A method and apparatus for screening for congenital scoliosis risk

Abstract

This invention discloses a method for screening the risk of congenital scoliosis. The method includes: acquiring a back surface image of the subject, which reflects the surface morphological distribution of the back region; preprocessing the back surface image to obtain a standardized surface image to reduce the influence of differences in acquisition posture, imaging conditions, or subject body shape; based on the standardized surface image, constructing surface morphological feature information for congenital scoliosis risk screening, which includes at least the left-right symmetry of the back, the state of spinal midline deviation, or the trend of surface morphological changes; inputting the surface morphological feature information into a pre-trained congenital scoliosis risk screening model to generate congenital scoliosis risk screening results. This invention can achieve objective and consistent congenital scoliosis risk screening.

Description

Technical Field

[0001] This invention relates to a biometric identification software method and apparatus based on machine learning or neural networks, and more particularly to a method for analyzing and judging surface image features for screening the risk of congenital scoliosis and its application software implementation. Background Technology

[0002] Congenital scoliosis is one of the important types of spinal deformities in childhood, and it is often included in the management framework of early-onset scoliosis (EOS) for follow-up and intervention in clinical practice. Figure 1 This is a schematic diagram of early-onset scoliosis (EOS) based on related technologies. For example... Figure 1 As shown, early-onset scoliosis (EOS) is managed by classification based on age of onset, etiology, and characteristics of deformity progression. Different types show significant differences in progression rate and intervention window. Common characteristics of EOS include young age of onset, potentially rapid deformity progression, difficulty in treatment, and impact on the normal development of the thoracic cavity and cardiopulmonary system. Without effective control, severe cases can pose high health risks; therefore, "early intervention" is crucial for controlling scoliosis progression and ensuring cardiopulmonary development. Furthermore, scoliosis is essentially a three-dimensional deformity encompassing coronal curvature, axial rotation, and sagittal curvature changes. As the deformity worsens, vertebral and thoracic structures can change, and the anatomical relationships between thoracic / abdominal organs and structures within the spinal canal may gradually become abnormal, leading to risks of organ dysfunction, particularly cardiopulmonary dysfunction.

[0003] Current scoliosis screening and assessment pathways typically involve visual inspection / flexion tests, followed by imaging confirmation when necessary. Taking adolescent idiopathic scoliosis (AIS) screening as an example, guidelines recommend combining visual inspection, the Adam flexion test, and trunk rotation angle (ATR) measurement to improve screening accuracy, and suggest using ATR > 5° as a positive screening indicator. The guidelines also indicate that posture assessment can employ non-radioactive methods such as physical examination, the Adam flexion test, ATR measurement, and surface morphology. Surface morphology is characterized by being non-invasive, radiation-free, and providing objective assessment, exhibiting high reliability and validity in identifying postural abnormalities caused by a certain degree of severity (e.g., Cobb angle ≥ 20°). However, from an engineering and scaling perspective, traditional visual inspection / flexion test relies on the examiner's experience and is highly subjective; ATR measurement and palpation have limitations in terms of standardization, repeatability, and scenario adaptability; although X-ray examination can accurately calculate the Cobb angle, it has limitations such as radiation exposure, high cost, and high equipment requirements, making it difficult to use as a routine initial screening method.

[0004] Furthermore, with the development of artificial intelligence and application software development technologies, biometric recognition software based on machine learning or neural networks is gradually being applied to medical image and body surface morphology analysis scenarios. By learning and recognizing morphological features in human body surface images, it achieves automated identification and risk alerts for specific abnormal body posture patterns. This type of biometric recognition software typically uses body surface images or videos as input, combining feature extraction, pattern recognition, and classification decision-making processes to output recognition results. It features high automation, batch processing capability, and ease of embedding into mobile terminals or cloud systems, thus possessing high promotional value in application software development fields such as primary care screening, telemedicine, and health management. However, body surface image data for children with congenital scoliosis exhibits characteristics such as inconsistent acquisition postures, significant differences in imaging conditions, and substantial individual body shape variations. Without targeted data standardization and model training strategies, the aforementioned biometric recognition software is prone to recognition instability or insufficient generalization ability when applied across institutions, devices, and populations, thereby affecting the consistency and reliability of screening results.

