Method and system for designing and manufacturing scoliosis brace

Abstract

The invention relates to the technical field of scoliosis brace design, in particular to a method and system for designing and manufacturing a scoliosis brace through deep learning, three-dimensional scanning and three-dimensional printing technologies. The method comprises the following steps: prefabricating and storing brace male mold models suitable for different scoliosis types; performing preprocessing and feature extraction on an X-ray film of a patient by using deep learning, automatically calculating a Cobb angle and a Risser feature, and matching a corresponding male mold according to the Cobb angle and the Risser feature; performing noise reduction, filtering and alignment on the point cloud data of the patient by adopting 3D scanning and an ICP algorithm to generate a personalized three-dimensional model; and then a genetic algorithm is combined with feedback data to optimize design parameters of the brace, a processing file is automatically generated, and finally precise manufacturing and real-time adjustment of the brace are achieved. The full-process automation of the brace from intelligent diagnosis to personalized design and manufacturing is realized, and the suitability and the treatment effect are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of scoliosis brace design, and particularly to a method and system for designing and manufacturing scoliosis braces using deep learning, three-dimensional scanning, and three-dimensional printing technologies. Background Art

[0002] Scoliosis refers to an abnormal curvature of the spine in the lateral direction, usually accompanied by rotational deformation of the spine, resulting in asymmetry of the trunk. This disease is common in the adolescent population, and if left untreated, it may lead to back pain, limited respiratory function, and other serious health problems. There are mainly two treatment methods for scoliosis: surgical treatment and conservative treatment. For many adolescent patients, especially those with mild to moderate scoliosis, non-surgical treatment (such as brace treatment) is the preferred option, aiming to control the progression of spinal deformity through the brace and avoid surgical intervention.

[0003] Currently, the design and manufacturing of scoliosis braces mostly rely on manual design and personalized customization. Traditional brace design methods usually manually adjust the brace model based on the experience of clinicians and the X-ray films of patients. This process is not only inefficient but also requires high experience of designers. In addition, brace design often lacks systematic optimization, resulting in significant differences in the adaptability, comfort, and treatment effect of braces.

[0004] Deficiencies of traditional design methods:

[0005] 1. Manual and inefficient design process: Traditional scoliosis brace design relies on designers to manually design braces based on patients' X-ray films, clinical data, and other medical images. The design process of each brace requires repeated modification and debugging, which leads to high labor costs and long waiting periods. At the same time, the quality and effect of brace design highly depend on the experience and skills of designers, lacking standardization and personalized precision.

[0006] 2. Lack of high-precision adaptability: Due to the large amount of manual intervention in the brace design process and over-reliance on designers, the precise matching of the brace to the patient's spine often cannot be fully achieved, and parameters such as the size and pressure point distribution of the brace often cannot be accurately adjusted according to the individual differences of patients. This results in poor wearing effects of the brace, which may cause discomfort or fail to achieve the best treatment effect.

[0007] 3. Uncertainty of treatment effect: One of the biggest problems of current brace design methods is the lack of optimization for brace design and spinal correction effect. The pressure distribution, shape, and interaction between the brace and the spine are crucial for the treatment effect, but traditional design methods often ignore these details, resulting in the brace failing to achieve the expected correction effect.

[0008] 4. Poor comfort and durability: Most of the existing scoliosis braces are made of traditional materials and processes, usually being relatively bulky and uncomfortable. Patients often find it difficult to wear the brace for a long time due to discomfort, which affects the treatment effect.

[0009] With the development of information technology and manufacturing technology, especially the progress of deep learning, 3D scanning and 3D printing technologies, the design and manufacturing of scoliosis braces have witnessed revolutionary innovations. In recent years, the introduction of artificial intelligence (AI) and 3D digital technologies has provided new solutions for the automated design, personalized adjustment and optimization of braces.

[0010] Although deep learning, 3D scanning and 3D printing technologies have provided innovative solutions for the design and manufacturing of scoliosis braces, there are still some challenges in this field, such as how to more efficiently integrate multiple data sources (such as X-ray films, 3D scans, patient historical data, etc.) for comprehensive analysis, how to improve the accuracy of deep learning algorithms, and how to better integrate these technologies into medical devices for wide application in clinical practice. Summary of the Invention

[0011] To solve the above technical problems, the present invention combines deep learning, 3D scanning and 3D printing technologies, and proposes a new method for the design and manufacturing of scoliosis braces, which realizes the precise design and efficient manufacturing of personalized braces in an automated and intelligent manner. This method can not only improve the adaptability and treatment effect of the brace, but also reduce the degree of manual intervention, thus accelerating the customization and production process of the brace, and has important application prospects and market value.

[0012] To achieve the above object, the present invention adopts the following technical solutions:

[0013] A method for the design and manufacturing of a scoliosis brace, the method comprising the following steps:

[0014] 1) Pre-fabrication and storage of the positive mold model of the brace: Generate multiple positive mold models of the brace customized according to different scoliosis types (C-type, S-type, mixed type) clinically by using 3D CAD software. When designing the positive mold model, determine the support points, pressure distribution areas and geometric curves according to biomechanical principles, and embed applicable parameters in the model, including Cobb angle range, Risser sign grading, etc. Store the model and parameters in a database in a standard format (such as STL or STEP file);

[0015] 2) Patient X-ray processing and deep learning feature analysis: Input the patient's X-ray in DICOM format into the system, and perform preprocessing such as gray-scale normalization, noise filtering, and contrast enhancement. Subsequently, input the preprocessed image into a pre-trained multi-task deep learning model to automatically output detection parameters reflecting the patient's scoliosis characteristics;

[0016] 3) Automatically match the positive mold model of the brace:

[0017] Based on the detection parameters obtained in step (2) (including the predicted Cobb angle, Risser sign grading, and scoliosis type), use a preset tolerance (Cobb angle ±2 degrees) and a fuzzy matching algorithm (such as parameter matching based on Euclidean distance) to automatically retrieve the positive mold model of the brace that matches the patient's detection parameters from the database;

[0018] 4) Patient 3D scan data acquisition, preprocessing, and point cloud alignment:

[0019] Use a 3D scanning device to collect the point cloud data of the patient's back and side, perform statistical outlier removal and voxel grid filtering on the data, and convert it into a three-dimensional mesh model or digital elevation model. Then, adopt the automatic extraction of key anatomical points (such as the scapula, spinal vertex) and the iterative closest point (ICP) algorithm to spatially align the patient's point cloud data with the data of the positive mold model retrieved in step (3) to generate a personalized three-dimensional reference model for the patient;

