A digital design method, medium and system for scoliosis orthopedic brace

Abstract

The present invention provides a digital design method, medium and system for scoliosis orthopedic braces, which belongs to the technical field of scoliosis orthopedic brace design, and includes: firstly extracting the contour image of each shooting angle and compressing it into a two-dimensional matrix. Then inputting these contour matrices into a pre-trained model to obtain the corresponding three-dimensional contour matrix. Next, traversing these three-dimensional matrices, rotating them according to the shooting angle to obtain the three-dimensional theoretical matrix of each angle, and performing similarity calculation with the corresponding three-dimensional generation matrix. Selecting the three-dimensional generation matrix with the highest similarity as a candidate, and correcting it using a difference regression function to obtain the final three-dimensional contour image. Finally, the algorithm inputs the three-dimensional contour image into a pre-trained scoliosis orthopedic brace parameter model, and outputs a three-dimensional model with information such as support points and pressure points marked as a scoliosis orthopedic design scheme. The problem that the existing counting relies on the doctor's experience and is prone to operational errors is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of scoliosis orthopedic brace design, and in particular, relates to a scoliosis orthopedic brace digital design method, medium and system. Background Art

[0002] Scoliosis is a common spinal orthopedic disease, the main feature of which is the abnormal curvature of the spine in the sagittal and coronal planes. This deformity may be caused by a variety of reasons, such as congenital factors, trauma, muscle imbalance, etc. If scoliosis is not corrected in time, it will lead to serious physical deformity, affecting the patient's appearance and quality of life. At present, scoliosis braces are often used as the main non-surgical treatment in clinical practice.

[0003] Traditional scoliosis braces are handmade by professional orthotists according to the specific conditions of the patients. The production process usually includes: first, the doctor needs to evaluate the deformity and deformation trend of the patient's spine based on medical images such as X-rays, CT scans or MRIs; then, the orthotist needs to use plaster or other materials to make a brace model that fits the spine contour for the patient according to the doctor's advice; finally, the model is converted into the final orthopedic brace, and the key parts of the brace such as support points and pressure points are adjusted according to the doctor's advice. The production process of this traditional brace production method requires the doctor to make a professional judgment on the patient's spinal condition, and the orthotist hand-makes the brace model based on experience, which requires high professional quality of personnel and is prone to operational errors. Summary of the invention

[0004] In view of this, the present invention provides a digital design method, medium and system for a scoliosis orthopedic brace, which can solve the technical problem that the existing method needs to rely on the doctor's experience and is prone to operating errors.

[0005] The present invention is achieved in that:

[0006] A first aspect of the present invention provides a method for digitally designing a scoliosis orthopedic brace, comprising the following steps:

[0007] S10, obtaining medical images of the patient's spine from at least two shooting angles;

[0008] S20, extracting the contour of the medical image at each shooting angle to obtain a contour image corresponding to the shooting angle;

[0009] S30, for each contour image, after size registration, set the areas without contours in the contour image to 0, set the areas with contours to 1, and compress them using run-length encoding to obtain a contour two-dimensional matrix;

[0010] S40, inputting each contour matrix into a pre-trained spine three-dimensional contour matrix generation model to obtain a corresponding contour three-dimensional matrix, which is recorded as a three-dimensional generation matrix;

[0011] S50, traversing each three-dimensional generator matrix as a basic generator matrix, rotating it according to the shooting angles of all medical images, obtaining a contour three-dimensional matrix of each shooting angle, recorded as a three-dimensional theoretical matrix; calculating the similarity between the three-dimensional theoretical matrix of each shooting angle and the three-dimensional generator matrix of the corresponding angle; after the traversal is completed, obtaining a similarity array of each three-dimensional generator matrix;

[0012] S60, selecting elements of the similarity array and the largest three-dimensional generator matrix as candidate generator matrices, performing two-dimensional extraction according to the shooting angles of all medical images, and obtaining a two-dimensional theoretical matrix for each shooting angle;

[0013] S70, calculating the difference regression function of each shooting angle based on the two-dimensional theoretical matrix and the corresponding contour two-dimensional matrix of each shooting angle; and using the difference regression function of each shooting angle, modifying the candidate generation matrix to obtain a target generation matrix;

[0014] S80, expanding the target generation matrix by using a run-length encoding inverse operation to obtain a three-dimensional contour image of the patient's spine;

[0015] S90, inputting the three-dimensional contour image of the patient's spine into a pre-trained scoliosis orthosis brace parameter model, obtaining a three-dimensional model of the inner surface of the scoliosis orthosis brace marked with support points, pressure points, release points, elastic points, buffer points, and traction points as a design solution and outputting it.

[0016] The support point is located on the inner surface of the brace, and is used to provide support for the spine to ensure the correction effect; the position and size of the support point will affect the way and intensity of applying the correction force.

[0017] The pressure points are located on the inner surface of the brace and are used to apply corrective pressure, thereby changing the shape of the spine; the position of the pressure points and the magnitude of the pressure are the key to achieving correction.

[0018] The release point is located on the inner surface of the brace, and is used to relieve pressure on specific parts of the spine and improve wearing comfort; the release point is designed to reduce the stimulation of the brace to the skin and muscles.

[0019] The elastic point is located on the inner surface of the brace and is made of elastic material to provide moderate elasticity to follow the deformation of the spine; the elastic point is used to increase the flexibility and fit of the brace and improve wearing comfort.

[0020] The buffer point is located on the inner surface of the brace and is made of soft material, and is used to reduce the pressure concentration when the brace contacts the skin; the buffer point is used to reduce the pressure of the brace on the skin and improve wearing comfort.

[0021] The traction point is located on the inner surface of the brace and is used to apply a corrective traction force to straighten the spine. The position and strength of the traction point will affect the correction effect.

[0022] The structure of the scoliosis orthosis brace parameter model adopts a deep neural network, including a convolution layer, a pooling layer and a fully connected layer, and the steps of obtaining training data specifically include:

[0023] Collect a large number of patients' 3D spinal contour data and corresponding orthotic brace design solutions;

[0024] Parametrically process the design scheme and extract key point positions and pressure parameters;

[0025] Pair the spinal contour data and brace parameters to form training samples;

[0026] Perform data augmentation and normalization on samples.

[0027] The training steps specifically include:

[0028] Step 1: Initialize network parameters;

[0029] Step 2: Use training data for forward propagation;

[0030] Step 3, calculating the loss between the predicted brace parameters and the true parameters;

[0031] Step 4: Back propagation updates network parameters;

[0032] Step 5: Repeat steps 2-4 until convergence or the preset number of iterations is reached;

[0033] Step 6: Use the validation set to evaluate model performance and adjust the network structure or hyperparameters if necessary.

[0034] On the basis of the above technical solution, the digital design method of a scoliosis orthopedic brace of the present invention can also be improved as follows:

[0035] The medical image is an X-ray, a CT scan or an MRI scan.

[0036] Furthermore, the method for extracting contours of the medical image at each shooting angle adopts a Canny edge detection algorithm or an active contour model algorithm.

[0037] Furthermore, the steps of calculating the similarity between the three-dimensional theoretical matrix of each shooting angle and the three-dimensional generated matrix of the corresponding angle are specifically:

[0038] Calculate the Euclidean distance between two matrices;

[0039] Calculate the cosine similarity of two matrices;

[0040] Calculate the structural similarity index of two matrices;

[0041] Combining the above three indicators, the weighted average is taken as the final similarity.

[0042] Among them, for each contour image, the step of performing size registration specifically includes:

[0043] Determine the key anatomical landmarks in the Zhanglun image, including the vertebral margin and intervertebral disc;

[0044] Calculate the zoom ratio and rotation angle of the wheel image according to the marker points;

[0045] According to the calculated image scaling ratio and rotation angle, the Zhanglun image is geometrically transformed accordingly to align it with the standard template;

[0046] The bilinear interpolation method is used to resample the images so that all the images have the same pixel resolution.

[0047] The structure of the spine three-dimensional contour matrix generation model adopts a 3D convolutional neural network, and the steps of training data acquisition specifically include:

[0048] Collect a large amount of spinal CT or MRI scan data;

[0049] Professional doctors will annotate the spinal contours on the scan data;

[0050] Perform data enhancement on the labeled data, such as rotation, scaling, and adding noise;

[0051] The processed data is divided into training set, validation set and test set.