[0005] Therefore, the technical problem to be solved by this patent is how to conduct a consistent, objective, and repeatable screening of "scoliosis risk" using easily accessible surface information without relying on high-dose or frequent imaging examinations in the scenario of congenital scoliosis risk screening, and to minimize the accumulation of errors caused by differences in posture, imaging conditions, and individual body types. This would enable the screening to have transferable and reusable engineering capabilities across different institutions, different equipment, and different examinee populations, and provide a more stable basis for subsequent referrals and imaging confirmation. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention aims to provide a method and apparatus for screening the risk of congenital scoliosis.

[0007] According to one aspect of the present invention, a method for screening the risk of congenital scoliosis is provided. The method includes: acquiring a back surface image of a subject, the back surface image reflecting the surface morphological distribution of the back region; preprocessing the back surface image to obtain a standardized surface image to reduce the influence of differences in acquisition posture, imaging conditions, or subject body shape; based on the standardized surface image, constructing surface morphological feature information for congenital scoliosis risk screening, the surface morphological feature information including at least the left-right symmetry of the back, the state of spinal midline deviation, or the trend of surface morphological changes; inputting the surface morphological feature information into a pre-trained congenital scoliosis risk screening model to generate congenital scoliosis risk screening results.

[0008] According to another aspect of the present invention, a congenital scoliosis risk screening device is also provided. The device includes: an acquisition module for acquiring a back surface image of a subject, the back surface image reflecting the surface morphological distribution of the back region; a preprocessing module for preprocessing the back surface image to obtain a standardized surface image, thereby reducing the influence of differences in acquisition posture, imaging conditions, or subject body shape; a construction module for constructing surface morphological feature information for congenital scoliosis risk screening based on the standardized surface image, the surface morphological feature information including at least the left-right symmetry of the back, the state of spinal midline deviation, or the trend of surface morphological changes; and a generation module for inputting the surface morphological feature information into a pre-trained congenital scoliosis risk screening model to generate congenital scoliosis risk screening results.

[0009] This invention acquires images of the back and preprocesses them to reduce the influence of posture, imaging conditions, and body shape differences. Under a unified standard, it constructs morphological feature information reflecting the left-right symmetry of the back, the midline deviation of the spine, and the trend of changes in the body surface morphology. The features are then input into a pre-trained risk screening model for analysis, thereby achieving objective and consistent risk screening for congenital scoliosis without the need for imaging examinations. This solves the problems of strong dependence on screening methods, insufficient stability, and lack of scalability in existing technologies. Attached Figure Description

[0010] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0011] Figure 1 This is a schematic diagram of early-onset scoliosis (EOS) based on related technologies;

[0012] Figure 2 This is a flowchart of a method for screening the risk of congenital scoliosis according to an embodiment of the present invention;

[0013] Figure 3 This is a schematic representation of the back of a person with congenital scoliosis according to Example 1 of the present invention; and

[0014] Figure 4 This is a schematic diagram of a congenital scoliosis risk screening system according to an embodiment of the present invention. Detailed Implementation

[0015] The following embodiments are merely examples to clearly illustrate the present invention and are not intended to limit the implementation of the invention. Those skilled in the art can make other variations or modifications based on the following description, and these variations, modifications, substitutions, and alterations arising from the principles and spirit of the present invention still fall within the protection scope of the present invention.

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0017] According to an embodiment of the present invention, a method for screening the risk of congenital scoliosis is provided. Figure 1 This is a flowchart of a method for screening the risk of congenital scoliosis according to an embodiment of the present invention. Figure 2 As shown, the steps include S202 to S208.

[0018] Step S202: Acquire a back surface image of the subject, which is used to reflect the surface morphology distribution of the back region.

[0019] Step S204: Preprocess the back surface image to obtain a standardized surface image to reduce the impact of differences in acquisition posture, imaging conditions or object size.

[0020] Step S206: Based on the standardized body surface image, construct body surface morphological feature information for risk screening of congenital scoliosis. The body surface morphological feature information includes at least the left and right symmetry of the back, the state of midline deviation of the spine, or the trend of changes in body surface morphology.

[0021] Step S208: Input the body surface morphological feature information into the pre-trained congenital scoliosis risk screening model to generate congenital scoliosis risk screening results.

[0022] Current screening methods for congenital scoliosis mainly rely on physical examinations, flexion tests, and imaging techniques. These methods suffer from high subjectivity, low standardization, and the need for repeated radiation examinations, making it difficult to achieve continuous, repeatable, and large-scale risk screening in early childhood. Furthermore, differences in body shape, posture, and imaging conditions among children can lead to unstable screening results based on surface information, limiting their application in primary care and follow-up settings. This invention collects back surface images and preprocesses them to reduce the influence of posture, imaging conditions, and body shape differences. Under a unified standard, it constructs morphological feature information reflecting the left-right symmetry of the back, spinal midline deviation, and trends in surface morphology. These features are then input into a pre-trained risk screening model for analysis. This achieves objective and consistent congenital scoliosis risk screening without the need for imaging examinations, solving the problems of high dependence on screening methods, insufficient stability, and limited scalability in existing technologies.

