphones to capture images in standard postures and generate sparse point clouds through attention U-Net segmentation.

[0014] Furthermore, the specific contents of the feature extraction and curve fitting module are as follows:

[0015] a) Surface mesh regularization

[0016] Input the original point cloud P = {pi}, remove outliers through α-shape shell clipping, and construct a bidirectional Laplace smooth mesh for vector update operation as follows:

[0017]

[0018] Where, η = 0.35, V i old is the original vector of node i, V j old is the original vector of node j, N(i) is the number of neighbors of node i;

[0019] b) Rough estimation of the main spinal line

[0020] The bone point hypothesis based on stacked Hough is used: the point density peak hm is found in every three-layer cross section, and then the rough curve is spliced by dynamic time warping (DTW) to make the bone point hypothesis h of different layers m Splice them into a thick curve in chronological order, allowing non-linear alignment, as follows:

[0021]

[0022] Among them, c and h m+1 , c represents the coordinate values of the m-th layer and the m+1-th layer sections in the c-th dimension respectively;

[0023] c) Minimization of high-order B-core curvature energy

[0024] Let the curve parameter form c(υ)=0blB(υ), the objective function

[0025]

[0026] in, represents the k-th order derivative, {γ1, γ2, γ3} = {0.45, 0.25, 0.30}.

[0027] Furthermore, the specific contents of the machine learning prediction module are as follows:

[0028] a) Parallel preprocessing of training data;

[0029] b) Basic learner layer;

[0030] c) Meta-fusion layer;

[0031] d) Online migration and confidence monitoring.

[0032] Furthermore, the specific content of the parallel preprocessing of the training data is as follows:

[0033] Data Sheet N = 1852, feature dimension K = 22. For continuous features, the Box-Cox transformation is used to eliminate skewness and optimize the distribution characteristics of the data, thereby improving the effect of subsequent modeling. The details are as follows:

[0034]

[0035] in, represents the jth original data point, Express The value after parametric transformation, λ is the parameter of the transformation.

[0036] Furthermore, the base learner layer can output three-way prediction y (1) ,y (2) ,y (3) ;

[0037] The meta-fuser layer is based on the input z=[y (1) ,y (2) ,y (3), Var(y)], Elastic-Net regression is used for model training, and the weight ω is optimized by minimizing the objective function, as follows:

[0038]

[0039] in, is the predicted value of the model, ω q is the weight of the qth feature, z q is the qth feature, ρ is the intensity coefficient;

[0040] The online migration and confidence monitoring can increase the sample 50 cases trigger the sliding window kernel deviation detection, which is used to evaluate whether the deviation between the model prediction value and the true value exceeds a certain threshold, as follows:

[0041]

[0042] Where ψ is the ratio of the error exceeding the threshold, is a sample set containing multiple data points. represents the error threshold, y w is the true value of the w-th sample, The predicted value of the w-th sample, that is, the output of the model for sample w.

[0043] Furthermore, the brace shell is made of a three-layer composite of elastic photosensitive resin + carbon fiber reinforced polyurethane, the inner lining is 1.5mm medical silicone; the middle layer is CF-PU elastic composite; and the outer layer is a breathable nylon mesh.

[0044] Furthermore, the real-time pressure F of the flexible resistive sensor ij The calculation formula of (t) is as follows:

[0045] F ij (t)=σ[R0 / R ij (t)-1]

[0046] Among them, R0 is the calibration resistor, R ij is the real-time resistance.

[0047] Furthermore, the specific contents of the adaptation verification and closed-loop module are as follows:

[0048] a) Modeling of 16×16 flexible piezoresistive array;

[0049] b) Using partitioned MPC + variable structure PID;

[0050] c) Rapid secondary scan verification;

[0051] d) Prediction of long-term efficacy.

[0052] Compared with the prior art, this application has at least the following beneficial effects:

[0053] Based on further analysis and research on existing technical problems, this application adopts multi-frequency phase structured light combined with adjustable auxiliary light threshold histogram equalization method to achieve submillimeter surface reconstruction, and uses the improved iterative closest point (I-ICP) algorithm to automatically extract key points of the spinal curve and 18 geometric indicators; uses layered stacking learning to establish anatomical feature-brace parameter mapping function, and outputs model, shell thickness function and partition pressure scheme; at the same time, in response to the problem of lack of three-dimensional scanning equipment in telemedicine scenarios, a sparse reconstruction and perspective correction algorithm based on multi-angle images of smartphones is proposed to improve measurement accuracy; in response to the problem of rapid changes in morphology of patients in the growth and development period, an online incremental learning and pressure closed-loop adjustment mechanism is designed to improve the orthotic force retention rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A schematic diagram of a module of a personalized adaptation system for a flexible scoliosis brace based on artificial intelligence analysis provided in one embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of this application more clear, this application is further described in detail below with reference to the accompanying drawings and embodiments.