[0020] 5) Personalized model optimization and parameter adjustment:

[0021] Based on the personalized model generated in step (4), use a deep learning module and a model optimization algorithm (genetic algorithm or particle swarm optimization algorithm) to automatically adjust the design parameters of the positive mold model of the brace. The design parameters include size, geometric curvature, and pressure point distribution, and introduce the patient's historical treatment data and brace wearing feedback data as weight factors during the optimization process to improve the adaptability of the model in terms of geometric matching and biomechanical performance;

[0022] 6) Generate processing files and manufacture the brace:

[0023] Compare the personalized brace model optimized in step (5) with the patient's three-dimensional data and X-ray images, and further optimize the design accuracy through manual or automatic deep learning algorithms to automatically generate processing files for 3D printing or CNC machining; Use 3D printing or CNC machining technology to manufacture the scoliosis brace, and install and adjust it to make the manufactured brace suitable for the patient to wear.

[0024] Preferably, the method further includes trimming the brace according to the patient's needs and smoothing the edges of the brace; and / or increasing the thickness and the hollow pattern holes of the brace to meet specific biomechanical requirements; and / or using an intelligent sensing device to monitor the fit between the brace and the patient's body in real time, and automatically adjusting the shape or pressure distribution of the brace through a feedback mechanism to improve the wearing experience and treatment effect.

[0025] Preferably, in step 2), the multi-task deep learning model includes:

[0026] (a) A regression branch for predicting the Cobb angle of the patient's spine;

[0027] (b) A classification branch for determining the Risser sign grade of the patient;

[0028] (c) An auxiliary branch for determining the type of scoliosis.

[0029] Preferably, in step 3), the matching algorithm is based on calculating the Euclidean distance between the predicted parameters and the parameters of each male mold model in the database, and selecting the model with the smallest distance and within a preset tolerance as the best matching model.

[0030] Preferably, in step 4), the preprocessing of the 3D scan data includes using a statistical outlier removal algorithm and voxel grid filtering to reduce the point cloud noise and the data volume, while retaining the key geometric features of the patient's three-dimensional structure.

[0031] Preferably, in step 5), the model optimization step uses a genetic algorithm to optimize the design parameters of the brace male mold model. The genetic algorithm includes steps of generating an initial population, fitness evaluation, selection, crossover, mutation, and generating a new generation of population. The objective function is a weighted combination of the geometric matching error, biomechanical performance error, and deviation penalty based on feedback data of each candidate design scheme;

[0032] and / or, in the automatic feature detection and data output step, the deep learning model integrates the output result after processing the X-ray film into a JSON format as the basis for subsequent matching of the brace male mold model.

[0033] Furthermore, the present invention also discloses a design system for a scoliosis brace. The system implements the method described above and includes:

[0034] A database management module for storing the brace male mold model and the patient's historical data; an image processing and deep learning module for preprocessing the patient's X-ray film and automatically extracting scoliosis features;

[0035] A 3D data acquisition module for acquiring the patient's 3D scan point cloud data;

[0036] A model generation and data alignment module, which is used to preliminarily and finely align the patient's point cloud data with the brace male mold model retrieved from the database to generate a personalized three-dimensional reference model;

[0037] A model optimization module, which is used to automatically adjust the brace design parameters based on the output parameters of deep learning and the feedback data;

[0038] A manufacturing module, which is used to generate processing files and manufacture scoliosis braces.

[0039] Preferably, the image processing and deep learning module adopts a multi-task convolutional neural network based on the ResNet or VGG architecture, which is respectively used to predict the Cobb angle, determine the Risser sign grade, and identify the scoliosis type.

[0040] Preferably, when aligning the patient's 3D scan data, the model generation and data alignment module adopts the Iterative Closest Point (ICP) algorithm and uses the automatically extracted key anatomical points as an aid to achieve the precise matching of the patient's point cloud data and the brace male mold model in the same coordinate system;

[0041] And / or, the model optimization module combines the patient's historical data, including wearing feedback and correction progress, to dynamically adjust the brace model to improve the correction effect;

[0042] And / or, the manufacturing module automatically generates numerical control processing paths or 3D printing slice data according to the personalized brace three-dimensional model output by the model optimization module, and docks with 3D printing or CNC processing equipment to realize the automatic manufacturing of braces and subsequent automated post-processing.

[0043] Preferably, the system further includes an intelligent sensing module, which is used to real-time monitor the adaptability between the brace and the patient's body, and adjust the shape or pressure distribution of the brace according to the deep learning algorithm through a feedback mechanism to further optimize the wearing experience.

[0044] Due to the adoption of the above technical solutions, the present invention proposes a brand-new scoliosis brace design and manufacturing method and system by combining deep learning, three-dimensional scanning and 3D printing technologies. Compared with the traditional technical methods, it has the following remarkable technical effects:

[0045] 1. Precise Diagnosis and Personalized Design: The present invention utilizes a deep learning model to preprocess and extract features from the patient's X-ray films, automatically and precisely calculates the Cobb angle and Risser sign, classifies them according to the type of scoliosis, and provides an accurate diagnostic basis for the design of the brace. At the same time, three-dimensional data of the patient is collected through 3D scanning technology, and automated data alignment (such as the iterative closest point algorithm) and key-point assisted registration are adopted to accurately match the patient's anatomical information with the prefabricated male mold model in the same coordinate system, ensuring that the generated three-dimensional reference model fully reflects the patient's individual physiological curve and realizing true personalized design.

[0046] 2. Automated and Intelligent Design and Manufacturing Process: The present invention constructs an integrated software system covering the entire process from data collection, image and point cloud preprocessing, deep learning feature analysis, model matching, data alignment, personalized model generation, optimization parameter adjustment, to automatic generation of processing files and brace manufacturing. Through automated algorithms and feedback closed-loop mechanisms, this system not only greatly reduces human operation errors and working hours, but also realizes the full-process intelligence from data input to product manufacturing, significantly improving the design and manufacturing efficiency.