[0023] According to one embodiment of the present invention, preprocessing the back surface image includes: detecting the pixel positions of at least two anatomical reference regions in the back surface image, the anatomical reference regions including a shoulder region and a pelvic region; constructing a human posture reference coordinate system based on the pixel positions of the anatomical reference regions; and performing a two-dimensional or three-dimensional affine transformation on the back surface image according to the human posture reference coordinate system, so that the back surface image undergoes rotation, translation, and scale normalization processing under the human posture reference coordinate system; wherein the affine transformation parameters are determined by minimizing the sum of squared Euclidean distances between the reference regions before and after the transformation.

[0024] In this embodiment of the invention, by constructing a pose coordinate system based on an anatomical reference region and performing an affine transformation, the back surface images under different acquisition poses are aligned under a unified geometric reference, reducing the impact of pose differences on subsequent feature construction and model input consistency.

[0025] According to one embodiment of the present invention, after performing a two-dimensional or three-dimensional affine transformation on the back surface image, the method further includes:

[0026] The image of the back surface is represented as a two-dimensional pixel matrix. And perform brightness normalization processing on the pixel matrix to make it meet the requirements.

[0027] , (x,y)∈Ω

[0028] in,

[0029] Ω is the set of pixels in the dorsal surface image that are involved in the statistics and calculations.

[0030] It is the pixel intensity before brightness normalization.

[0031] It is the pixel intensity after brightness normalization.

[0032] and Let represent the mean and standard deviation of the pixel intensity of the back surface image, respectively, and be categorized by... Statistical analysis of internal pixels;

[0033] After completing brightness normalization, this pixel matrix... Apply a spatial smoothing filter operator to reduce high-frequency noise.

[0034] In this embodiment of the invention, pixel-level statistical normalization and smoothing are used to reduce the impact of different imaging conditions and device parameters on the grayscale distribution of the image, providing a unified data foundation for the stable extraction of subsequent morphological features.

[0035] According to one embodiment of the present invention, preprocessing the back surface image further includes:

[0036] The back surface image after brightness normalization and smoothing filtering is represented as a scalar pixel function I'(x,y) defined on two-dimensional spatial coordinates (x,y);

[0037] Based on the first and second partial derivatives of the scalar pixel function in the two-dimensional spatial coordinates, differential features of the body surface morphology are constructed, wherein...

[0038] Taking the first partial derivative of the pixel function I'(x,y) yields the gradient vector field.

[0039] ∇I'(x,y)=

[0040] Take the second-order partial derivative of the pixel function I'(x,y) and construct the Hessian matrix.

[0041]

[0042] in exist The second order is continuously differentiable, such that ,

[0043] The pixel function I'(x,y), the gradient vector field ∇I'(x,y), and the Hessian matrix H(I'(x,y)) are concatenated along the feature dimension to form a unified body surface morphology differential feature tensor.

[0044] T(x,y)=[I′(x,y), ∇I′(x,y), H(I′(x,y))]

[0045]

[0046] The differential feature tensor of the body surface morphology is used as a component of the standardized body surface image for the subsequent construction of body surface morphology feature information.

[0047] This invention characterizes the variation trend and local structural characteristics of the back surface morphology in different spatial directions at a continuous mathematical level. Specifically, the gradient vector field formed by first-order partial differentials reflects the direction and magnitude of change in the back surface morphology along the horizontal and vertical directions, while the Hessian matrix constructed by second-order partial differentials further describes the curvature distribution and intensity of change of the surface morphology in local regions. Based on this, a unified surface morphology differential feature tensor is formed, which helps to reduce the information deficiency caused by simply relying on the intensity of the original image, allowing the subtle morphological changes related to scoliosis on the back surface to be completely preserved in a structured form.

[0048] According to one embodiment of the present invention, after inputting the body surface morphological feature information into a pre-trained congenital scoliosis risk screening model, the method further includes:

[0049] The congenital scoliosis risk screening model represents the body surface morphology differential feature tensor as defined in the back body surface domain. tensor field on ;

[0050] This congenital scoliosis risk screening model includes a tensor coding backbone network and a global discriminant head network.