[0056] like Figure 1 As shown, the present application provides a personalized adaptation system for a flexible scoliosis brace based on artificial intelligence analysis, including a three-dimensional scanning module, a feature extraction and curve fitting module, a machine learning prediction module, and an adaptation verification and closed-loop module.

[0057] The 3D scanning module uses multi-frequency phase structured light projection and dual-band laser-assisted calibration to achieve point cloud capture with 0.1mm accuracy. The phase calculation process uses a four-step phase shift + multi-frequency phase unwrapping strategy. The specific calculation formula for the four-step phase shift is as follows:

[0058]

[0059] Among them, P m (x, y) is the mth phase-shifted image, C and D are the background and modulation depth, Multi-frequency phase unwrapping uses triple frequency differences: fa = 60 Hz, fb = 120 Hz, and fc = 240 Hz. The absolute phase is calculated using the least squares remainder theorem. This is then decoded using the least squares remainder theorem to resolve the wrapped phase ambiguity. This provides high-fidelity 3D body surface data for subsequent modules, supporting precise feature modeling and brace adaptation.

[0060] At the same time, the 3D scanning module supports smartphones (such as iPhone 13Pro) to shoot images in standard postures and generate sparse point clouds through attention U-Net segmentation.

[0061] The 3D scanning module also includes denoising and pre-alignment functions;

[0062] Denoising was performed using bilateral filtering and α-shape hull clipping to remove outliers and preserve anatomical features.

[0063] Pre-alignment performs preliminary registration of the original point cloud based on the improved iterative closest point I-ICP algorithm to eliminate the posture offset error.

[0064] The specific contents of the super-resolution phase reconstruction algorithm of the 3D scanning module are as follows:

[0065] a) Multi-layer frequency domain fringe coding

[0066] The scanning projector is loaded with three sets of mutually prime fringe periods {Λ1, Λ2, Λ3}, which are 5.6mm, 4.8mm, and 4.0mm respectively. Projection frame sequence g r The formula for (x, y) is as follows:

[0067]

[0068] Among them, L is the ambient light component and M is the modulation depth map.

[0069] After the three frames are phase-shifted, the initial wrapped phase is obtained by subtracting the cosine, as follows:

[0070]

[0071] b) Congruence Residue Decoding

[0072] Let integer phase level k n (x, y) = round[θ◇Λ n / 2π], and find the unique non-wrapped phase using the Chinese remainder theorem, as follows:

[0073]

[0074] in, is the modular antielement, Φ=Π n Λ n At the same time, in order to prevent high-frequency noise from being amplified, a phase credibility mask is introduced, and local bidirectional interpolation compensation is used in the χ = 0 region. The details are as follows:

[0075]

[0076] c) Depth regression and sub-pixel resampling

[0077] Map the absolute phase to 3-D space:

[0078]

[0079] Among them, ν i and coefficients The coefficients are obtained by least squares method using a double calibration plate.

[0080] Then perform square root iteration to obtain sub-pixel coordinates as follows:

[0081]

[0082] Where J is the Jacobian and f is the image gradient constraint. Three iterations are enough to converge the Z error to 8.5×10-5m.

[0083] The feature extraction and curve fitting module uses an improved I-ICP algorithm to accurately match the patient's 3D spinal image with the reference anatomical model Mref. It automatically detects 32 bony landmarks and 10 soft tissue contours in scoliosis patients, extracts 3D spinal morphological features, and calculates 18 geometric quantities in three planes. This module decomposes complex 3D spinal deformities into quantifiable geometric parameters, providing objective parameters for subsequent diagnosis, brace design, and efficacy evaluation.

[0084] At the same time, the spinal centerline was fitted with a third-order B-spline curve, combined with curvature energy minimization (weight ratio λ1:λ2:λ3=6:3:1) to ensure curve smoothness and anatomical consistency.

[0085] The spine centerline is represented by a third-order B-spline curve q(s), and its fitting energy function is defined as:

[0086]

[0087] where p t is the sampling point, ε represents the point-curve distance, k * is the curvature, λ1:λ2:λ3=6:3:1.