[0047] 3. Optimization Adaptation and Feedback Regulation Mechanism: On the basis of generating the personalized brace model, the present invention further introduces a genetic algorithm or a particle swarm optimization algorithm to automatically optimize and adjust the key design parameters of the brace, such as size, geometric curvature, and pressure point distribution. At the same time, by comprehensively considering the patient's historical treatment data and the feedback of brace wearing, the dynamic weight of the objective function is adjusted, so that the optimization result better meets the actual needs of the patient, improves the wearing comfort and treatment effect, and realizes the real-time monitoring and secondary correction of the manufactured brace through the feedback module, forming a perfect closed-loop regulation system.

[0048] 4. Improving Manufacturing Precision and Reducing Production Cycle: The 3D printing or CNC machining equipment is controlled by the automatically generated CAD processing file to realize the high-precision automatic manufacturing of the brace. In the whole process, through noise reduction, filtering and voxel grid downsampling of the point cloud data, the data redundancy is effectively reduced, and the subsequent mesh reconstruction and simulation precision are improved, so as to ensure that the manufactured brace highly matches the patient's individual anatomical features in shape and structure. At the same time, the cycle from design to manufacturing is greatly shortened, the production cost is reduced, and the consistency of product quality during mass production is guaranteed.

[0049] 5. Innovative Applications of Multi-Technology Integration: The present invention organically integrates a number of advanced technologies such as deep learning, 3D scanning, automatic matching, fine registration, optimization algorithms, and manufacturing equipment control, realizing the full-process automation of scoliosis braces from intelligent diagnosis to personalized design, and then to automated manufacturing and real-time feedback adjustment. This multi-technology integration solution not only breaks through the limitation of relying on a large amount of manual experience in the traditional brace design process, but also significantly improves the product adaptability and clinical treatment effect, with remarkable innovation and practical value.

[0050] In summary, through the integration of advanced data processing and intelligent optimization technologies, the present invention realizes the precise and personalized design and manufacturing of scoliosis braces, which not only improves the adaptability and comfort of the braces, but also greatly enhances the manufacturing efficiency and quality stability, providing a new, efficient, and intelligent auxiliary tool for scoliosis treatment. Brief Description of the Drawings

[0051] Figure 1 : Schematic diagram of the overall process of scoliosis brace design and manufacturing;

[0052] Figure 2 : Schematic diagram of each main functional module and data flow;

[0053] Figure 3 : Flow chart of deep learning model data processing and optimization;

[0054] Figure 4 : Schematic diagram of preprocessing of patient 3D scan data and point cloud alignment. Detailed Description of the Preferred Embodiment

[0055] Next, in combination with the embodiments of the present invention, the technical solutions in the embodiments will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0056] Embodiment 1: Method for Designing and Manufacturing Scoliosis Braces Based on Deep Learning and 3D Data Processing

[0057] In this embodiment, a set of integrated software systems is used as a carrier to elaborate on the method for designing and manufacturing scoliosis braces. The entire process includes steps such as data acquisition, data processing, model matching, parameter optimization, generation of processing files, and final manufacturing and installation adjustment. The specific steps are as follows:

[0058] Step 1: Classification, Customization, and Database Storage of the Positive Mold Model of the Brace

[0059] 1) Preparation of the Positive Mold Model of the Brace

[0060] According to the common types of scoliosis clinically (such as mild, moderate and severe scoliosis, as well as type C, type S, mixed type, etc.), a variety of brace male mold models are designed and generated. When designing each male mold model, the support points, pressure distribution areas and applicable geometric curves are determined based on biomechanical principles. The male mold models in the standard STL or STEP file format are generated using 3D CAD software (Rodin 4D), and at the same time, annotation information (such as the applicable spinal curvature range, the position of reference pressure points, thickness preset, etc.) is embedded in the models.

[0061] 2) Database establishment and storage

[0062] Using a MySQL or MongoDB database, an information library of brace male mold models is established. Each model data includes the file storage path, applicable parameters (such as Cobb angle range, Risser sign grading), and other auxiliary description information.

[0063] Step 2: Import of patient X-ray films and deep learning feature analysis

[0064] 1) X-ray film data acquisition and preprocessing

[0065] The patient's X-ray films are transmitted into the system in DICOM format. If they are traditional films, they need to be digitized and scanned first. Preprocessing is performed on the input images, including operations such as gray normalization, noise filtering (such as using Gaussian filtering), and contrast enhancement, to ensure that the subsequent deep learning model can obtain clear image features.

[0066] 2) Deep learning model design and training

[0067] A convolutional neural network (CNN), a network based on the ResNet or VGG architecture, is used to extract features from the preprocessed X-ray films. The network is trained with a large amount of labeled data (including manual annotations of Cobb angle and Risser sign). During the training process, the cross-entropy loss function and the Adam optimizer are used, and the number of training epochs is determined according to the size of the dataset (between 50 and 200 epochs).

[0068] The present invention uses a multi-task deep learning network, and the overall network can be divided into two major branches:

[0069] Regression branch: Used to predict the value of the Cobb angle.

[0070] Classification branch: Used to judge the grading of the Risser sign.

[0071] 2.1) Basic convolutional neural network (CNN)

[0072] Select a classic network structure (such as ResNet, VGG, or DenseNet) as the feature extractor. The first few layers of the network gradually extract low-level and high-level features of the image through multiple convolutional layers, ReLU activation functions, and pooling layers.

[0073] 2.2) Task Branch Design

[0074] Regression Branch: Add a fully connected layer after the output of the base network, and use a linear activation to output a continuous value as the predicted value of the Cobb angle.

[0075] Classification Branch: Based on the shared convolutional features, add a fully connected layer and connect it to a Softmax activation layer to output the probability distribution of each Risser grade.

[0076] 2.3) Auxiliary Branch (Optional)

[0077] For the task of generating a high-precision 3D spine model, a generative adversarial network (GAN) structure can be designed. The generator receives features extracted from the CNN or a noise vector and generates 3D model data, while the discriminator judges the difference between the generated model and the real model to enhance the generation quality.

[0078] 2.4) Loss Function Design

[0079] Use the mean squared error (MSE) as the regression loss for Cobb angle prediction:

[0080]

[0081] where, and are the predicted and real Cobb angle values, respectively.

[0082] Use the cross-entropy loss function as the Risser sign classification loss:

[0083]

[0084] where K is the number of Risser grade categories.

[0085] The overall loss function can be in the form of a weighted combination:

[0086] L total = λ reg L reg + λ cls L cls

[0087] where, λ reg and λ cls are weight coefficients, and the optimal values are determined through experiments to balance the impacts of the regression and classification tasks.