[0052] The training steps specifically include:

[0053] Step 1: Initialize network parameters;

[0054] Step 2: Use training data for forward propagation;

[0055] Step 3: Calculate the loss between the predicted result and the true label;

[0056] Step 4: Back propagation updates network parameters;

[0057] Step 5: Repeat steps 2-4 until convergence or the preset number of iterations is reached;

[0058] Step 6: Use the validation set to evaluate model performance and adjust hyperparameters if necessary.

[0059] The basic generation matrix is ​​rotated by using a three-dimensional rotation matrix transformation method, and the specific steps include:

[0060] Calculate the rotation angle based on the shooting angle;

[0061] Construct the corresponding rotation matrix;

[0062] Transforms each point in the base generator matrix by applying a rotation matrix.

[0063] The candidate generation matrix is ​​subjected to two-dimensional extraction according to the shooting angle, and the specific steps include:

[0064] Determine the projection plane based on the shooting angle;

[0065] Perform projection transformation on each point in the three-dimensional matrix;

[0066] The projected points are gathered into a two-dimensional matrix;

[0067] Interpolate and smooth two-dimensional matrices;

[0068] Calculating the difference regression function of the shooting angles, the specific steps include:

[0069] Calculate the pixel-level difference between the 2D theoretical matrix and the 2D matrix of the contour;

[0070] Statistical analysis was performed on the differences to obtain the difference distribution;

[0071] Use polynomial regression or support vector regression method to fit the difference distribution and obtain the regression function.

[0072] Wherein, the step S10 specifically includes: obtaining medical images of the patient's spine at least two different angles, and commonly used medical imaging methods include X-rays, CT scans and MRI scans, etc. These technologies can obtain two-dimensional images of the patient's spine at different angles. For the same patient, it is usually necessary to collect image data from multiple different angles to fully describe the three-dimensional morphology of the spine.

[0073] Wherein, the step S20 specifically includes: extracting the contour of the medical image at each shooting angle to obtain the contour image of the corresponding shooting angle. The contour extraction method can adopt the Canny edge detection algorithm or the active contour model (Snake algorithm), etc. These algorithms can effectively segment the edge contour of the spine from the medical image.

[0074] Wherein, the step S30 specifically includes: for each contour image, after size registration, the areas without contours in the contour image are set to 0, the areas with contours are set to 1, and run-length encoding is used for compression to obtain a contour two-dimensional matrix. The specific steps of size registration include: determining key anatomical landmarks in the image, calculating the image scaling and rotation angle, geometrically transforming the image to align it with the standard template, and resampling the image using bilinear interpolation.

[0075] Wherein, the step S40 specifically includes: inputting each contour matrix into a pre-trained spine three-dimensional contour matrix generation model to obtain a corresponding contour three-dimensional matrix, which is recorded as a three-dimensional generation matrix. The spine three-dimensional contour matrix generation model adopts a 3D convolutional neural network structure, and the training data includes a large amount of spine CT or MRI scan data, spine contours annotated by professional doctors, and the data is enhanced and divided.

[0076] Among them, the step S50 specifically includes: traversing each three-dimensional generation matrix as a basic generation matrix, rotating it according to the shooting angles of all medical images, and obtaining a contour three-dimensional matrix of each shooting angle, recorded as a three-dimensional theoretical matrix; calculating the similarity between the three-dimensional theoretical matrix of each shooting angle and the three-dimensional generation matrix of the corresponding angle, including Euclidean distance, cosine similarity and structural similarity index (SSIM), etc., and comprehensively obtaining the final similarity.

[0077] Wherein, the step S60 specifically includes: selecting the similarity array elements and the largest three-dimensional generation matrix as candidate generation matrices, performing two-dimensional extraction according to the shooting angles of all medical images, and obtaining a two-dimensional theoretical matrix for each shooting angle. The two-dimensional extraction step includes: determining the projection plane, performing projection transformation on the points in the three-dimensional matrix, gathering the projected points into a two-dimensional matrix, and performing interpolation and smoothing processing.

[0078] Wherein, the step S70 specifically includes: based on the two-dimensional theoretical matrix of each shooting angle and the corresponding contour two-dimensional matrix, calculating the difference regression function of the shooting angle; using these difference regression functions, modifying the candidate generation matrix to obtain the target generation matrix. The calculation of the difference regression function includes: calculating the pixel-level difference between the two-dimensional theoretical matrix and the contour matrix, performing statistical analysis on the difference and fitting the regression function.

[0079] Wherein, the step S80 specifically includes: applying the inverse operation of run-length encoding to the target generation matrix to expand it into a complete three-dimensional contour image. The step S90 specifically includes: inputting the three-dimensional contour image of the patient's spine into a pre-trained scoliosis orthotic brace parameter model, which uses a deep neural network structure to automatically generate a three-dimensional orthotic brace model with key points annotated as a design solution.

[0080] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned digital design method of a scoliosis orthopedic brace.

[0081] A third aspect of the present invention provides a digital design system for a scoliosis orthopedic brace, which includes the above-mentioned computer-readable storage medium.

[0082] Compared with the prior art, the digital design method, medium and system of scoliosis brace provided by the present invention utilizes advanced 3D reconstruction algorithms to extract the 3D contour of the patient's spine from medical images, and inputs it into a pre-trained parametric brace design model to automatically generate a personalized brace solution that fits the spine. Compared with traditional manual production, the solution of the present invention has the following technical effects:

[0083] 1. Improved the fit and comfort of the brace. Through 3D scanning and modeling technology, the overall shape of the patient's spine can be captured more accurately, and a brace that better fits the spine contour can be designed to reduce local pressure concentration and improve wearing comfort.

[0084] 2. Shortened the brace production cycle. Since the design process is digitized and automated, the time from medical image acquisition to brace manufacturing can be greatly shortened, which is conducive to timely correction of spinal deformity.

[0085] 3. Achieved quantitative evaluation of the correction effect. The digital design method can objectively analyze the pressure distribution and force conditions of the brace, providing a reliable basis for the quantitative evaluation of the correction effect, which is conducive to the targeted optimization of the brace design.

[0086] 4. Reduced dependence on professionals. The digital design process is more standardized and does not need to rely too much on the experience and judgment of doctors and orthotists, which can improve production efficiency and stability.

[0087] In summary, the present invention solves the technical problem that the existing method needs to rely on the doctor's experience and is prone to operating errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 A flow chart of the method provided by the present invention;

[0089] Figure 2 It is the spine contour extraction diagram of the anteroposterior X-ray;

[0090] Figure 3 The resulting image is a 3D reconstruction of the patient's spine. DETAILED DESCRIPTION

[0091] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0092] like Figure 1 FIG. 1 is a flow chart of a digital design method for a scoliosis orthopedic brace provided by the first aspect of the present invention. The method comprises the following steps:

[0093] S10, obtaining medical images of the patient's spine from at least two shooting angles;

[0094] S20, extracting the contour of the medical image at each shooting angle to obtain a contour image corresponding to the shooting angle;

[0095] S30, for each contour image, after size registration, set the areas without contours in the contour image to 0, set the areas with contours to 1, and compress them using run-length encoding to obtain a contour two-dimensional matrix;

[0096] S40, inputting each contour matrix into a pre-trained spine three-dimensional contour matrix generation model to obtain a corresponding contour three-dimensional matrix, which is recorded as a three-dimensional generation matrix;

[0097] S50, traversing each three-dimensional generator matrix as a basic generator matrix, rotating it according to the shooting angles of all medical images, obtaining a contour three-dimensional matrix of each shooting angle, recorded as a three-dimensional theoretical matrix; calculating the similarity between the three-dimensional theoretical matrix of each shooting angle and the three-dimensional generator matrix of the corresponding angle; after the traversal is completed, obtaining a similarity array of each three-dimensional generator matrix;

[0098] S60, selecting elements of the similarity array and the largest three-dimensional generator matrix as candidate generator matrices, performing two-dimensional extraction according to the shooting angles of all medical images, and obtaining a two-dimensional theoretical matrix for each shooting angle;

[0099] S70, calculating the difference regression function of the shooting angles based on the two-dimensional theoretical matrix of each shooting angle and the corresponding two-dimensional contour matrix; and using the difference regression function of each shooting angle, correcting the candidate generation matrix to obtain the target generation matrix;

[0100] S80, expanding the target generation matrix by means of run-length coding inverse operation to obtain a three-dimensional contour image of the patient's spine;

[0101] S90, input the three-dimensional contour image of the patient's spine into a pre-trained scoliosis orthosis brace parameter model, obtain a three-dimensional model of the inner surface of the scoliosis orthosis brace marked with support points, pressure points, release points, elastic points, buffer points, and traction points as a design solution and output it.