[0088] The specific contents of the feature extraction and curve fitting module are as follows:

[0089] a) Surface mesh regularization

[0090] Input the original point cloud P = {pi}, remove outliers through α-shape shell clipping, and construct a bidirectional Laplace smooth mesh for vector update operation as follows:

[0091]

[0092] Where, η = 0.35, V i oldis the original vector of node i, V j old is the original vector of node j, and N(i) is the number of neighbors of node i.

[0093] b) Rough estimation of the main spinal line

[0094] The bone point hypothesis based on stacked Hough is used: the point density peak hm is found in every three-layer cross section, and then the rough curve is spliced by dynamic time warping (DTW) to make the bone point hypothesis h of different layers m Splice them into a rough curve in chronological order, allowing non-linear alignment. The details are as follows:

[0095]

[0096] Among them, c and h m+1 , c represents the coordinate values of the m-th layer and the m+1-th layer sections in the c-th dimension respectively.

[0097] c) Minimization of high-order B-core curvature energy

[0098] Let the curve parameter form c(υ)=0blB(υ), the objective function

[0099]

[0100] in, Denotes the kth derivative, {γ1, γ2, γ3} = {0.45, 0.25, 0.30}. A conjugate gradient-Newton hybrid search was used, with convergence achieved after 20 iterations; the convergence threshold ∈ = 10-7. The final output was 34 spinal feature points and 22 additional surface indices.

[0101] The machine learning prediction module is based on a hierarchical stacking ensemble, integrating three types of basis learners: gradient boosting tree, extreme random forest, and radial basis kernel SVM. It optimizes hyperparameters through tree-structured Bayesian search and outputs the brace model, shell thickness function, and 16×16 pressure matrix.

[0102] The underlying layered stacked ensemble model framework includes gradient boosting trees (LightGBM, learning rate 0.05, depth 6, forest of 200 trees), extreme random forests (ExtraTrees, forest of 500 trees), and radial basis kernel support vector machines (γ = 0.1, C = 1.2); the top layer is a generalized linear aggregator with L2 regularization. Hyperparameters are automatically tuned using tree-structured Bayesian search (TPE), with the overall objective function being:

[0103] Θ(ω)=β1Acc+β2Sens+β3Spec, β1:β2:β3=0.5:0.3:0.2

[0104] Where Acc is accuracy, Sens is sensitivity, and Spec is specificity. Accuracy, sensitivity, and specificity are weighted together using a comprehensive objective function to optimize model performance across multiple dimensions. The weights of β1, β2, and β3 are used to control the direction of model optimization and adapt to different clinical needs.

[0105] The specific contents of the machine learning prediction module are as follows:

[0106] a) Parallel preprocessing of training data

[0107] Data Sheet N = 1852, feature dimension K = 22. For continuous features, the Box-Cox transformation is used to eliminate skewness and optimize the distribution characteristics of the data, thereby improving the effect of subsequent modeling. The details are as follows:

[0108]

[0109] in, represents the jth original data point, Express The value after parametric transformation, λ is the parameter of the transformation.

[0110] Classification features are encoded with target and jittered with noise.

[0111] b) Basic learner layer

[0112] CatBoost: tree depth 7, learning rate 0.03, L2 regularization 3.0;

[0113] ObliviousRandomForest: 1024 layers, MSE split;

[0114] Radial basis kernel Gauss-SVR: σ = 0.25, C = 2.8, ε = 0.12

[0115] Output three-way prediction y (1) ,y (2) ,y (3) .

[0116] c) Meta-Fuser Layer

[0117] Input z=[y (1) ,y (2) ,y (3) , Var(y)], Elastic-Net regression is used for model training, and the weight ω is optimized by minimizing the objective function, as follows:

[0118]

[0119] in, is the predicted value of the model, ω q is the weight of the qth feature, z q is the qth feature, and ρ is the intensity coefficient.

[0120] We set ρ1 = 0.002 and ρ2 = 0.001. A five-fold cross-validation test yielded a mean absolute error of 2.7N and an F-β score (β = 0.5) of 0.91.

[0121] d) Online migration and confidence monitoring

[0122] Incremental Sample 50 cases trigger the sliding window kernel deviation detection, which is used to evaluate whether the deviation between the model prediction value and the true value exceeds a certain threshold, as follows:

[0123]

[0124] Where ψ is the ratio of the error exceeding the threshold, is a sample set containing multiple data points. represents the error threshold, y w is the true value of the w-th sample, The predicted value of the w-th sample, that is, the output of the model for sample w.