[0088] 2.5) Data Augmentation and Expansion

[0089] To improve the generalization ability of the model and alleviate the problem of limited medical image data, the following data augmentation techniques are adopted:

[0090] Rotation: Randomly rotate the image within a certain angle range.

[0091] Scaling: Randomly adjust the image size and crop it.

[0092] Translation: Randomly translate the image horizontally and vertically.

[0093] Mirror Flip: Horizontally or vertically flip the image.

[0094] Noise Addition: Simulate the noise in the real scenario and expand the dataset by adding random noise.

[0095] 2.6 Model Training Process

[0096] Dataset Division: Divide the processed dataset into a training set, a validation set, and a test set, usually in the ratio of 70%:15%:15%.

[0097] Batch Size: Select according to the video memory situation (e.g., 16 or 32).

[0098] Learning Rate: Initially set to be small (e.g., 1e-4 or 1e-5), and adopt an adaptive learning rate scheduler.

[0099] Optimizer: Use the Adam optimizer or the SGD optimizer and set the momentum parameter.

[0100] Number of Training Epochs: Usually between 50 and 200 epochs, and stop early (Early Stopping) according to the performance on the validation set.

[0101] Training Process: In each training cycle, perform forward propagation on each batch of data, calculate the joint loss L total (and the adversarial loss, if 3D model generation is involved), and then perform backpropagation and parameter update. After each Epoch ends, evaluate the model performance on the validation set, monitor the regression error and classification accuracy, and prevent overfitting. The best model parameters can be saved using model checkpoints (Checkpoint).

[0102] Use the root mean square error (RMSE) to evaluate the prediction accuracy of the Cobb angle; use metrics such as accuracy, recall rate, and F1 score to evaluate the classification effect of the Risser sign.

[0103] Adjust hyperparameters and network structure according to the evaluation results. When necessary, introduce regularization (such as Dropout) and data augmentation strategies to further optimize the model performance.

[0104] 2.7) Generative Adversarial Network (GAN) Training (Optional)

[0105] The generator and discriminator are alternately trained, and the network parameters are iteratively updated using the adversarial training method. When training the generator, the reconstruction loss, adversarial loss, and feature matching loss are combined to ensure that the generated 3D spine model has high detail and realism. The discriminator continuously improves its discrimination ability between real data and generated data to ensure the continuous improvement of the quality of the generator's output.

[0106] 3) Automatic Feature Detection and Data Output

[0107] The system runs the trained CNN model in real time on the imported X-ray films, automatically detects the type of scoliosis, accurately calculates the Cobb angle value, and grades the Risser sign. The output results are saved in JSON format as the basis for subsequent model matching.

[0108] The pre-trained CNN model has been trained with a large amount of labeled data (including Cobb angle and Risser sign labels) and saved in a standard format (such as TensorFlow SavedModel, PyTorch's.pt file, etc.). The system loads this model into memory during initialization to ensure efficient and real-time subsequent inference processes.

[0109] The overall model adopts a multi-task network design, including the following output branches:

[0110] Regression Branch: Outputs a continuous value as the predicted value of the Cobb angle.

[0111] Classification Branch: Outputs various probabilities. After passing through the Softmax layer, the Risser sign grade is determined (for example, the probability distribution of grades 0 - 5), and the system selects the category with the highest probability as the final prediction.

[0112] Scoliosis Type Judgment: The type of scoliosis (such as type C, type S, etc.) can be determined based on the above output or an additional designed branch.

[0113] The preprocessed image data is passed into the model for forward propagation. The model outputs the prediction results of each task, usually returning a multi-dimensional array or dictionary. The data format returned by the model may be:

[0114] pred_cobb_angle: A floating-point value representing the predicted Cobb angle;

[0115] pred_risser_probs: A vector representing the probability distribution of each Risser grade;

[0116] pred_scoliosis_type: A class label or probability vector to determine the final scoliosis type through subsequent processing;

[0117] For the Risser sign grading, usually np.argmax(pred_risser_probs) is used to determine the final grading value.

[0118] Integrate the output results of each model through post - processing into a unified data structure:

[0119] Cobb angle: Convert the prediction result to a standard floating - point number, retaining an appropriate number of decimal places (e.g., 18.5°);

[0120] Risser sign grading: Take the index of the maximum probability from the probability vector and convert it to an integer category (e.g., grade 2);

[0121] Scoliosis type: Determine the final category according to a preset threshold or classification result (e.g., "C - type" or "S - type");

[0122] Package the integrated data into JSON format for easy calling by subsequent modules. In specific implementation, the JSON library provided by the programming language (the json module in Python) can be used to generate a string, which can be written to a file or directly passed to subsequent modules. The generated JSON data can be directly returned to the model matching module of the system through an interface as the basis for automatically matching the brace male mold model. At the same time, the JSON file can be stored in a database or on a local disk for data tracking, log recording, and subsequent review.

[0123] Step 3: Automatically retrieve and match the brace male mold model

[0124] 1) Implementation of the matching algorithm

[0125] The system compares the parameters output in Step 2 with the male mold models stored in the database. Using a fuzzy matching algorithm (such as parameter matching based on Euclidean distance) or a rule - based matching method, determine the brace male mold model that best matches the patient's data. The matching process considers the parameter tolerance (Cobb angle ±2 degrees) to ensure the reliability of the matching.

[0126] The following presents an implementation solution for a matching algorithm based on parameter differences and Euclidean distance, which is used to automatically detect the output parameters (Cobb angle and Risser sign grading) from X-ray films and find the brace male mold model that best matches the patient's data in the database. This solution also takes into account the preset tolerance range (±2 degrees for the Cobb angle), and on the premise of meeting the tolerance conditions, the best matching model is selected by calculating the Euclidean distance.

[0127] 1.1 The algorithm idea is as follows

[0128] Input data, the detection results obtained from the deep learning module include:

[0129] predicted_cobb: The predicted Cobb angle value (e.g., 18.5°);

[0130] predicted_risser: The predicted Risser sign grading (e.g., 2);

[0131] predicted_scoliosis_type: The predicted scoliosis type (e.g., "C type" or "S type").

[0132] Database model data structure, each brace male mold model record stored in the database contains the following key information:

[0133] model_id: Model identifier;

[0134] model_type: The scoliosis type corresponding to the brace (e.g., "C type");

[0135] cobb_angle: The reference Cobb angle (or central value) adapted by the model;

[0136] risser_grade: The reference Risser sign grading adapted by the model;

[0137] file_path: Other information such as the model file path.