[0102] The specific implementation methods of the above steps are described in detail below:

[0103] Step S10, obtaining medical images of the patient's spine from at least two shooting angles. The purpose of this step is to collect three-dimensional information of the patient's spine to provide a data basis for subsequent three-dimensional reconstruction and brace design. Common medical imaging methods include X-rays, CT scans, and MRI scans, which can obtain two-dimensional images of the patient's spine at different angles. For the same patient, it is usually necessary to collect imaging data from multiple different angles to fully describe the three-dimensional morphology of the spine.

[0104] Step S20, extract the contour of the medical image at each shooting angle to obtain the contour image of the corresponding shooting angle. The purpose of this step is to extract the contour information of the spine from the two-dimensional medical image, laying the foundation for the subsequent three-dimensional reconstruction. The contour extraction method can adopt the Canny edge detection algorithm or the active contour model (Snake algorithm), etc. These algorithms can effectively segment the edge contour of the spine from the medical image.

[0105] Step S30, for each contour image, after size registration, the areas without contours in the contour image are set to 0, the areas with contours are set to 1, and run-length encoding is used to compress the contour image to obtain a contour two-dimensional matrix. The purpose of this step is to standardize and compress the extracted two-dimensional contour image to provide a unified data format for subsequent three-dimensional reconstruction and feature matching. The specific steps of size registration include: determining key anatomical landmarks in the image, such as vertebral edges, intervertebral discs, etc.; calculating the image scaling and rotation angle based on these landmarks; performing corresponding geometric transformations on the image to align it with the standard template; and resampling the image using bilinear interpolation so that all images have the same pixel resolution. The areas without contours in the registered image are set to 0, the areas with contours are set to 1, and then run-length encoding is used to compress the image to obtain a compact contour two-dimensional matrix.

[0106] Step S40, input each contour matrix into a pre-trained spine three-dimensional contour matrix generation model to obtain a corresponding contour three-dimensional matrix, which is recorded as a three-dimensional generation matrix. The purpose of this step is to use a deep learning model to reconstruct a corresponding three-dimensional contour matrix from a two-dimensional contour matrix. The structure of the spine three-dimensional contour matrix generation model adopts a 3D convolutional neural network, and the training data acquisition step includes: collecting a large amount of spine CT or MRI scan data, having professional doctors annotate the scan data with spine contours, performing data enhancement on the annotated data, such as rotation, scaling, adding noise, etc., and dividing the processed data into a training set, a validation set, and a test set. The training step includes: initializing network parameters, using training data for forward propagation, calculating the loss between the predicted result and the true label, back-propagating to update the network parameters, repeating the above steps until convergence or reaching a preset number of iterations, and finally using the validation set to evaluate the model performance and adjust the hyperparameters if necessary. Through such a training process, the model can learn the mapping relationship from the two-dimensional contour matrix to the three-dimensional contour matrix, thereby realizing the three-dimensional reconstruction of the input two-dimensional contour matrix.

[0107] Step S50, traverse each three-dimensional generator matrix as the basic generator matrix, rotate it according to the shooting angle of all medical images, obtain the contour three-dimensional matrix of each shooting angle, recorded as the three-dimensional theoretical matrix; calculate the similarity between the three-dimensional theoretical matrix of each shooting angle and the three-dimensional generator matrix of the corresponding angle; after the traversal is completed, obtain the similarity array of each three-dimensional generator matrix. The purpose of this step is to evaluate the degree of consistency between different three-dimensional generator matrices and actual medical images, and provide a basis for the subsequent selection of the best three-dimensional matrix. The specific steps of rotation include: calculating the rotation angle according to the shooting angle, constructing the corresponding rotation matrix, and applying the rotation matrix to transform each point in the basic generator matrix. The calculation of similarity includes: calculating the Euclidean distance, cosine similarity and structural similarity index (SSIM) of the two matrices, and taking the weighted average of these three indicators as the final similarity.

[0108] Step S60, select the elements of the similarity array and the largest three-dimensional generator matrix as candidate generator matrices, perform two-dimensional extraction according to the shooting angles of all medical images, and obtain a two-dimensional theoretical matrix for each shooting angle. The purpose of this step is to select the best candidate matrix from all three-dimensional generator matrices and project it onto a two-dimensional plane according to the shooting angle so as to compare it with the original two-dimensional contour matrix. The specific steps of two-dimensional extraction include: determining the projection plane according to the shooting angle, performing projection transformation on each point in the three-dimensional matrix, gathering the projected points into a two-dimensional matrix, and interpolating and smoothing the two-dimensional matrix.

[0109] Step S70, based on the two-dimensional theoretical matrix of each shooting angle and the corresponding contour two-dimensional matrix, calculate the difference regression function of the shooting angle; and use the difference regression function of each shooting angle to correct the candidate generation matrix to obtain the target generation matrix. The purpose of this step is to further optimize the three-dimensional candidate generation matrix by analyzing the difference between the two-dimensional theoretical matrix and the measured two-dimensional contour matrix, so that it better fits the actual spinal morphology. The calculation of the difference regression function includes: calculating the pixel-level difference between the two-dimensional theoretical matrix and the contour two-dimensional matrix, performing statistical analysis on the difference to obtain the difference distribution, and fitting the difference distribution using methods such as polynomial regression or support vector regression to obtain the regression function. Using these difference regression functions, the candidate generation matrix can be corrected to obtain a target generation matrix that is more in line with reality.

[0110] Step S80, the target generation matrix is ​​expanded by the inverse operation of run-length coding to obtain a three-dimensional contour image of the patient's spine. The purpose of this step is to restore the compressed three-dimensional contour matrix to a complete three-dimensional contour model, and provide the required three-dimensional spine morphology information for subsequent brace design. The specific steps are: apply the inverse operation process of run-length coding to the target generation matrix to expand it into a complete three-dimensional contour image.

[0111] Step S90, input the three-dimensional contour image of the patient's spine into a pre-trained scoliosis orthopedic brace parameter model, obtain a three-dimensional model of the inner surface of the scoliosis orthopedic brace marked with support points, pressure points, release points, elastic points, buffer points, and traction points as a design plan and output it. The purpose of this step is to automatically generate a suitable orthopedic brace design plan based on the three-dimensional contour information of the patient's spine using a deep learning model. The structure of the scoliosis orthopedic brace parameter model adopts a deep neural network, including a convolution layer, a pooling layer, and a fully connected layer. The training data acquisition step includes: collecting a large number of patients' spine three-dimensional contour data and corresponding orthopedic brace design plans, parameterizing the design plans, extracting key point positions and pressure parameters, pairing the spine contour data and brace parameters to form training samples, and performing data enhancement and normalization on the samples. The training step includes: initializing network parameters, using training data for forward propagation, calculating the loss between the predicted brace parameters and the actual parameters, back-propagating to update the network parameters, repeating the above steps until convergence or reaching a preset number of iterations, and finally using a validation set to evaluate the model performance, and adjusting the network structure or hyperparameters if necessary. Through such a training process, the model can learn the mapping relationship from the three-dimensional contour of the spine to the brace design parameters, thereby automatically generating an orthopedic brace design plan that meets the patient's needs.

[0112] In summary, this digital design method for scoliosis braces includes multiple key steps, including medical image acquisition, contour extraction, 3D reconstruction, feature analysis, parameter optimization, etc. Through advanced algorithms such as deep learning, the entire process from 2D medical images to 3D spinal models to personalized brace design can be digitized, greatly improving the efficiency and accuracy of the design process.

[0113] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned digital design method of a scoliosis orthopedic brace.

[0114] A third aspect of the present invention provides a digital design system for a scoliosis orthopedic brace, which includes the above-mentioned computer-readable storage medium.

[0115] Specifically, the principle of the present invention is to use deep learning technology to achieve rapid reconstruction of three-dimensional spine contour information from two-dimensional medical images, and automatic generation of personalized brace parameters based on three-dimensional spine models. Its technical principle mainly includes the following key steps:

[0116] 1. Two-dimensional contour extraction and size registration. First, this method needs to extract the contour information of the spine from the patient's two-dimensional X-ray or CT / MRI images. Here, algorithms such as Canny edge detection can be used to preprocess the medical images and extract clear spine contour images. For subsequent three-dimensional reconstruction and feature matching, these contour images need to be size-aligned, including determining key anatomical landmarks, calculating scaling and rotation angles, performing geometric transformations, etc., so that all images have the same standardized scale.