[0051] Specifically, the tensor encoding backbone performs multi-scale tensor convolution and tensor shrinking on the tensor field to obtain the latent variable tensor representation. The tensor is shrunk and satisfies

[0052]

[0053] in, It is the number of input tensor channels. It is the output channel index. It is the input channel index;

[0054] It is related to the output channel Input Channel The corresponding spatial weight kernel / attention weight;

[0055] It is the first after the merger One latent variable component.

[0057] A morphological consistency mapping based on a Riemannian metric is introduced into the latent variable space. This Riemannian metric is derived from a symmetric positive definite matrix field. Given that for any two points The geodetic distance satisfies

[0058]

[0059] in, It is the body surface area The above two points;

[0060] ;

[0061] A symmetric positive definite matrix field gives a local metric;

[0062] It is the path derivative;

[0063] It is transpose.

[0064] Furthermore, the global discriminant head network outputs a risk score vector based on the latent variable tensor representation.

[0065] This invention, through explicit modeling of the input as a tensor field on the body surface domain and the introduction of tensor shrinkage and geodesic distance constraints, enables the model to handle spatial variations and cross-individual differences in body surface morphology under a unified geometric metric. This reduces the inconsistency in representation caused by relying solely on Euclidean pixel space and enhances the modeling ability for continuous changes in back morphology.

[0066] According to one embodiment of the present invention, the congenital scoliosis risk screening model is trained by minimizing a variational composite objective function, which is dependent on the sample distribution. The expectation is expressed as

[0067]

[0068] in,

[0069] This is a risk screening model for congenital scoliosis.

[0070] It is the set of trainable parameters for the model.

[0071] It represents information about the body surface morphology.

[0072] This indicates a risk label corresponding to the body's surface morphological characteristics.

[0073] Among the regularization terms To define in the domain A weighted combination of the Sobolev seminorm and mirror consistency constraints satisfies

[0074]

[0075] in,

[0076] and Let these represent the first-order gradient operator and the Hessian operator with respect to spatial coordinates, respectively.

[0077] It is an input that corresponds to the morphological features of the body surface.

[0078] It is a surveillance label used to represent risk level / positive / negative results.

[0079] It is the training distribution.

[0080] (T) is the model's output scoring field on the body surface.

[0081] This represents the mirror coordinate mapping point about the midline of the spine.

[0082] It is the Frobenius norm.

[0083] These are hyperparameters used to balance various constraints.

[0084] It is a mirror coordinate mapping about the midline of the spine.

[0085] The embodiments of the present invention suppress non-smooth oscillations in the model output in space by writing the training target as a variational functional that integrates over the domain and introducing first-order and second-order spatial regularization terms. This makes the model output more consistent with the prior of continuous changes in back morphology. At the same time, the mirror consistency term provides a stable constraint on the difference between left and right morphology, thereby improving the generalization stability under cross-pose and cross-imaging conditions.

[0086] According to one embodiment of the present invention, the congenital scoliosis risk screening model is trained using a two-layer optimization framework, wherein,

[0087] Inner layer optimization is used to update model parameters. The inner layer optimization satisfies:

[0088] ;

[0089] Outer layer optimization is used to update the hyperparameter set. The outer layer optimization satisfies:

[0090] ;

[0091] In the inner layer iterative update process, the parameter update rule obtained by discretizing the stochastic differential equation is adopted, satisfying... ,

[0092] in,

[0093] In Dimensions and same

[0094] It is a set of hyperparameters.

[0095] It is the step size parameter.

[0096] It's a temperature parameter.

[0097] It is a Gaussian random perturbation term.

[0098] Is with parameters An identity matrix with consistent dimensions.

[0099] The embodiments of the present invention achieve adaptive selection of hyperparameters through two-layer optimization, enabling the training process to achieve a more stable trade-off between classification error and morphological constraints. At the same time, the update rule in the form of stochastic differential equations improves the search capability under complex non-convex targets and reduces the probability of getting trapped in bad local extrema, thereby improving the training robustness and transferability of the pre-trained model under different data distribution conditions.

[0100] According to one embodiment of the present invention, after completing the forward inference of the pre-trained congenital scoliosis risk screening model, a model defined on the back surface domain is constructed based on the model's response to input body surface morphological features. On the continuous risk scoring field ;

[0101] The risk scoring field is derived from the intermediate latent variables output by the model via a mapping function. The transformation yields the result and satisfies the following conditions:

[0102]

[0103] in This represents the latent variable representation of the model at the corresponding location on the dorsal body surface.