[0125] If ψ>0.08, retrain the worst performing base learner.

[0126] The top six features of the model, ranked by SHAP importance, were: principal curve angle (0.37), vertebral rotation angle (0.23), chest circumference (0.14), costovertebral angle difference (0.11), waist circumference (0.09), and trunk length (0.06). Ultimately, the model achieved 92.3% accuracy in a five-fold cross-validation, with a Kappa of 0.89. This allows for the analysis and selection of key features, ensuring transparent and reliable decision-making. The model also supports incremental learning, triggering a sliding window kernel deviation check (threshold ψ = 0.08) every 50 new data points, retraining the worst-performing base model to adapt to patient growth and development needs.

[0127] The fit verification and closed-loop module collects real-time pressure distribution data through a 16×16 flexible resistive sensor array, combining log-Cosh filtering to eliminate transient wear noise. It employs partitioned model predictive control (MPC) in conjunction with a variable structure PID controller to achieve a 0.42-second step response and a pressure error of ±5mm, ensuring precise and stable correction force.

[0128] At the same time, it has a secondary scanning verification function: a quick 3D scan is performed 5 minutes after the brace is worn, the average point-to-surface distance from the initial model is calculated, and the shell thickness of the out-of-tolerance area is adjusted in the direction of maximum density.

[0129] Closed-loop iterative optimization: Real-time data is encrypted and uploaded to the cloud via the MQTT protocol for incremental training and parameter migration, controlling and improving long-term orthopedic force retention.

[0130] The brace shell is a three-layer composite of elastic photosensitive resin and carbon fiber reinforced polyurethane. The inner lining is 1.5mm medical silicone (Shore 30A); the middle layer is a CF-PU elastic composite (E = 1.2GPa, variable thickness); and the outer layer is a breathable nylon mesh. The airbag-sensor module has a response time of 0.42s and is resistant to donning and doffing cycles. A 16×16 array of flexible resistive sensors is embedded in the predicted pressure zone. Real-time pressure F ij The calculation formula of (t) is as follows:

[0131] F ij (t)=σ[R0 / R ij (t)-1]

[0132] Among them, R0 is the calibration resistor, R ij is the real-time resistance.

[0133] The PID airbag driver is controlled by a micro pump group and a solenoid valve, and its regulation rules are as follows:

[0134]

[0135] The step response time is 0.42s when {Kp, Ki, Kd} = {0.8, 0.15, 0.05}. Closed-loop data is uploaded to the cloud via MQTT for incremental training, with a threshold drift adaptive correction cycle of 50 cases.

[0136] The specific contents of the adaptation verification and closed-loop module are as follows:

[0137] a) Modeling of 16×16 flexible piezoresistive array

[0138] Single inductor resistance R u，v (t) = R min +(κ-P u，v (t)) δ , δ = 1.65. The array is converted to a pressure matrix P(t), and log-Cosh filtering is used to remove instantaneous spikes caused by manual wearing. This is used to find the optimal Q matrix that minimizes the difference with the target matrix P, as follows:

[0139]

[0140] Among them, P * is the optimal Q matrix, Q u，v is the element in the uth row and vth column of the matrix Q, P u，v is the element in the uth row and vth column of the matrix P.

[0141] b) Use partitioned MPC + variable structure PID.

[0142] MPC prediction model:

[0143]

[0144] The prediction window is H=10, Q=diag(50), R=diag(0.3). When the solution time exceeds 20ms, it is automatically downgraded to PID-S. The details are as follows:

[0145]

[0146] c) Rapid secondary scan verification

[0147] 5 minutes after the adaptation is completed, a quick scan is performed to generate the point cloud P′. The average distance from the initial model point to surface is as follows:

[0148]

[0149] in, is the size of the point cloud P′, and ||qf|| is the distance between two points q and f.

[0150] like The system locally expands or reduces the wall thickness in the direction of maximum density, as follows:

[0151]

[0152] d) Prediction of long-term efficacy

[0153] Historical Pressure-Cobb Series The six-month control rate II is predicted using the gated recurrent unit-self-attention stack as follows:

[0154] u τ =σ(U1P τ +U2φ τ +b u ), r τ =σ(R1P τ +R2φ τ +b r )

[0155] h τ =u τ ☉h τ-1 +(1-u τ )☉tanh(WP τ +V(r τ ☉h τ-1 ))

[0156] The self-attention layer is as follows:

[0157]

[0158] final AUC=0.93, mean absolute error 4.5%.