[0117] 2. Three-dimensional spine contour reconstruction. After obtaining the standardized two-dimensional contour image, this method inputs it into a pre-trained three-dimensional contour generation model, and uses the powerful representation ability of deep learning to restore the three-dimensional contour matrix of the patient's spine from the two-dimensional information. Here, a 3D convolutional neural network is used as the reconstruction model. Through supervised learning of a large amount of spinal CT / MRI data, the model can effectively learn the mapping relationship from the two-dimensional contour to the three-dimensional spinal morphology.

[0118] 3. Optimization of the three-dimensional contour matrix. In order to further improve the accuracy of three-dimensional contour reconstruction, this method evaluates the similarity between each three-dimensional contour matrix and the original two-dimensional contour image, and selects the matrix with the highest similarity as the candidate result. Then, based on the two-dimensional contour image of each shooting angle, the difference between the two-dimensional projection of the candidate matrix and the measured contour is calculated, and the candidate matrix is ​​corrected using the difference regression function to obtain the final target three-dimensional contour matrix. This joint optimization method based on multi-view information can effectively improve the accuracy and robustness of three-dimensional reconstruction.

[0119] 4. Personalized brace parameter generation. With accurate three-dimensional spinal contour information, this method inputs it into a pre-trained scoliosis orthosis brace parameter generation model. The model uses a deep neural network structure and can automatically output a personalized brace three-dimensional model with key parameters such as support points, pressure points, and release points based on the input three-dimensional spinal contour. This end-to-end parametric design method greatly improves the automation and pertinence of brace design, and provides an important basis for subsequent brace manufacturing and wearing.

[0120] In general, the core innovation of the method of the present invention is to make full use of the powerful ability of deep learning in the field of computer vision, and realize the automation of the whole process from two-dimensional medical imaging to three-dimensional spine reconstruction and then to personalized brace design. Among them, the three-dimensional contour reconstruction model can effectively restore the precise three-dimensional shape of the spine from the two-dimensional information, and the personalized brace parameter generation model can automatically generate a brace design that meets the patient's needs based on the three-dimensional spine model.

[0121] In order to better understand and implement the present invention, a specific embodiment 1 of the present invention is provided below. The steps of this embodiment 1 are described as follows: First, in step S10, it is necessary to obtain medical images of at least two shooting angles of the patient's spine. Assume that the number of images obtained is N, and the size of each image is H×W pixels. For the image I of the i-th shooting angle i , which can be expressed as:

[0122] I i = {I i (x,y)|x=1,2,...,H; y=1,2,...,W};

[0123] Among them, I i (x, y) represents the gray value at the coordinate (x, y) in the i-th image. These medical imaging data provide the basis for subsequent three-dimensional reconstruction and brace design.

[0124] In step S20, it is necessary to extract the contour of the medical image at each shooting angle to obtain the contour image of the corresponding shooting angle. The Canny edge detection algorithm can be used here, and its basic steps are as follows:

[0125] 1. Gaussian filtering: Use Gaussian kernel function to filter the input image I i Perform smoothing filtering to reduce the impact of noise and obtain the smoothed image G i The Gaussian kernel function can be expressed as:

[0126]

[0127] Among them, σ is the standard deviation of the Gaussian kernel, which is a hyperparameter that needs to be adjusted.

[0128] 2. Calculate the gradient magnitude and direction: for the smoothed image G i Calculate the horizontal gradient G x and the vertical gradient G y , the gradient amplitude M is obtained by the following formula i and direction θ i :

[0129]

[0130] 3. Non-maximum suppression: along the gradient direction θ i For the gradient magnitude M i Suppress and obtain candidate edge points.

[0131] 4. Dual threshold detection: set two thresholds T h and T l , any gradient amplitude greater than T h Points with a value less than T are directly judged as edge points. l The points between the two are directly judged as non-edge points, and the points between them need further judgment.

[0132] 5. Hysteresis processing: For points between two thresholds, if they are connected to the determined edge points, they are also determined as edge points; otherwise, they are determined as non-edge points.

[0133] Through the above Canny algorithm, we can get i Extract the corresponding contour image C i , where C i (x,y)={0,1}, 1 represents a contour point, and 0 represents a non-contour point.

[0134] In step S30, each contour image C i Perform size registration and compress using run-length encoding to obtain the contour two-dimensional matrix M iThe specific steps of size registration include:

[0135] 1. Determine the key anatomical landmarks in the image, such as the vertebral edge, intervertebral disc, etc., and record them as {(x k ,y k )|k=1,2,...,K}.

[0136] 2. Calculate the image scaling s and rotation angle θ:

[0137]

[0138] Where (x′ 1 ,y′ 1 ) and (x′ 2 ,y′ 2 ) are the coordinates of the corresponding marking points on the standard template.

[0139] 3. For image C i Perform geometric transformation to align it with the standard template:

[0140]

[0141] Where (x 0 ,y 0 ) is the image translation amount.

[0142] 4. Resample image C′ using bilinear interpolation i , so that they have the same pixel resolution.

[0143] The registered contour image C′ i The places without contours are set to 0, and the places with contours are set to 1, and then run-length encoding is used for compression to obtain a compact contour two-dimensional matrix M. i :

[0144] M i ={(l j ,n j )|j=1,2,...,J};

[0145] Among them, l j represents the length of the jth run, n j Indicates the number of pixels with a value of 1 in this run.

[0146] In step S40, each contour matrix M needs to be i Input into the pre-trained spine 3D contour matrix generation model to obtain the corresponding contour 3D matrix T i , recorded as a three-dimensional generation matrix. The model adopts a 3D convolutional neural network structure, and its input is a two-dimensional contour matrix M i , the output is a three-dimensional contour matrix Ti .

[0147] The acquisition of training data includes:

[0148] 1. Collect a large amount of spinal CT or MRI scan data, denoted as {V k |k=1,2,...,K}, where V k The three-dimensional volume data representing the k-th volume data.

[0149] 2. The scan data is reviewed by a professional doctor k Manually mark the spine contour and obtain the corresponding three-dimensional contour model T k .

[0150] 3. For the labeled data {T k} Perform data enhancement, such as rotation, scaling, adding noise, etc., to obtain the enhanced training samples {T′ k}.

[0151] 4. Divide the augmented data into training set, validation set and test set.

[0152] The training sample is denoted as {(M i ,T i )}, where M i is the two-dimensional contour matrix, T i is the corresponding three-dimensional contour matrix. The training process of the neural network is as follows:

[0153] 1. Initialize network parameters Θ.

[0154] 2. Perform forward propagation on the training samples to obtain the predicted three-dimensional contour matrix

[0155]

[0156] Among them, f(·; Θ) is the mapping function of the neural network, and the parameter is Θ.

[0157] 3. Calculate the prediction results With the true label T i The loss function L between them, such as mean square error loss:

[0158]

[0159] 4. Back-propagate the loss function L and update the network parameters Θ to minimize the loss:

[0160]

[0161] Among them, η is the learning rate.

[0162] 5. Repeat steps 2-4 until convergence or the preset number of iterations is reached.

[0163] 6. Use the validation set to evaluate model performance and adjust the network structure or hyperparameters if necessary.

[0164] Through such a training process, the model can learn the mapping relationship from the two-dimensional contour matrix to the three-dimensional contour matrix, thereby realizing three-dimensional reconstruction of the input two-dimensional contour matrix.

[0165] In step S50, it is necessary to traverse each three-dimensional generator matrix T i As the basic generation matrix, it is rotated according to the shooting angle of all medical images to obtain the contour three-dimensional matrix of each shooting angle Recorded as the three-dimensional theoretical matrix; calculate the three-dimensional theoretical matrix for each shooting angle and the three-dimensional generator matrix T of the corresponding angle j After the traversal is completed, the similarity array of each three-dimensional generator matrix is ​​obtained {S i}.

[0166] The specific steps of rotation are as follows: Assume that the i-th three-dimensional generator matrix T i is the basic matrix, and the rotation angle of the jth shooting angle is θ j ,φ j ,ψ j , we can construct the following 3D rotation matrix R j :

[0167]

[0168] For the basic matrix T i Apply the rotation matrix R to each point (x, y, z) in j By transforming, we can get the three-dimensional theoretical matrix of the jth shooting angle

[0169]

[0170] Next, we need to calculate the three-dimensional theoretical matrix for each shooting angle and the three-dimensional generator matrix T corresponding to the angle j The similarity S i,j There are many similarity metrics that can be used here, such as Euclidean distance, cosine similarity, and structural similarity index (SSIM):

[0171] Euclidean distance:

[0172] Cosine similarity:

[0173] SSIM:

[0174] Among them, μ T and T j and The average value, and Variance is the covariance, c 1 and c 2 is a constant used to avoid the denominator being 0.