[0138] 1.2 The matching process is as follows:

[0139] (1) Preliminary screening

[0140] Traverse all models in the database. First, it is required that the type of the model is consistent with the predicted scoliosis type; for each model, calculate the absolute difference between the predicted Cobb angle and the model reference Cobb angle, and the absolute difference between the predicted Risser sign grading and the model reference Risser sign grading; if both are within the preset tolerance range (i.e., the difference ≤ 2 degrees and ≤ 1 grade), then this model is considered a candidate model.

[0141] (2) Euclidean distance calculation

[0142] For each candidate model, calculate the Euclidean distance:

[0143]

[0144] As a measure of the quality of the match.

[0145] (3) Select the best model

[0146] Among all candidate models, select the model with the smallest distance as the best matching model. If the candidate models are empty, an empty result can be returned or a tolerance extension strategy can be triggered.

[0147] 1.3 Output result

[0148] Return the information of the best matching model (such as model_id, file_path, etc.) for the subsequent personalized brace design process. After successful matching, the system automatically reads the corresponding 3D CAD file and enters the subsequent personalized design stage.

[0149] Step 4: Acquisition of patient 3D data and generation of personalized model

[0150] 1) 3D scan data acquisition

[0151] Use the built-in 3D scan application of the iPad or a portable structured light 3D scanner to scan the patient's back and side to obtain high-precision point cloud data. During the acquisition process, multi-angle and continuous frame acquisition are adopted, and the overall scan accuracy can be improved by using real-time data fusion technology (SLAM algorithm).

[0152] 2) Point cloud data preprocessing

[0153] Perform noise reduction, filtering (using the statistical outlier removal algorithm) and voxel grid filtering on the obtained point cloud data to reduce the data volume while maintaining the key geometric features. Convert the preprocessed point cloud data into a 3D mesh model or a digital elevation model (DEM).

[0154] The following gives a specific method for implementing point cloud data preprocessing based on Python and the Open3D library. This method mainly includes the following steps:

[0155] 2.1) Load point cloud data

[0156] Read the point cloud data from a file (such as PCD or PLY format).

[0157] 2.2) Statistical Outlier Removal

[0158] Using a statistical outlier removal algorithm, calculate the mean distance of each point to its surrounding neighborhood points, and filter out the outliers with a relatively large distance according to the preset parameters. This algorithm usually requires setting parameters:

[0159] nb_neighbors: The number of neighborhood points used for statistics.

[0160] std_ratio: The standard deviation ratio. Points with a distance mean greater than the mean plus the standard deviation of this ratio will be removed.

[0161] 2.3) Voxel Grid Filtering

[0162] Divide the point cloud data into cubic grids of a fixed size in space. The points within each grid are represented by the center or mean within the grid, thereby achieving downsampling, reducing the data volume, and at the same time maintaining the overall geometric features.

[0163] voxel_size: The side length of the voxel, which determines the accuracy of downsampling.

[0164] 2.4) Convert to a 3D mesh model or Digital Elevation Model (DEM)

[0165] If generating a 3D mesh model, a triangular mesh can be generated using Poisson reconstruction or the Ball Pivoting Algorithm (BPA) after point cloud normal estimation; if generating a DEM, a two-dimensional regular grid needs to be constructed according to the point cloud distribution, and the height (Z value) is calculated for each grid cell (pixel). Common methods include nearest neighbor or weighted average.

[0166] 3) Data Alignment and Personalized Model Generation

[0167] Use the Iterative Closest Point (ICP) algorithm to spatially align the patient's 3D data with the male mold model retrieved in step 3 to ensure that the two are in the same coordinate system. During the alignment process, the system automatically extracts key points (such as the scapula, the vertex of the spine, etc.) for auxiliary registration. After alignment, the deep learning module (using a dedicated regression network) further analyzes the point cloud data, automatically identifies the local curvature of the spine, the rotation angle, and the relative position between the spine and the pelvis. Finally, a 3D reference model of the patient's personalized spinal brace is generated. This model retains the unique physiological curve information of the patient and generates an editable 3D file in CAD software.

[0168] 3.1 Automatic Key Point Extraction and Preliminary Registration

[0169] To improve the accuracy of subsequent ICP (Iterative Closest Point) algorithm registration, the system automatically extracts several key anatomical points from the patient's point cloud data, such as the edges of the scapula and the vertices of the spine. These key points have high anatomical landmarks and can be used as references for auxiliary registration. The extracted key points are initially matched with the corresponding key points in the reference male mold model to preliminarily determine the relative position and orientation between the two sets of data, thereby providing a relatively reasonable initial transformation matrix. This initial transformation matrix helps the subsequent ICP algorithm to accurately optimize the registration result within a smaller search range.

[0170] 3.2 Using the Iterative Closest Point (ICP) algorithm to achieve fine registration

[0171] The system uses the ICP algorithm to perform fine registration on the preprocessed patient point cloud data and the point cloud data of the male mold model. The specific process is as follows:

[0172] Initialization: Using the aforementioned automatic key point extraction and preliminary registration results, an initial rigid transformation matrix is obtained to roughly place the patient point cloud data in the same coordinate system as the reference male mold.

[0173] Iterative optimization: Based on the initial transformation, the ICP algorithm repeatedly performs the following steps: For each point in the patient point cloud, find the corresponding point in the reference male mold that is closest to it; Calculate the rigid transformation matrix (including rotation and translation) that minimizes the registration error based on all corresponding point pairs, so as to maximize the coincidence degree of the two sets of point clouds; Apply this rigid transformation to the patient point cloud data and iterate and update until the error drops to a preset threshold or reaches the maximum number of iterations.

[0174] Through the above steps, the system can automatically solve an accurate ICP registration transformation matrix to accurately align the patient point cloud data and the male mold model in space, ensuring that the two are in the same coordinate system.