[0159] Example 1

[0160] Standard body type: 15-year-old female, 35° thoracic curve. EinScan H-Plus has a scanning distance of 80cm, a resolution of 0.1mm, and collects a 2.6M effective point cloud. Bilateral filter kernel parameters:

[0161]

[0162] The three-dimensional Cobb angle θ is obtained by curve fitting 3D = 34.7° ± 0.3°. The model recommends Boston-M, with primary pressure zones T7–T9 and an expected correction force of 28 N. Scanned immediately after first wear: Cobb angle reduced to 19°; sensor matrix peak value 29.3 kPa, meeting ISO 13485 standards.

[0163] Example 2

[0164] Remote mobile phone image adaptation: 13-year-old male. 6 images were taken with iPhone13Pro in a standard pose, and 50k sparse point clouds were generated through attention U-Net segmentation; Poisson reconstruction after perspective correction:

[0165]

[0166] Estimated chest circumference: 762±8mm, waist circumference: 632±6mm (error <3%). TLSO-S is recommended via the cloud; it is delivered to the user and fine-tuned remotely via video. The airbag pressure error is ±5mm; the total adaptation period is 48 hours.

[0167] Example 3

[0168] Complex double-curve 3D-printed brace for a 12-year-old female with a 28° thoracic curve and a 32° lumbar curve. Printed using Form3L, Shore75A elastic resin. The shell thickness function is as follows:

[0169]

[0170] The maximum correction force simulated by the stress field was 19 N for the thoracic region and 22 N for the lumbar region. The stress uniformity index was 0.87. Six-week follow-up revealed that the thoracic curvature was 15° and the lumbar curvature was 18°. The average daily wearing time was 20.3 h.

[0171] In the above-mentioned personalized adaptation system of flexible scoliosis brace based on artificial intelligence analysis, multi-frequency phase structured light combined with adjustable auxiliary light threshold histogram equalization method is used to achieve submillimeter-level body surface reconstruction, and the improved iterative closest point (I-ICP) algorithm is used to automatically extract the key points of the spinal curve and 18 geometric indicators; layered stacking learning is used to establish the anatomical feature-brace parameter mapping function, and output the model, shell thickness function and partition pressure scheme; at the same time, in response to the problem of lack of three-dimensional scanning equipment in telemedicine scenarios, a sparse reconstruction and perspective correction algorithm based on multi-angle images of smartphones is proposed to improve measurement accuracy; in response to the problem of rapid changes in the morphology of patients during the growth and development period, an online incremental learning and pressure closed-loop adjustment mechanism is designed to improve the correction force retention rate.

[0172] The 3D scanning module provides basic data, the feature extraction module analyzes anatomical structures, the machine learning module generates adaptation parameters, the manufacturing module produces the brace, the verification module provides real-time feedback on the results, and the closed-loop module continuously optimizes the system, forming a complete chain of "acquisition → modeling → prediction → manufacturing → verification → iteration." By integrating AI with flexible sensing technology, personalized, high-precision, and dynamically adaptive scoliosis correction can be achieved, improving patient compliance and treatment outcomes.

[0173] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A personalized adaptation system for flexible scoliosis braces based on artificial intelligence analysis, characterized in that: It includes 3D scanning module, feature extraction and curve fitting module, machine learning prediction module, and adaptation verification and closed-loop module; The 3D scanning module is used to achieve point cloud capture with 0.1mm accuracy using multi-frequency phase structured light projection and dual-band laser assisted calibration; The feature extraction and curve fitting module is used to accurately match the patient's spine 3D image with the reference anatomical model Mref using an improved I-ICP algorithm, automatically detect 32 bony landmarks and 10 soft tissue contours of scoliosis patients, extract the 3D morphological features of the spine, and calculate 18 geometric quantities in three planes; The machine learning prediction module is based on a hierarchical stacking ensemble, integrating three types of base learners: gradient boosting tree, extreme random forest, and radial basis kernel SVM. It optimizes hyperparameters through tree-structured Bayesian search and outputs the brace model, shell thickness function, and 16×16 pressure matrix. The adaptation verification and closed-loop module collects pressure distribution data in real time through a 16×16 flexible resistive sensor array, and uses partitioned model predictive control (MPC) in collaboration with variable structure PID to achieve a 0.42-second step response. At the same time, it can encrypt and upload real-time data to the cloud, triggering incremental learning to optimize model parameters.