[0175] Combining the above three similarity indicators, we can get the final similarity S i,j :

[0176]

[0177] Among them, α, β, and γ are the weight coefficients of the corresponding indicators, which need to be adjusted according to actual conditions.

[0178] Traverse all three-dimensional generating matrices T i and shooting angle j, we can get the similarity array of each three-dimensional generation matrix {S i}.

[0179] In step S60, it is necessary to select a similarity array {S i} and the largest three-dimensional generator matrix T c As a candidate generation matrix, two-dimensional extraction is performed according to the shooting angles of all medical images to obtain the two-dimensional theoretical matrix for each shooting angle The specific steps of two-dimensional extraction are as follows:

[0180] 1. According to the rotation angle θ of the jth shooting angle j ,φ j ,ψ j , construct the corresponding 3D rotation matrix R j .

[0181] 2. Generate the candidate three-dimensional matrix T c Apply the rotation matrix R to each point (x,y,z) in j Transform and obtain the projected two-dimensional coordinates (x′, y′):

[0182]

[0183] 3. Collect all projected points into a two-dimensional matrix in Indicates the existence of contour points, Indicates that there is no contour point.

[0184] 4. For two-dimensional matrix Interpolation and smoothing are performed to obtain a more continuous and smooth two-dimensional theoretical contour matrix.

[0185] In step S70, a two-dimensional theoretical matrix based on each shooting angle is required. The corresponding contour two-dimensional matrix M j , calculate the difference regression function f of the shooting angle j , and use these difference regression functions to generate the candidate matrix T c Correction is performed to obtain the target generation matrix T * .

[0186] The calculation steps of the difference regression function are as follows:

[0187] 1. Calculate the two-dimensional theoretical matrix With the contour two-dimensional matrix M j The pixel-level difference between i,j :

[0188]

[0189] Where (x i ,y i ) represents the coordinates of the i-th pixel.

[0190] 2. For the difference {d i,j}Perform statistical analysis to obtain its distribution characteristics, such as histogram, mean and standard deviation.

[0191] 3. Use polynomial regression or support vector regression to fit the regression function f of the difference distribution j (d), This function can describe the difference relationship between the two-dimensional theoretical matrix and the measured contour matrix.

[0192] With the difference regression function of each shooting angle {f j}, we can generate the candidate matrix T c Modify it to get a more realistic target generation matrix T * The specific steps are as follows:

[0193] 1. According to the shooting angle j, T c Projected onto the corresponding two-dimensional plane, we get the two-dimensional theoretical matrix

[0194] 2. Calculation and the corresponding contour matrix M j The pixel-level difference d i,j .

[0195] 3. Using the difference regression function f j(d) Correct the difference of each pixel:

[0196] T * (x,y,z)=T c (x,y,z)+f j (d i,j );

[0197] Through this correction process, we can obtain a target generation matrix T that is more in line with the actual spine morphology. * .

[0198] In step S80, the target generation matrix T * Perform the inverse operation of run-length encoding to obtain a complete three-dimensional contour image of the patient's spine. The specific steps are as follows:

[0199] 1. Traverse T * The position (x,y,z) of each non-zero element in .

[0200] 2. For consecutive non-zero elements, record their length l and value n to form a run-length encoding sequence (l, n).

[0201] 3. Perform inverse operations on all run-length coding sequences (l,n) and expand them into a complete three-dimensional contour image.

[0202] In this way, the compressed three-dimensional contour matrix can be restored to a complete three-dimensional contour model, providing the required three-dimensional spinal morphology information for subsequent brace design.

[0203] Finally, in step S90, the three-dimensional contour image of the patient's spine needs to be input into a pre-trained scoliosis orthotic brace parameter model to obtain a three-dimensional model of the inner surface of the scoliosis orthotic brace marked with support points, pressure points, release points, elastic points, buffer points, and traction points as a design solution and output it.

[0204] The brace parameter model also uses a deep neural network structure, and the input is a three-dimensional contour image T * , the output is the key parameters of brace design, such as the location and pressure of various key points. The acquisition of training data includes:

[0205] 1. Collect 3D spine profile data from a large number of patients And the corresponding orthotic brace design {P k}.

[0206] 2. Design Scheme P k Parametric processing is performed to extract key parameters such as key point locations and pressure magnitudes.

[0207] 3. Spine profile data and brace parameter Pk Pairing to form training samples

[0208] 4. Perform data augmentation and normalization on training samples.

[0209] The training process of the network is similar to the above:

[0210] 1. Initialize network parameters Φ.

[0211] 2. Input spine outline image Get predicted brace parameters Where g(·;Φ) is the mapping function of the network.

[0212] 3. Calculate prediction parameters With the true parameter P k The loss function L between them, such as mean square error loss:

[0213]

[0214] 4. Back-propagation updates the network parameters Φ to minimize the loss function L.

[0215] 5. Repeat steps 2-4 until convergence or the preset number of iterations is reached.

[0216] 6. Use the validation set to evaluate model performance and adjust the network structure or hyperparameters if necessary.

[0217] Through such a training process, the model can learn the mapping relationship from the three-dimensional contour of the spine to the brace design parameters, thereby automatically generating an orthotic brace design plan that meets the patient's needs. The plan contains information such as the location of key points and the size of pressure, providing an important basis for subsequent brace manufacturing and wearing.

[0218] In order to further better understand and implement the present invention, the following is a specific application scenario of the present invention, Example 2: A 15-year-old patient was admitted to a hospital and was diagnosed with mild scoliosis. The doctor recommended the use of a non-surgical treatment method of orthopedic brace and required the production of a set of personalized orthopedic brace that fits the patient's spine contour.

[0219] According to the digital scoliosis orthopedic brace design method proposed in the present invention, the hospital's rehabilitation doctors and engineers began to implement a specific brace design process for the patients.

[0220] In the first step, the doctor conducted a comprehensive spinal examination for the patient and collected multimodal medical imaging data including anteroposterior X-rays, lateral X-rays, and CT scans. The dimensions of these imaging data are shown in Table 0:

[0221] Table 0 Image size table

[0222] Image type Size (pixels) Frontal X-ray 2048x1536 Lateral X-ray 2048x1536 CT scan 512x512x256

[0223] The doctor carefully examined the patient's spine and found that she had mild thoracic scoliosis and a Cobb angle of about 25 degrees. In order to design a suitable correction brace, it is necessary to obtain accurate three-dimensional morphological information of the patient's spine.

[0224] In the second step, the image data is processed in the brace design algorithm. First, the contour extraction and size registration of each two-dimensional image are performed. Taking the anteroposterior X-ray as an example, the Canny edge detection algorithm is used to pre-process it and extract the clear contour of the spine, such as Figure 2 shown.

[0225] In order to make these contour images have the same standardized scale, size alignment is required. The specific steps are as follows:

[0226] 1. Determine the key anatomical landmarks in the image, such as the vertebral edge, intervertebral disc, etc., and mark 10 main key points.

[0227] 2. Based on the coordinates of these key points, the image scaling ratio is calculated to be 1.2 and the rotation angle is 5 degrees.

[0228] 3. Perform geometric transformation on the original X-ray image to align it with the standard template.

[0229] 4. Use bilinear interpolation method to resample the transformed image to obtain a standardized contour image with a resolution of 1024x768.

[0230] After the above preprocessing, the contour images of all two-dimensional images are converted into a unified standard scale, preparing for subsequent three-dimensional reconstruction.

[0231] In the third step, the standardized two-dimensional contour image is input into a pre-trained three-dimensional spinal contour reconstruction model. This model uses a 3D convolutional neural network structure and can effectively restore the patient's spine's precise three-dimensional morphology from two-dimensional information.

[0232] When training the reconstruction model, the research team collected a large amount of spinal CT / MRI scan data, and professional doctors manually annotated these three-dimensional volume data to obtain the corresponding three-dimensional model of the spine contour. After data enhancement and other processing, a total of 1,000 training samples were formed. During the training process, the model's loss function uses mean square error, and the network parameters are continuously optimized through back propagation until the reconstruction accuracy on the validation set meets the requirements.

[0233] By inputting the patient's two-dimensional contour image into the trained three-dimensional reconstruction model, the three-dimensional contour matrix of the patient's spine can be obtained, such as Figure 3The deformity of the patient's spine can be clearly observed from the three-dimensional reconstruction results, including the obvious left convexity of the thoracic segment.