[0104] The risk scoring field is located in the dorsal body surface area. Perform spatial integration and normalization to obtain the global risk scalar. Its satisfaction

[0105]

[0106] in This indicates the measurement of the dorsal body surface area;

[0107] Based on this global risk scalar And combined with a pre-set set of risk assessment rules By comparing the global risk scalar with the threshold in the set of judgment rules, the risk screening result of congenital scoliosis corresponding to the examinee is determined.

[0108] This invention extends the model output from discrete decision values ​​to a continuous risk score field defined on the body surface, and further converges it into a global risk scalar through integration and normalization. This gives the screening result generation process a clear spatial mathematical basis, which helps to obtain stable and consistent risk assessment results under complex body surface morphology distribution conditions.

[0109] According to one embodiment of the present invention, the step of generating the congenital scoliosis risk screening result further includes:

[0110] This risk scoring field Considered as defined in the dorsal body surface region A random field is used to construct a risk probability measure based on this random field. ;

[0111] Among them, for any measurable sub-region This risk probability measure satisfies

[0112]

[0113] Based on this risk probability measure, the expected value of risk is defined as follows:

[0114]

[0115] Furthermore, based on this risk probability measure, the risk variance term is defined as follows:

[0116]

[0117] Based on this risk expectation Risk variance and the preset uncertainty constraint function Construct a joint decision function

[0118]

[0119] in:

[0120] This represents the first spatial gradient of the risk scoring field.

[0121] It is the second-order Wasserstein distance.

[0122] It is a predefined normal body surface risk distribution.

[0123] and Represents the weight parameters used in the decision phase.

[0124] By using this joint decision function The results are compared with the preset judgment range, and the risk screening results of congenital scoliosis for the subject are output.

[0125] This invention introduces statistical measures such as probability measures, expectation and variance to model the risk scoring field, so that the screening results are not only based on the average risk level, but also reflect the uncertainty of the risk distribution in the body surface area. This provides a more stable screening judgment basis when facing complex body surface morphology or significant local abnormalities.

[0126] According to one embodiment of the present invention, after constructing the joint decision function Ψ, the method further includes: introducing a probabilistically constrained stochastic optimization process in the risk determination stage to perform a secondary discrimination on the joint decision function, wherein the secondary discrimination process satisfies a constrained optimization model of the following form:

[0127]

[0128]

[0129] in,

[0130] The joint decision function is constructed according to claim 9;

[0131] The risk offset control variable is introduced to describe the adjustment range of the judgment threshold under random disturbances;

[0132] This is the upper bound for determining the risk level;

[0133] This is the preset risk tolerance probability;

[0134] Furthermore, the constraint optimization process solves the problem by transforming the probabilistic constraints into an equivalent Lagrangian form, whose Lagrangian function is expressed as:

[0135]

[0136] in For Lagrange multipliers, The indicator function is used to determine the final judgment interval corresponding to the risk screening result by satisfying the corresponding KKT conditions.

[0137] The embodiments of the present invention introduce a stochastic optimization model with probability constraints on the basis of the joint decision function, so that the risk screening results have clear mathematical decision boundaries while considering the uncertainty of the risk distribution on the body surface, which helps to avoid unstable decision-making by a single threshold under complex or noisy conditions.

[0138] The implementation process of the above embodiments of the present invention will be described in detail below with reference to examples.

[0139] Example 1: Basic Risk Screening Example Based on a Single Back Surface Image

[0140] This example provides a method for screening the risk of congenital scoliosis based on a single back surface image, suitable for scenarios such as pediatric health clinics, school physical examinations, or primary care screenings. This example is characterized by its ease of operation and low equipment requirements, enabling preliminary risk assessment without relying on any imaging examinations.

[0141] In this example, images are captured while the subject is standing naturally. The subject stands with feet naturally apart, arms hanging naturally at their sides, and back fully exposed, avoiding any obstruction by clothing. The acquisition device can be a standard visible light camera or a mobile terminal camera module. The interpretation of back posture can be found in [reference needed]. Figure 3 .

[0142] The acquired back surface images are represented as coordinates defined in two-dimensional space. pixel function To reduce the impact of different lighting conditions and shooting environments on subsequent analysis, the pixel function is normalized to obtain a standardized pixel function. It satisfies:

[0143] in, This represents the pixel region used for analysis in the image of the back of the body. and These represent the mean and standard deviation of pixel intensity calculated within this region, respectively.