2. The personalized adaptation system for flexible scoliosis braces based on artificial intelligence analysis according to claim 1, characterized in that: The three-dimensional scanning module also includes denoising and pre-alignment functions; The denoising process is used to remove outliers and retain anatomical features; The pre-alignment performs preliminary registration on the original point cloud based on the improved iterative closest point (I-ICP) algorithm to eliminate the posture offset error.

3. The flexible scoliosis brace personalized adaptation system based on artificial intelligence analysis according to claim 1, characterized in that: The 3D scanning module supports smartphones to capture images in standard postures and generates sparse point clouds through attention U-Net segmentation.

4. The personalized adaptation system for flexible scoliosis braces based on artificial intelligence analysis according to claim 1, characterized in that: The specific contents of the feature extraction and curve fitting module are as follows: a) Surface mesh regularization Input the original point cloud P = {pi}, remove outliers through α-shape shell clipping, and construct a bidirectional Laplace smooth mesh for vector update operation as follows: Where η = 0.35, is the original vector of node i, is the original vector of node j, N(i) is the number of neighbors of node i; b) Rough estimation of the main spinal line The bone point hypothesis based on stacked Hough is used: the point density peak hm is found in every three-layer cross section, and then the rough curve is spliced by dynamic time warping (DTW) to make the bone point hypothesis h of different layers m Splice them into a thick curve in chronological order, allowing non-linear alignment, as follows: Among them, c and h m+1 , c represents the coordinate values of the m-th layer and the m+1-th layer sections in the c-th dimension respectively; c) Minimization of high-order B-core curvature energy Let the curve be parametric Objective function in, represents the k-th order derivative, {γ1, γ2, γ3} = {0.45, 0.25, 0.30}.

5. The personalized adaptation system for flexible scoliosis braces based on artificial intelligence analysis according to claim 1, characterized in that: The specific contents of the machine learning prediction module are as follows: a) Parallel preprocessing of training data; b) Basic learner layer; c) Meta-fusion layer; d) Online migration and confidence monitoring.

6. The personalized adaptation system for flexible scoliosis braces based on artificial intelligence analysis according to claim 5, characterized in that: The specific content of the parallel preprocessing of training data is as follows: Data Sheet N = 1852, feature dimension K = 22. For continuous features, Box-Cox transformation is used to eliminate skewness and optimize the distribution characteristics of the data, thereby improving the effect of subsequent modeling. The details are as follows: in, represents the jth original data point, Express The value after parametric transformation, λ is the parameter of the transformation.

7. The personalized adaptation system for flexible scoliosis braces based on artificial intelligence analysis according to claim 5, characterized in that: The base learner layer can output a three-way prediction y (1) ,y (2) ,y (3) ; The meta-fuser layer is based on the input z=[y (1) ,y (2) ,y (3) , Var(y)], Elastic-Net regression is used for model training, and the weight ω is optimized by minimizing the objective function, as follows: in, is the predicted value of the model, ω q is the weight of the qth feature, z q is the qth feature, ρ is the intensity coefficient; The online migration and confidence monitoring can increase the sample 50 cases trigger the sliding window kernel deviation detection, which is used to evaluate whether the deviation between the model prediction value and the true value exceeds a certain threshold, as follows: Where ψ is the ratio of the error exceeding the threshold, is a sample set containing multiple data points. represents the error threshold, y w is the true value of the w-th sample, The predicted value of the w-th sample, that is, the output of the model for sample w.

8. The flexible scoliosis brace personalized adaptation system based on artificial intelligence analysis according to claim 1, characterized in that: The brace shell is made of a three-layer composite of elastic photosensitive resin + carbon fiber reinforced polyurethane, the inner lining is 1.5mm medical silica gel; the middle layer is CF-PU elastic composite; and the outer layer is a breathable nylon mesh.

9. The flexible scoliosis brace personalized adaptation system based on artificial intelligence analysis according to claim 1, characterized in that: The real-time pressure F of the flexible resistive sensor ij The calculation formula of (t) is as follows: F ij (t)=σ[R0 / R ij (t)-1] Among them, R0 is the calibration resistor, R ij is the real-time resistance.

10. The flexible scoliosis brace personalized adaptation system based on artificial intelligence analysis according to claim 1, characterized in that: The specific contents of the adaptation verification and closed-loop module are as follows: a) Modeling of 16×16 flexible piezoresistive array; b) Using partitioned MPC + variable structure PID; c) Rapid secondary scan verification; d) Prediction of long-term efficacy.

Images