[0175] 3.3 Feature extraction of the aligned point cloud by the deep learning module

[0176] After the alignment is completed, the system will use a pre-trained deep learning model to further analyze the patient's point cloud data and extract geometric parameters that are crucial for personalized brace design. This deep learning module usually adopts a point cloud-based regression network (such as an architecture based on PointNet or PointNet++), and its specific functions include:

[0177] Local curvature extraction: Automatically identify the local bending conditions of different regions of the spine to obtain the curvature distribution information of each key region;

[0178] Measurement of the overall rotation angle: Calculate the overall rotation angle of the patient's spine relative to the vertical axis or other reference axes to reflect the lateral rotation state of the spine;

[0179] Determination of the relative position between the spine and the pelvis: Extract the spatial relationship between the spine and the pelvis, usually described by a 3D displacement vector or relative coordinates, which reflects the precise relative position between the two.

[0180] Through the analysis of the deep learning module, the system can obtain a series of numerical parameters that truly reflect the physiological structure and characteristics of the patient's spine.

[0181] 3.4 Generation of personalized brace model

[0182] Based on the parameters extracted by the deep learning module, the system makes personalized adjustments to the pre-stored standard brace male mold to generate a dedicated three-dimensional reference model of the brace suitable for the patient. The specific method is as follows:

[0183] a) Global geometric transformation: Using the rotation angle and relative displacement information output by the deep learning module, perform a rigid transformation on the brace male mold, that is, rotate first and then translate, so that the brace model matches the patient's spine in the overall position.

[0184] b) Local shape adjustment: On the basis of the rigid transformation, according to the local curvature information of the spine, make fine adjustments to the local structure of the brace male mold. This process can use Free-Form Deformation (FFD) technology or a mesh deformation algorithm based on physical simulation to make the key support surfaces of the brace model fit the actual curve of the patient's spine more closely.

[0185] c) Generate an editable 3D file: After the global and local adjustments, the finally generated personalized brace model retains the unique physiological curve information of the patient. This model will be converted into a CAD-compatible 3D file format (such as STL or STEP file) for further detailed modification in CAD software or directly used for manufacturing and processing.

[0186] Step 5: Optimization of the brace model and adjustment of design parameters

[0187] 1) Optimization algorithm and parameter adjustment

[0188] Based on the personalized model, use a model optimization algorithm (genetic algorithm or particle swarm optimization algorithm) to automatically adjust the various design parameters of the brace male mold model, including size, geometric curvature, and pressure point distribution. During the optimization process, introduce feedback data (such as the patient's historical treatment data, the usage record of the last worn brace) as a weight factor to make the optimization result more in line with the actual needs of the patient.

[0189] The following provides a detailed textual description of an optimization and parameter adjustment method based on the Genetic Algorithm (GA). This method aims to automatically optimize and adjust the key design parameters (such as size, geometric curvature, and pressure point distribution) of the model based on a preliminarily generated personalized brace model, so that the final design better meets the actual needs of the patient. At the same time, combined with the patient's historical treatment data and wearing records as feedback information, the weight factors in the objective function are adjusted. The specific implementation is described as follows:

[0190] 1.1 Parametric Representation and Design Variable Definition

[0191] First of all, it is necessary to parametrically describe the positive mold model of the brace, and abstract the design features of the model into a parameter vector. Let the design parameter vector be x = [x1, x2, …, xn]; where each component represents the following content respectively:

[0192] Size parameters: brace length, width, thickness, etc.;

[0193] Geometric curvature parameters: control point coordinates or curvature radii describing the bending shape of the brace;

[0194] Pressure point distribution parameters: define the position, size, and shape of the pressure distribution area where the brace contacts the patient's body.

[0195] The value ranges of these parameters should be preset with reasonable upper and lower bounds according to clinical requirements and manufacturing processes.

[0196] 1.2 Objective Function Design

[0197] The objective function (Fitness Function) is used to measure the gap between the candidate design scheme and the ideal design goal. This objective function is usually a weighted combination of multiple indicators, mainly considering the following aspects:

[0198] Geometric matching error: Calculate the matching error between the candidate brace model and the patient's physiological three-dimensional data (such as spinal curve, relative positions of scapula and pelvis) after preliminary alignment. The error can be measured by metrics such as Euclidean distance and curve fitting error, reflecting the fitting degree between the brace model and the patient's anatomical structure.

[0199] Biomechanical indicators: Calculate indicators such as stress distribution and pressure point rationality of the brace during wearing according to finite element analysis or empirical formulas, ensuring that the brace has sufficient supporting force and does not generate excessive pressure at key parts.

[0200] Feedback data constraint: Introduce historical treatment data and the last wearing record. For example, if the patient felt discomfort in a certain area during the last wearing, a higher penalty weight is assigned to the excessive pressure in that area in the objective function; if the patient's treatment progress is slow, the optimization direction can also be adjusted through the feedback data. The feedback data can be obtained through questionnaires, sensor monitoring data, or physician records, and is introduced into the objective function through preset weights.

[0201] Considering the above factors, the objective function can be expressed as:

[0202] J(x) = w 1 ·E geo (x) + w 2 ·E mech (x) + w 3 ·E feedback (x)

[0203] Where:

[0204] E geo (x) represents the geometric matching error;

[0205] E mech (x) represents the biomechanical performance error;

[0206] E feedback (x) represents the deviation penalty based on the feedback data;

[0207] w1, w2, w3 are the corresponding weight coefficients, and the selection of weights can be adjusted through expert experience or experiments.

[0208] 1.3 Implementation steps of the genetic algorithm

[0209] (1) Generation of the initial population

[0210] Randomly generate several candidate design solutions, and each candidate solution is an instance of a design parameter vector x. The population size (50 - 100 individuals) should be large enough to cover the design parameter space.

[0211] (2) Fitness evaluation

[0212] For each candidate solution in the population, calculate the fitness value according to the objective function J(x). The lower (or higher, depending on the definition method) the fitness value, the closer the candidate design is to the ideal goal.

[0213] (3) Selection operation

[0214] Adopt methods such as roulette wheel selection, tournament selection, or ranking selection to select individuals with higher fitness from the current population as the breeding pool. This can make excellent design solutions have a higher breeding probability.

[0215] (4) Crossover operation

[0216] Randomly select individuals in pairs from the breeding pool and perform gene crossover operations. The crossover method can adopt single-point crossover, multi-point crossover or uniform crossover, and exchange some design parameters of the parent individuals to generate new candidate solutions. This step aims to explore the parameter space and combine the excellent characteristics of different individuals.

[0217] (5) Mutation operation

[0218] Apply random mutations to the candidate solutions generated by crossover, that is, randomly adjust some design parameters with a certain low probability to prevent the population from premature convergence and local optimality. The mutation amplitude can set an appropriate step size according to the parameter value range.