[0144] In obtaining the standardized pixel function Then, a first-order partial differential operation is performed on it in spatial coordinates to obtain the gradient vector field:

[0145]

[0146] This gradient vector field reflects the direction and intensity of changes in brightness on the back surface, and can be used to describe left-right asymmetry and local morphological abrupt changes on the back.

[0147] Furthermore, a second-order partial derivative operation is performed on the pixel function to construct a Hessian matrix:

[0148]

[0149] This matrix is ​​used to reflect the bending and changing trends of the body surface morphology in local areas.

[0150] The system will , and By cascading along the feature dimension, a differential feature tensor of body surface morphology is formed:

[0151]

[0152] This feature tensor serves as the model input for subsequent risk analysis.

[0153] The body surface morphology differential feature tensor is input into a pre-trained congenital scoliosis risk screening model. The model is in the body surface domain upper output hidden variable representation And further mapped to a risk scoring field:

[0154]

[0155] By performing spatial integration and normalization on the risk scoring field, a global risk scalar is obtained:

[0156]

[0157] Finally, by comparing this global risk scalar with a preset threshold, the risk screening results for congenital scoliosis of the subjects are output.

[0158] Example 2: An example of enhanced risk screening based on left-right symmetry constraints

[0159] Building upon Example 1, this example introduces analysis of left-right symmetry of the back to further enhance the ability to identify the risk of mild or early scoliosis. This example is particularly suitable for subjects with no obvious clinical manifestations but potential structural abnormalities.

[0160] In this example, besides constructing the body surface morphology differential feature tensor In addition, a mirror mapping of the spinal midline is introduced during the model inference process. Let... Representing a mirror coordinate mapping about the midline of the spine, for any point within the body surface region... Its mirror point is represented as .

[0161] Risk score field output by the model Both training and inference processes satisfy the left-right symmetry constraint, and its constraint terms can be expressed as:

[0162] This constraint is used to quantify the degree of difference in risk distribution between the left and right back regions.

[0163] In this example, the final risk screening result is not only related to the global risk scalar In addition, the degree of left-right asymmetry is also taken into account. When the difference in left-right symmetry exceeds a preset range, the system will mark the corresponding inspected object as a high-risk object.

[0164] Example 3: Statistical Screening Example Based on Risk Distribution Uncertainty

[0165] This example provides a risk screening method for congenital scoliosis based on the statistical characteristics of risk distribution. It is suitable for scenarios where the stability of screening results needs to be evaluated, such as follow-up management or long-term monitoring.

[0166] In this example, the risk scoring field Considered to be defined in the body surface domain Nonnegative random fields are used to construct risk probability measures. Its definition is:

[0167] in, For body surface domain Any subregion.

[0168] Based on the aforementioned probability measure, the system calculates the expected value and variance of risk:

[0169]

[0170] In this example, the system constructs a joint decision functional:

[0171]

[0172] The joint decision functional is then compared with a preset interval to obtain the final risk screening result.

[0173] Example 4: Multiple screening based on follow-up data

[0174] In another example, the congenital scoliosis risk screening method of the present invention can also be applied to scenarios involving multiple body surface image acquisitions of the same subject at different time points. The system can compare and analyze the risk score fields obtained at different time points to help determine the trend of risk changes.

[0175] Set at time The risk score fields obtained are as follows: The system can further calculate the rate of change of risk:

[0176] In addition, the risk scalar sequence obtained from multiple screenings is used to comprehensively assess the risk changes of the tested subjects.

[0177] According to an example of the present invention, a congenital scoliosis risk screening device is also provided. For example... Figure 4 As shown, the device includes: an acquisition module 42 for acquiring a back surface image of the subject, which reflects the surface morphology distribution of the back region; a preprocessing module 44 for preprocessing the back surface image to obtain a standardized surface image, thereby reducing the influence of differences in acquisition posture, imaging conditions, or subject body shape; a construction module 46 for constructing surface morphological feature information for congenital scoliosis risk screening based on the standardized surface image, which includes at least the left-right symmetry of the back, the state of spinal midline deviation, or the trend of surface morphological changes; and a generation module 48 for inputting the surface morphological feature information into a pre-trained congenital scoliosis risk screening model to generate congenital scoliosis risk screening results.