[0234] To further improve the accuracy of 3D reconstruction, the research team optimized each 3D contour matrix. Specifically, they first rotated the 3D reconstruction matrix according to the shooting angle of the patient's image to obtain 2D projections at different angles. Then, these 2D projections were compared with the original 2D contour images to calculate the degree of difference between them. By fitting this difference distribution through polynomial regression, the difference regression function for each shooting angle can be obtained. Finally, these difference functions are used to correct the 3D reconstruction matrix to further optimize its consistency with the actual spinal morphology.

[0235] After the above optimization steps, the research team obtained the target three-dimensional contour matrix of the patient's spine. This three-dimensional model can more accurately reflect the actual shape of the patient's spine, including the deformity characteristics of the left convexity of the thoracic segment.

[0236] In the fourth step, the optimized three-dimensional spinal contour model is input into the pre-trained orthopedic brace parameter generation model. This model also uses a deep neural network structure and can automatically generate a personalized brace design that meets the patient's needs based on the three-dimensional spinal morphology.

[0237] When training the parameter generation model, the research team collected a large amount of patients' 3D spine contour data and their corresponding orthopedic brace designs, parametrically processed the design schemes, extracted key parameters such as key point positions and pressure levels, and formed training samples. Through end-to-end training of deep learning, the model learned the complex mapping relationship from the 3D spine morphology to the brace parameters.

[0238] By inputting the patient's three-dimensional spine model into the trained parameter generation model, a set of personalized orthopedic brace design solutions with key parts such as support points, pressure points, and release points can be automatically obtained. For example, the main support point of the brace is designed at the convex part of the left thoracic segment to provide sufficient support to correct the deformity in this area; the pressure point is concentrated near the concave vertebral body, and a certain amount of pressure is applied in the hope of correcting the overall curvature of the spine; the release points are mainly distributed in the shoulder blades and hips, which are prone to pressure, to reduce stimulation to the skin and muscles; in addition, elastic points and buffer points are also set inside the brace to further increase the flexibility and comfort of the brace. At the same time, the parameter generation model also calculates the specific pressure of each key point, which provides an important basis for the subsequent brace manufacturing and debugging.

[0239] Through the above digital design process, the clinician and engineer team customized a set of personalized orthopedic braces for the patient that fully conforms to the actual morphology of the spine. Compared with the traditional manual design method, this digital method based on deep learning has many advantages:

[0240] 1. Accurate 3D spine reconstruction can be achieved by relying on conventional 2D medical images instead of expensive 3D scanning equipment. This greatly reduces the cost of medical treatment for patients.

[0241] 2. The design process is highly automated. From image acquisition to brace parameter generation, the entire process only takes a few hours, greatly shortening the production cycle and facilitating timely correction of spinal deformities.

[0242] 3. The brace can accurately fit the patient's spinal contour, reduce local pressure concentration, and improve wearing comfort. At the same time, the parameter generation model can also output pressure information at each key point, providing a basis for subsequent brace debugging.

[0243] 4. This method can quantitatively evaluate the correction effect and provide a basis for further optimization of the design by analyzing the stress conditions of the brace.

[0244] The following is a second specific application scenario of the present invention, Example 3, which focuses on the matrix and the calculation process: Patient Xiao Ming, 14 years old, was diagnosed with scoliosis during a physical examination. After a preliminary examination, it was suspected to be idiopathic adolescent scoliosis. The doctor decided to use the method of the present invention to design a customized scoliosis orthosis brace for the patient.

[0245] Step S10: Acquire medical images of the patient's spine from multiple shooting angles

[0246] The doctor took three X-rays of Xiao Ming at different angles, namely the front (AP), the side (LAT) and the 45-degree oblique (OBL). The resolution of each X-ray is 2048×2048 pixels, and the grayscale value range is 0-255. These three X-rays can be represented as:

[0247] I AP = {I AP (x,y)|x=1,2,...,2048; y=1,2,...,2048};

[0248] I LAT = {I LAT (x,y)|x=1,2,...,2048; y=1,2,...,2048};

[0249] I OBL = {I OBL(x,y)|x=1,2,...,2048; y=1,2,...,2048};

[0250] Among them I AP (x,y),I LAT (x,y) and I OBL (x,y) represents the grayscale value at the coordinate (x,y) in the frontal, lateral and oblique X-ray films respectively.

[0251] Step S20: Extract contours of medical images at each shooting angle

[0252] The Canny edge detection algorithm is used to extract the contours of the three X-ray films. First, each X-ray film is Gaussian filtered using a 5×5 Gaussian kernel with a standard deviation of σ=1.4. The Gaussian kernel function is:

[0253]

[0254] The gradient magnitude and direction are calculated for the filtered image. Taking the front X-ray as an example, the gradient magnitude matrix M is obtained. AP and the direction matrix θ AP :

[0255]

[0256]

[0257] Among them G x and G y are the gradients in the horizontal and vertical directions respectively.

[0258] Next, non-maximum suppression is performed, and the high and low thresholds are set to T h =100 and T l = 50. After double threshold detection and hysteresis processing, the contour image C of the frontal X-ray film is obtained. AP :

[0259] C AP ={C AP (x,y)|x=1,2,...,2048; y=1,2,...,2048; C AP (x,y)∈{0,1}}

[0260] Similarly, the contour images C of the lateral and oblique X-ray films can be obtained. LAT and C OBL .

[0261] Step S30: Size registration and compression coding of each contour image

[0262] Taking the frontal X-ray as an example, first determine the four key anatomical landmarks:

[0263] 1. The midpoint of the lower edge of the seventh cervical vertebra: (1024,256);

[0264] 2. The midpoint of the lower edge of the 12th thoracic vertebra: (1024,1024);

[0265] 3. The midpoint of the lower edge of the fifth lumbar vertebra: (1024,1792);

[0266] 4. Midpoint of the upper edge of the pelvis: (1024,2048)

[0267] Calculate the scaling and rotation angles:

[0268]

[0269] For the contour image C AP Perform geometric transformations:

[0270]

[0271] Where (x 0 ,y 0 ) is the translation amount to align the image center.

[0272] Then use bilinear interpolation to resample the image C′ AP , so that its resolution is unified to 1024×1024.

[0273] Run-length coding is performed on the registered contour image to obtain the contour two-dimensional matrix M AP :

[0274] M AP ={(l j ,n j )|j=1,2,...,J};

[0275] For example, M AP The first few elements of may be:

[0276] {(100,0),(50,1),(200,0),(30,1),...};

[0277] Similarly, the lateral and oblique X-rays are processed to obtain M LAT and M OBL .

[0278] Step S40: Input the contour matrix into the spine 3D contour matrix generation model

[0279] Use the pre-trained 3D convolutional neural network model to transform the contour matrix M AP 、M LAT and M OBLAs input, we get the corresponding three-dimensional generator matrix T AP 、T LAT and T OBL .

[0280] The structure of the neural network is as follows:

[0281] 1. Input layer: 1024×1024×3 (contour matrix of three viewing angles);

[0282] 2.3D convolution layer 1: 64 3×3×3 convolution kernels, stride 1;

[0283] 3.3D max pooling layer 1: 2×2×2 pooling window, stride 2;

[0284] 4.3D convolution layer 2: 128 3×3×3 convolution kernels, stride 1;

[0285] 5.3D max pooling layer 2: 2×2×2 pooling window, stride 2;

[0286] 6.3D convolution layer 3: 256 3×3×3 convolution kernels, stride 1;

[0287] 7.3D deconvolution layer 1: 128 3×3×3 convolution kernels, stride 2;

[0288] 8.3D deconvolution layer 2: 64 3×3×3 convolution kernels, stride 2;

[0289] 9.3D convolution layer 4: 1 1×1×1 convolution kernel, stride 1, activation function is sigmoid;

[0290] 10. Output layer: 256×256×256 three-dimensional matrix;

[0291] The forward propagation function of the model can be expressed as:

[0292]

[0293] Where Θ is the model parameter, is the predicted three-dimensional contour matrix.

[0294] Assume that the three three-dimensional generating matrices are:

[0295] T AP ={T AP (x,y,z)|x,y,z=1,2,...,256;T AP (x,y,z)∈[0,1]};

[0296] T LAT ={T LAT(x,y,z)|x,y,z=1,2,...,256; T LAT (x,y,z)∈[0,1]};

[0297] T OBL ={T OBL (x,y,z)|x,y,z=1,2,...,256; T OBL (x,y,z)∈[0,1]};

[0298] Step S50: Calculate the similarity of the three-dimensional generator matrix

[0299] T AP Generate a matrix based on the three shooting angles and rotate it to get three three-dimensional theoretical matrices and

[0300] The rotation matrices are:

[0301]

[0302] To T AP Apply the rotation matrix to each point (x,y,z) in , resulting in three theoretical matrices.