[0219] (6) Generate a new generation of population and iterate

[0220] Combine the new candidate solutions after crossover and mutation with some excellent individuals (elitist retention) to form a new generation of population. Repeat the steps of "fitness evaluation - selection - crossover - mutation" until the termination condition is met (such as reaching the maximum number of generations or the fitness change is lower than the preset threshold).

[0221] 4. Introduction of feedback data

[0222] In the objective function, adjust the weight factor through feedback data to make the optimization process more in line with the actual needs of patients. The specific methods include:

[0223] Conduct statistical analysis on the patient's historical treatment data to determine which regions or parameters have caused discomfort;

[0224] According to the feedback record, impose stricter penalties on the corresponding design parameters (such as the thickness or local curvature of the brace in a specific area), and increase the weight of this part in the objective function;

[0225] Dynamically update the weights: During the optimization iteration process, the w3 or other weight parameters can be adjusted according to the latest clinical feedback information, so that the subsequent candidate solutions are more in line with the individual needs of patients.

[0226] 1.5. Output and application

[0227] After several generations of iteration, the genetic algorithm will converge to one or more optimal candidate solutions. The system selects the design parameter vector with the highest fitness (or the lowest error) as the final optimization result. Subsequently, apply this parameter vector to the CAD design of the personalized brace model to generate the final 3D model file. This model not only precisely matches the patient's anatomical structure in terms of geometric shape, but also is fully optimized in terms of pressure distribution and biomechanical performance to improve wearing comfort and treatment effect.

[0228] 2) Simulation and verification

[0229] Perform finite element simulation (FEA) on the optimized brace design to simulate the stress distribution during wearing and ensure that each key part can withstand the expected biomechanical load. According to the simulation results, automatically fine-tune the brace design again until the simulation indicators meet the preset requirements.

[0230] Step 6: Machining file generation and brace manufacturing

[0231] 1) Machining file generation

[0232] Convert the final optimized 3D model of the brace into standard machining files (such as STL, OBJ files), and automatically generate CNC machining paths or 3D printing slicing data. During the file generation process, automatically add necessary machining allowances, support structures, and post-processing process parameters (such as surface smoothing processing instructions).

[0233] 2) Manufacturing and post-processing

[0234] Use a 3D printer or CNC machining center to manufacture a physical brace according to the generated machining files. After manufacturing, smooth the edges of the brace through an automated post-processing device, and increase the local thickness of the brace or open hollow pattern holes if necessary to meet the biomechanical requirements.

[0235] Step 7: Brace installation, real-time adjustment, and feedback closed-loop

[0236] 1) Installation and preliminary debugging

[0237] Hand over the manufactured brace to a clinician for preliminary installation. Use intelligent sensing devices (pressure sensors, flexible sensors) to monitor the fit between the brace and the patient's body and collect real-time data.

[0238] 2) Closed-loop feedback adjustment

[0239] Feed the monitored data back to the feedback module of the system. The system automatically determines whether there are any discomfort or uneven pressure distribution in the local area of the brace based on the feedback data. If a problem is detected, automatically generate an adjustment plan (such as local shape fine-tuning, pressure point redistribution), and guide technicians or directly drive the manufacturing equipment to make subsequent corrections to the brace until the best fit effect is achieved.

[0240] Embodiment 2: Scoliosis Brace Design and Manufacturing System

[0241] Refer to Figure 2 , this embodiment details the hardware composition and software functions of each module of the system to achieve the automated operation of the above method.

[0242] 1. System hardware composition

[0243] 1) Data processing host

[0244] Adopt a high-performance computing platform, configure GPU acceleration cards (such as NVIDIA series) to support the real-time operation of deep learning models. The host is installed with a Linux or Windows operating system and runs dedicated control software to achieve data transmission and process control between modules.

[0245] 2) 3D scanning device

[0246] The built-in 3D scanning function of the iPad or a dedicated portable 3D scanner can be used, which supports high-definition point cloud acquisition and has data preprocessing functions.

[0247] 3) Image acquisition device

[0248] A digital X-ray acquisition instrument or a DICOM interface device to ensure the input of high-quality medical images.

[0249] 4) Manufacturing equipment

[0250] It includes a 3D printer (supporting SLA, SLS or FDM technology) and a numerically controlled CNC machine tool, which can achieve high-precision manufacturing of the brace entity according to the processing file.

[0251] 5) Intelligent sensing device

[0252] Equipped with pressure sensors and displacement sensors, which are used to detect the mechanical state on the contact surface between the brace and the patient in real time and wirelessly transmit the data to the feedback module.

[0253] 2. System software architecture

[0254] 1) Database management module

[0255] Responsible for the storage and retrieval of the brace male mold model and the patient's historical data, and achieving seamless docking with external data interfaces.

[0256] 2) Image processing and deep learning module

[0257] Integrate CNN and GAN network models to process X-ray images and 3D point cloud data. Figure 3 Details the overall process of data preprocessing, feature extraction, model prediction and subsequent data feedback.

[0258] 3) Data alignment and model generation module

[0259] Realize the automatic alignment of the patient's three-dimensional scan data and the male mold model, and generate a personalized three-dimensional reference model of the brace. Built-in ICP algorithm and adaptive data registration algorithm to ensure the alignment accuracy at the sub-millimeter level.

[0260] 4) Model optimization and simulation module

[0261] Optimize the brace design parameters using genetic algorithms, particle swarm algorithms, etc. Integrate finite element simulation software to conduct stress and deformation analysis on the brace design and achieve virtual verification.

[0262] 5) Manufacturing control and post - processing module

[0263] Control 3D printing or CNC machining equipment according to the generated machining files to achieve automated manufacturing. Support the configuration of post - processing process parameters (such as laser polishing, chemical polishing, etc.).

[0264] 6) Feedback and adjustment closed - loop module

[0265] Receive real - time data from intelligent sensing devices, compare the design parameters with the actual usage status. Automatically generate adjustment suggestions, and support secondary machining or brace adjustment after manual confirmation.