[0178] Furthermore, for the sake of clarity of this invention, all parameters, symbols, operators, and sets are uniformly organized as follows. (1) Spatial and volumetric domain related parameters The dorsal surface region represents the analysis area of ​​the subject's back in a two-dimensional coordinate system; in discrete implementation, it corresponds to the set of pixels involved in the calculation, and in continuous representation, it corresponds to the area of ​​the body surface in a two-dimensional coordinate system. Defined in the body surface domain Two-dimensional spatial coordinates are used to represent pixel positions or normalized coordinate positions in a body surface image.

[0179] (2) Parameters related to body surface images and preprocessing Original back surface image in coordinates Pixel intensity function at Original body surface image The standardized pixel function obtained after brightness normalization and smoothing In the body surface area Inner pixel function Calculated pixel intensity mean In the body surface area Inner pixel function Calculated pixel intensity standard deviation

[0180] (3) Differential operators and surface morphological parameters Pixel function A first-order gradient vector field with respect to spatial coordinates is used to describe the direction and magnitude of changes in surface brightness. Pixel function The Hessian matrix, composed of second-order partial differentials, is used to describe the curvature changes of the body surface morphology within local regions. The L2 norm is used to measure the magnitude of a vector or gradient. The Frobenius norm is used to measure the overall strength of a matrix (such as a Hessian matrix).

[0181] (4) Body surface morphological differential feature tensor The differential feature tensor of body surface morphology, by , and Concatenation along the feature dimension yields The differential feature tensor of body surface morphology in the th Components on each channel Number of channels in the differential feature tensor of body surface morphology

[0182] (5) Model and parameter related symbols Congenital scoliosis risk screening model The set of trainable parameters for the risk screening model The set of hyperparameters used in model training and evaluation Risk screening model in the body surface domain The implicit variable representation of the output above Model output at the Latent variable response on each channel The spatial weighting function used in the model to weight and combine the tensors of body surface morphological features

[0183] (6) Parameters related to risk scores and screening results Mapping latent variables to risk scores using a mapping function. Defined in the body surface domain The continuous risk scoring field on the body surface represents the scoliosis risk level at different locations on the body surface. Risk scoring field In the body surface area The global risk scalar obtained by integration and normalization Threshold set used for risk level determination Number of risk thresholds or risk levels

[0184] (7) Parameters related to probability and statistical modeling Based on risk scoring field Constructed risk probability measure body surface area any measurable subregion Risk probability measurement The calculated expected value of risk Risk probability measurement The calculated risk variance A preset reference risk distribution is used to compare with the risk distribution of the current subject being examined.

[0185] (8) Joint decision functional and distance metric parameters The joint judgment functional is used to generate the final screening judgment by comprehensively considering risk intensity, spatial variation, and distribution differences. Risk scoring field First-order spatial gradient The second-order Wasserstein distance is used to measure the difference between the current risk distribution and the reference distribution. Weight parameters used to balance the smoothing constraints of the risk field Weighting parameters used to balance the distribution distance term

[0186] (9) Symmetry and mapping related parameters A mirror coordinate mapping function for the spinal midline is used to describe the left-right symmetry of the back.

[0187] (10) Training process Sample dataset for model training Datasets used for model validation or hyperparameter optimization Classification or regression loss function of risk screening model Step size parameter for model parameter updates Random disturbances or temperature parameters Random perturbation term following a Gaussian distribution Identity matrix with the same dimensions as model parameters

[0188] In summary, according to the above embodiments of the present invention, a method and device for screening the risk of congenital scoliosis are provided. The present invention acquires images of the back surface and preprocesses these images to reduce the influence of posture, imaging conditions, and body shape differences. Under a unified standard, it constructs morphological feature information reflecting the left-right symmetry of the back, spinal midline deviation, and trends in surface morphological changes. These features are then input into a pre-trained risk screening model for analysis. This achieves objective and consistent screening for the risk of congenital scoliosis without the need for imaging examinations, solving the problems of strong dependence on existing screening methods, insufficient stability, and limited scalability.

[0189] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for screening the risk of congenital scoliosis, characterized in that, include: Acquire images of the back surface of the subject, which are used to reflect the morphological distribution of the back region. The back surface image is preprocessed to obtain a standardized surface image, so as to reduce the influence of differences in acquisition posture, imaging conditions or object body size. Based on the standardized body surface images, body surface morphological feature information for screening the risk of congenital scoliosis is constructed. The body surface morphological feature information includes at least the left and right symmetry of the back, the state of midline deviation of the spine, or the trend of changes in body surface morphology. The surface morphological features are input into a pre-trained congenital scoliosis risk screening model to generate congenital scoliosis risk screening results.