[0303] Calculate the similarity between the three-dimensional theoretical matrix of each shooting angle and the three-dimensional generated matrix of the corresponding angle:

[0304] Euclidean distance:

[0305]

[0306] Cosine Similarity:

[0307]

[0308]

[0309] SSIM:

[0310]

[0311] Comprehensive similarity (weight: α = 0.3, β = 0.4, γ = 0.3):

[0312] S AP =0.3×0.15+0.4×0.98+0.3×0.96=0.737;

[0313] S LAT =0.3×0.25+0.4×0.92+0.3×0.90=0.713;

[0314] S OBL=0.3×0.20+0.4×0.95+0.3×0.93=0.725;

[0315] Similarly, calculate T LAT and T OBL Similarity array used as the basis for generating the matrix.

[0316] Step S60: Select candidate generation matrix and perform two-dimensional extraction

[0317] According to the similarity array, select T AP As a candidate generation matrix (similarity and maximum). AP Perform two-dimensional extraction to obtain the two-dimensional theoretical matrix of three shooting angles and

[0318] by For example, the two-dimensional extraction process is as follows:

[0319] 1. Construct the projection matrix

[0320]

[0321] 2. T AP To each point (x,y,z) in , apply the projection matrix:

[0322]

[0323] 3. Get the two-dimensional theoretical matrix

[0324] Step S70: Calculate the difference regression function and modify the candidate generation matrix

[0325] For each shooting angle, calculate the pixel-level difference between the 2D theoretical matrix and the corresponding contour 2D matrix. Take the front angle as an example:

[0326]

[0327] Difference d AP (x, y) to obtain its distribution characteristics. Assuming that the difference follows a normal distribution, calculate the mean μ AP and standard deviation σ AP :

[0328]

[0329] Use polynomial regression to fit the difference distribution and get the difference regression function f AP (d):

[0330] f AP (d) = 0.03 + 0.8d - 1.5d2 +0.7d 3 ;

[0331] Similarly, the regression function f of the difference between the lateral and oblique positions can be obtained: LAT (d) and f OBL (d).

[0332] Use these difference regression functions to generate the candidate matrix T AP Make corrections and get the target generation matrix T * :

[0333] T * (x,y,z)=T AP (x,y,z)+w AP f AP (d AP )+w LAT f LAT (d LAT )+

[0334] w OBL f OBL (d OBL );

[0335] where w AP 、w LAT and w OBL is the weight of each angle, which can be set according to the similarity, for example:

[0336] w AP =0.4,w LAT =0.3,w OBL =0.3;

[0337] Step S80: Perform run-length coding inverse operation on the target generator matrix

[0338] Generate the target matrix T * Perform the run-length encoding inverse operation to obtain a complete three-dimensional contour image of the patient's spine. Set the threshold t = 0.5 and set T * Elements in that are larger than the threshold are considered as contour points.

[0339] Traverse T * The position (x, y, z) of each non-zero element in , records the consecutive non-zero elements, and forms a run-length encoding sequence. For example, the following sequence may be obtained:

[0340] {(10,0),(5,1),(15,0),(8,1),...};

[0341] These run-length encoding sequences are inversely operated and expanded into a complete three-dimensional contour image V:

[0342] V={V(x,y,z)|x,y,z=1,2,...,256; V(x,y,z)∈{0,1}};

[0343] Step S90: Inputting the scoliosis orthopedic brace parameter model

[0344] The 3D contour image V of the patient's spine is input into the pre-trained scoliosis brace parameter model. The model adopts a deep neural network structure, including 3D convolutional layers, pooling layers and fully connected layers.

[0345] The model structure is as follows:

[0346] 1. Input layer: 256×256×256×1;

[0347] 2.3D convolution layer 1: 32 3×3×3 convolution kernels, stride 1, activation function ReLU;

[0348] 3.3D max pooling layer 1: 2×2×2 pooling window, stride 2;

[0349] 4.3D convolution layer 2: 64 3×3×3 convolution kernels, stride 1, activation function ReLU;

[0350] 5.3D max pooling layer 2: 2×2×2 pooling window, stride 2;

[0351] 6.3D convolution layer 3: 128 3×3×3 convolution kernels, stride 1, activation function ReLU;

[0352] 7.3D max pooling layer 3: 2×2×2 pooling window, stride 2;

[0353] 8. Fully connected layer 1: 1024 neurons, activation function ReLU;

[0354] 9. Fully connected layer 2: 512 neurons, activation function ReLU;

[0355] 10. Output layer: 6 categories (support point, pressure point, release point, elastic point, buffer point, traction point), 3D coordinates of 256 points in each category

[0356] The forward propagation function of the model can be expressed as:

[0357] O = g(V; Φ);

[0358] Where Φ is the model parameter and O is the predicted brace parameter matrix.

[0359] Assume that the brace parameter matrix O output by the model has the following structure:

[0360] O={O k|k=1,2,...,6};

[0361] Among them, k Represents the coordinate matrix of the k-th point:

[0362] O k ={(x i ,y i ,z i )|i=1,2,...,256};

[0363] Specifically, the meaning and quantity of each type of points are as follows:

[0364] 1. Support point (O 1 ): 256 points, mainly distributed in the back and waist areas;

[0365] 2. Pressure point (O 2 ): 256 points, concentrated on the ribs and waist area on the convex side;

[0366] 3. Release point (O 3 ): 256 points, mainly distributed in the concave side and breathing area;

[0367] 4. Elastic point (O 4 ): 256 points, distributed in the edge area of ​​the brace;

[0368] 5. Buffer point (O 5 ): 256 points, distributed on bony prominences and sensitive areas;

[0369] 6. Pull point (O 6 ): 256 points, distributed in the shoulder and hip areas;

[0370] Based on these parameters, a 3D model of the inner surface of the scoliosis brace can be generated. For example, some coordinates of the support points may be as follows:

[0371] O 1 ={(120,150,50),(121,150,51),(122,151,51),...};

[0372] Some coordinates of a pressure point might look like this:

[0373] O 2 ={(130,140,60),(131,141,61),(132,141,62),...};

[0374] The coordinates of these points will be used to construct the geometry and functional areas of the brace.

[0375] Based on the output brace parameter matrix O, a CAD model of the brace can be generated using 3D modeling software (such as SolidWorks or Fusion360). The specific steps are as follows:

[0376] 1. Import the patient's spine 3D contour image V as a reference.

[0377] 2. According to the support point O 1 Create the basic outline of the brace.

[0378] 3. At pressure point O 2 Add raised or thickened areas to locations.

[0379] 4. At release point O 3 Create a depression or opening in the desired location.

[0380] 5. At elastic point O 4 Design elastic structure or use elastic material at the location.

[0381] 6. At buffer point O 5 Add cushioning or pressure relief structures to the location.

[0382] 7. At traction point O 6 Position design of traction device or fixed structure.

[0383] The resulting 3D model of the inner surface of the brace will contain the following features:

[0384] 1. The overall shape matches the patient's spinal contour

[0385] 2. The convex side area has appropriate pressure point distribution

[0386] 3. There is enough space on the concave side and breathing area

[0387] 4. The edge area has a certain degree of elasticity

[0388] 5. Bone protrusions and sensitive areas are cushioned and protected

[0389] 6. The shoulders and hips have traction and fixation structures

[0390] In order to better demonstrate the design effect of the brace, we can generate some data to design the key parameters of the brace, as shown in the following tables:

[0391] Table 1 Support point density distribution:

[0392] area Upper back Middle back Lower back waist density(%) 20 30 25 25

[0393] Table 2 Pressure point distribution: Position: Right ribs Left ribs Right waist Left waist

[0394]

[0395] Table 3 Release point area ratio:

[0396] area Thoracic cavity abdomen Left Right Proportion(%) 15 10 8 5

[0397] Table 4 Elastic point distribution:

[0398] Location Neck edge Armpits Hip Thigh quantity 60 40 80 76

[0399] Table 5 Buffer point thickness distribution:

[0400] Location shoulder blade Spinal process Iliac crest Sacrum Thickness(mm) 5 4 6 8

[0401] Table 6 Force distribution at traction point:

[0402] Location Left shoulder Right shoulder Left hip Right hip Force (N) 20 25 30 35

[0403] Based on these data, corresponding charts can be generated to intuitively display the distribution characteristics of each functional point of the brace. These data and charts will help doctors and patients better understand the design principles and expected effects of the brace.