[0266] 3. System integration and data flow

[0267] As Figure 2 shown, data is transmitted between the system modules through a high - speed network:

[0268] Patient data (X - ray films, 3D scan point clouds) is transmitted to the data - processing host via the acquisition device;

[0269] The data - processing host calls the deep - learning module for image and point - cloud analysis;

[0270] The analysis results are transmitted to the database module for matching the male mold model;

[0271] The matched male mold model and patient data are synthesized into a personalized design in the data alignment module;

[0272] The optimized design is transmitted to the 3D printing / CNC equipment through the manufacturing control module to manufacture the physical brace;

[0273] After manufacturing, the intelligent sensing device collects real - time usage feedback data, and the closed - loop module guides subsequent adjustments.

[0274] The above is the description of the embodiments of the present invention. Through the above description of the disclosed embodiments, those skilled in the art can implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel points disclosed herein.

Claims

1. A method for designing and manufacturing a scoliosis brace, characterized in that: The method comprises the following steps: 1) Prefabrication and storage of brace positive mold models: A plurality of brace positive mold models customized according to different clinical scoliosis types are generated using three-dimensional CAD software. The positive mold model is designed based on the biomechanical principles to determine the support point, pressure distribution area and geometric curve, and applicable parameters are embedded in the model, including the Cobb angle range and Risser sign. The model and parameters are stored in a database in a standard format; 2) Patient X-ray processing and deep learning feature analysis: The patient's X-ray film is input into the system in DICOM format and preprocessed by grayscale normalization, noise filtering and contrast enhancement. The preprocessed image is then input into a pre-trained multi-task deep learning model to automatically output detection parameters reflecting the patient's scoliosis characteristics. 3) Automatically match the brace positive mold model: Based on the detection parameters obtained in step 2), including the predicted Cobb angle, Risser sign grade and scoliosis type, a brace positive mold model matching the patient's detection parameters is automatically retrieved from the database using a preset tolerance and a fuzzy matching algorithm; 4) Patient 3D scan data collection, preprocessing and point cloud alignment: Using a 3D scanning device to collect point cloud data of the patient's back and side, performing statistical outlier removal and voxel grid filtering on the data, and converting it into a three-dimensional grid model or a digital elevation model, and then using automatic extraction of key anatomical points and iterative closest point (ICP) algorithm to spatially align the patient's point cloud data with the positive model data retrieved in step 3); 5) Personalized model optimization and parameter adjustment: Based on the personalized model generated in step 4), the deep learning module and the model optimization algorithm are used to automatically adjust the design parameters of the brace positive mold model, wherein the design parameters include size, geometric curvature and pressure point distribution, and the patient's historical treatment data and brace wearing feedback data are introduced as weight factors during the optimization process to improve the adaptability of the model in terms of geometric matching and biomechanical performance; 6) Processing file generation and brace manufacturing: The personalized brace model optimized in step 5) is compared with the patient's three-dimensional data and X-ray images, and the design accuracy is further optimized through manual or automatic deep learning algorithms, and processing files for 3D printing or CNC processing are automatically generated; the scoliosis brace is manufactured using 3D printing or CNC processing technology, and is installed and adjusted so that the manufactured brace is suitable for the patient to wear.

2. The method according to claim 1, characterized in that The method also includes cutting the brace according to the patient's needs and smoothing the edges of the brace; and / or increasing the thickness and hollow pattern holes of the brace to meet specific biomechanical requirements; and / or using intelligent sensing equipment to monitor the adaptability of the brace to the patient's body in real time, and automatically adjusting the shape or pressure distribution of the brace through a feedback mechanism to enhance the wearing experience and treatment effect.

3. The method according to claim 1, characterized in that The multi-task deep learning model in step 2) comprises: (a) a regression branch for predicting the Cobb angle of the patient's spine; (b) a classification branch for determining the patient's Risser sign grade; (c) an auxiliary branch for determining the type of scoliosis.

4. The method according to claim 1, characterized in that: The matching algorithm described in step 3) is based on calculating the Euclidean distance between the predicted parameters and the parameters of each positive model in the database, and selecting the model with the smallest distance within a preset tolerance as the best matching model.

5. The method according to claim 1, characterized in that: The preprocessing of the 3D scan data in step 4) includes using a statistical outlier removal algorithm and voxel grid filtering to reduce point cloud noise and reduce data volume while retaining key geometric features of the patient's three-dimensional structure.

6. The method according to claim 1, characterized in that The model optimization step in step 5) uses a genetic algorithm to optimize the design parameters of the brace positive model, the genetic algorithm includes initial population generation, fitness evaluation, selection, crossover, mutation and new generation population generation steps, and the objective function is a weighted combination of the geometric matching error, biomechanical performance error and deviation penalty based on feedback data of each candidate design scheme; And / or, in the automatic feature detection and data output step, the deep learning model integrates the output results after X-ray film processing into JSON format as a basis for subsequent brace positive mold model matching.

7. A scoliosis brace design system, the system implementing the method of any one of claims 1 to 6, comprising: Database management module, used to store brace positive models and patient historical data; image processing and deep learning module, used to pre-process patient X-rays and automatically extract scoliosis features; 3D data acquisition module, used to collect 3D scanning point cloud data of patients; Model generation and data alignment module, used to perform preliminary and fine alignment of the patient's point cloud data with the brace positive mold model retrieved from the database to generate a personalized 3D reference model; Model optimization module, used to automatically adjust brace design parameters based on deep learning output parameters and feedback data; Manufacturing module for generating machining files and manufacturing scoliosis braces.

8. The system according to claim 7, characterized in that The image processing and deep learning module adopts a multi-task convolutional neural network based on ResNet or VGG architecture, which is used to predict the Cobb angle, determine the Risser sign grade and identify the scoliosis type respectively.

9. The system according to claim 7, characterized in that The model generation and data alignment module uses an iterative closest point (ICP) algorithm when aligning the patient's 3D scan data, and uses automatically extracted key anatomical points as an aid to achieve accurate matching of the patient's point cloud data and the brace positive model in the same coordinate system; And / or, the model optimization module dynamically adjusts the brace model in combination with the patient's historical data, including wearing feedback and correction progress, to improve the correction effect; And / or, the manufacturing module automatically generates a CNC machining path or 3D printing slice data according to the personalized brace three-dimensional model output by the model optimization module, and connects with the 3D printing or CNC processing equipment to realize automatic manufacturing of the brace and subsequent automated post-processing.

10. The system according to claim 7, characterized in that The system also includes an intelligent sensing module for real-time monitoring of the fit of the brace to the patient's body, and through a feedback mechanism, adjusts the shape or pressure distribution of the brace based on a deep learning algorithm to further optimize the wearing experience.

Images