[0404] After the brace design is completed, virtual fitting and simulation analysis are required to evaluate the correction effect and comfort of the brace. Finite element analysis software (such as ANSYS) can be used to perform the following simulations:

[0405] 1. Structural strength analysis: Ensure that the brace will not be deformed or damaged during use.

[0406] 2. Pressure distribution analysis: Check whether the pressure of the brace on various parts of the body is reasonable.

[0407] 3. Thermal comfort analysis: Evaluate the ventilation and heat dissipation performance of the brace.

[0408] 4. Kinematic analysis: Simulate the performance of the brace during the patient's daily activities.

[0409] The results of the simulation analysis may be as follows:

[0410] Structural strength analysis:

[0411] Maximum stress: 2.5MPa (located in the right rib pressure point area);

[0412] Maximum deformation: 0.8mm (located in the lumbar support area);

[0413] Safety factor: 3.2;

[0414] Pressure distribution analysis:

[0415] Maximum pressure: 30kPa (located at the right rib bulge);

[0416] Average pressure: 12kPa;

[0417] Pressure uniformity: 85%;

[0418] Thermal comfort analysis:

[0419] Maximum temperature: 35°C (located in the center of the back);

[0420] Average temperature: 32℃

[0421] Heat exchange efficiency: 75%;

[0422] Kinematic analysis:

[0423] Improvement of scoliosis Cobb angle: initial 40°, 30° after correction;

[0424] Range of motion restrictions: flexion and extension decreased by 15%, lateral bending decreased by 25%, and rotation decreased by 20%;

[0425] Based on the results of these analyses, minor adjustments to the brace design may be necessary. For example:

[0426] 1. Increase material thickness or use stronger material in the area of ​​maximum stress.

[0427] 2. Adjust the pressure point distribution to make the pressure more uniform.

[0428] 3. Add ventilation holes in high temperature areas to improve heat dissipation.

[0429] 4. Optimize the design of the active joints of the brace to improve the freedom of movement while ensuring the correction effect.

[0430] Finally, based on the optimized design, a 3D printing technology was used to make a brace prototype. Selective laser sintering (SLS) technology was used to print using PA12 nylon powder material. The printing parameters are as follows:

[0431] Layer thickness: 0.1mm;

[0432] Printing temperature: 180℃;

[0433] Printing speed: 50mm / s;

[0434] Fill rate: 80%;

[0435] After printing, post-processing is required, including surface grinding, cleaning and disinfection. Then, physical test assembly is carried out to evaluate the following indicators:

[0436] 1. Fit: How well the brace fits the patient's body;

[0437] 2. Pressure distribution: The actual pressure distribution is measured by a pressure sensor;

[0438] 3. Comfort: subjective rating by the patient (1-10 points);

[0439] 4. Degree of activity limitation: Measure the patient's range of motion when wearing the brace;

[0440] 5. Correction effect: X-ray review to measure the improvement of Cobb angle;

[0441] The test results may be as follows:

[0442] Fit: 95% (only slight gap in the underarm area)

[0443] Pressure distribution: consistent with the simulation analysis results, the maximum pressure is 32kPa;

[0444] Comfort: Patient score: 8 (slight discomfort, but acceptable);

[0445] Activity restriction degree: basically consistent with the simulation analysis results;

[0446] Correction effect: The Cobb angle was improved from the initial 40° to 28°, which was better than the simulation prediction;

[0447] Based on the trial fitting results, make final fine-tuning of the brace:

[0448] 1. Adjust the contours of the underarm area for a better fit.

[0449] 2. Slightly reduce pressure on the right rib area.

[0450] 3. Increase the elasticity of lumbar support and improve comfort.

[0451] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A digital design method for a scoliosis orthopedic brace, characterized in that: The following steps are involved: S10, obtaining medical images of the patient's spine from at least two shooting angles; S20, extracting the contour of the medical image at each shooting angle to obtain a contour image corresponding to the shooting angle; S30, for each contour image, after size registration, set the areas without contours in the contour image to 0, set the areas with contours to 1, and compress them using run-length encoding to obtain a contour two-dimensional matrix; S40, inputting each contour matrix into a pre-trained spine three-dimensional contour matrix generation model to obtain a corresponding contour three-dimensional matrix, which is recorded as a three-dimensional generation matrix; S50, traversing each three-dimensional generation matrix as a basic generation matrix, rotating it according to the shooting angles of all medical images, and obtaining a contour three-dimensional matrix of each shooting angle, recorded as a three-dimensional theoretical matrix; Calculate the similarity between the three-dimensional theoretical matrix of each shooting angle and the three-dimensional generated matrix of the corresponding angle; After the traversal is completed, the similarity array of each three-dimensional generator matrix is ​​obtained; S60, selecting elements of the similarity array and the largest three-dimensional generator matrix as candidate generator matrices, performing two-dimensional extraction according to the shooting angles of all medical images, and obtaining a two-dimensional theoretical matrix for each shooting angle; S70, calculating a difference regression function of each shooting angle based on the two-dimensional theoretical matrix of each shooting angle and the corresponding two-dimensional contour matrix; And using the difference regression function of each shooting angle, the candidate generation matrix is ​​corrected to obtain the target generation matrix; S80, expanding the target generation matrix by means of run-length coding inverse operation to obtain a three-dimensional contour image of the patient's spine; S90, inputting the three-dimensional contour image of the patient's spine into a pre-trained scoliosis orthosis brace parameter model, obtaining a three-dimensional model of the inner surface of the scoliosis orthosis brace marked with support points, pressure points, release points, elastic points, buffer points, and traction points as a design solution and outputting the model; The candidate generation matrix is ​​subjected to two-dimensional extraction according to the shooting angle, and the specific steps include: Determine the projection plane according to the shooting angle; Perform projection transformation on each point in the three-dimensional matrix; The projected points are gathered into a two-dimensional matrix; Interpolate and smooth two-dimensional matrices; Calculating the difference regression function of the shooting angles, the specific steps include: Calculate the pixel-level difference between the 2D theoretical matrix and the contour 2D matrix; The differences were statistically analyzed to obtain the difference distribution; Use polynomial regression or support vector regression method to fit the difference distribution and obtain the regression function.

2. A method for digital design of a scoliosis orthopedic brace according to claim 1, characterized in that: The medical image is an X-ray, a CT scan or an MRI scan.

3. A digital design method for scoliosis orthopedic brace according to claim 2, characterized in that: The method for extracting contours of medical images at each shooting angle adopts the Canny edge detection algorithm or the active contour model algorithm.

4. A method for digital design of a scoliosis orthopedic brace according to claim 3, characterized in that: The specific steps of calculating the similarity between the three-dimensional theoretical matrix of each shooting angle and the three-dimensional generated matrix of the corresponding angle are: Calculate the Euclidean distance between two matrices; Calculate the cosine similarity of two matrices; Calculate the structural similarity index of two matrices; Combining the above three indicators, the weighted average is taken as the final similarity.

5. The method for digital design of a scoliosis orthopedic brace according to claim 1, characterized in that: For each contour image, the steps of size registration include: Determine the key anatomical landmarks in the Zhanglun image, including the vertebral margin and intervertebral disc; Calculate the zoom ratio and rotation angle of the wheel image according to the marker points; According to the calculated image scaling ratio and rotation angle, the Zhanglun image is geometrically transformed accordingly to align it with the standard template; The bilinear interpolation method is used to resample the images so that all the images have the same pixel resolution.

6. The method for digital design of a scoliosis orthopedic brace according to claim 1, characterized in that: The structure of the spine three-dimensional contour matrix generation model adopts a 3D convolutional neural network, and the steps of training data acquisition specifically include: Collect a large amount of spinal CT or MRI scan data; Professional doctors will annotate the spinal contours on the scan data; Perform data enhancement on the labeled data, such as rotation, scaling, and adding noise; The processed data is divided into training set, validation set and test set.

7. The method for digital design of a scoliosis orthopedic brace according to claim 1, characterized in that: The basic generation matrix is ​​rotated by using a three-dimensional rotation matrix transformation method. The specific steps include: Calculate the rotation angle based on the shooting angle; Construct the corresponding rotation matrix; Transforms each point in the base generator matrix by applying a rotation matrix.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, which, when executed, are used to execute the digital design method for a scoliosis orthopedic brace according to any one of claims 1 to 7.

9. A digital design system for scoliosis orthopedic braces, characterized in that: Contains the computer-readable storage medium of claim 8.

Images