2. The method for screening the risk of congenital scoliosis according to claim 1, characterized in that, Preprocessing of the back surface image includes: The pixel locations of at least two anatomical reference regions, including the shoulder region and the pelvic region, are detected in the back surface image. A human pose reference coordinate system is constructed based on the pixel positions of the anatomical reference region; Based on the human body posture reference coordinate system, perform a two-dimensional or three-dimensional affine transformation on the back surface image to make the back surface image rotate, translate and scale normalize under the human body posture reference coordinate system. The affine transformation parameters are determined by minimizing the sum of squared Euclidean distances between the reference regions before and after the transformation.

3. The method for screening the risk of congenital scoliosis according to claim 2, characterized in that, After performing a two-dimensional or three-dimensional affine transformation on the back surface image, the method further includes: The back surface image is represented as a two-dimensional pixel matrix. The pixel matrix is ​​then subjected to brightness normalization to ensure that it meets the following requirements. ,(x,y)∈Ω in, Ω is the set of pixels in the back surface image that are involved in the statistics and calculations. It is the pixel intensity before brightness normalization. It is the pixel intensity after brightness normalization. and These represent the mean and standard deviation of the pixel intensity of the back surface image, respectively, and are calculated according to... Statistical analysis of internal pixels; After completing the brightness normalization, the pixel matrix... Apply a spatial smoothing filter operator to reduce high-frequency noise.

4. The method for screening the risk of congenital scoliosis according to claim 3, characterized in that, Preprocessing of the back surface image also includes: The back surface image after brightness normalization and smoothing filtering is represented as a scalar pixel function I'(x,y) defined on two-dimensional spatial coordinates (x,y); Based on the first and second partial derivatives of the scalar pixel function in the two-dimensional spatial coordinates, differential features of the body surface morphology are constructed, wherein... Taking the first partial derivative of the pixel function I'(x,y) yields the gradient vector field. ∇I'(x,y)= Take the second-order partial derivative of the pixel function I'(x,y) and construct the Hessian matrix. in exist The second order is continuously differentiable, such that , The pixel function I'(x,y), the gradient vector field ∇I'(x,y), and the Hessian matrix H(I'(x,y)) are concatenated along the feature dimension to form a unified body surface morphology differential feature tensor. T(x,y)=[I′(x,y), ∇I′(x,y), H(I′(x,y))] The differential feature tensor of the body surface morphology is used as a component of the standardized body surface image for the subsequent construction of body surface morphology feature information.

5. The method for screening the risk of congenital scoliosis according to any one of claims 1 to 4, characterized in that, After inputting the aforementioned body surface morphological feature information into the pre-trained congenital scoliosis risk screening model, the method further includes: The congenital scoliosis risk screening model represents the body surface morphology differential feature tensor as defined in the back body surface domain. tensor field on ; The congenital scoliosis risk screening model includes a tensor coding backbone network and a global discriminant head network. The tensor encoding backbone network performs multi-scale tensor convolution and tensor shrinking on the tensor field to obtain the latent variable tensor representation. The tensor shrinks and satisfies in, It is the number of input tensor channels. It is the output channel index. It is the input channel index; It is related to the output channel Input Channel The corresponding spatial weight kernel / attention weight; It is the first after the merger One latent variable component. A morphological consistency mapping based on a Riemannian metric is introduced into the latent variable space, wherein the Riemannian metric is composed of a symmetric positive definite matrix field. Given that for any two points The geodetic distance satisfies in, It is the body surface area The above two points; ； A symmetric positive definite matrix field gives a local metric; It is the path derivative; It is transpose. Furthermore, the global discriminant head network outputs a risk score vector based on the latent variable tensor representation.

6. The method for screening the risk of congenital scoliosis according to claim 5, characterized in that, The congenital scoliosis risk screening model is trained by minimizing a variational composite objective function, which is applied to the sample distribution. The expectation is expressed as in, This is the congenital scoliosis risk screening model. It is the set of trainable parameters of the model. It represents information about the body surface morphology. This indicates a risk label corresponding to the aforementioned body surface morphological features. Among the regularization terms To define in the domain A weighted combination of the Sobolev seminorm and mirror consistency constraints satisfies in, and Let these represent the first-order gradient operator and the Hessian operator with respect to spatial coordinates, respectively. It is an input that corresponds to the morphological features of the body surface. It is a surveillance label used to represent risk level / positive / negative results. It is the training distribution. (T) is the model's output scoring field on the body surface. This represents the mirror coordinate mapping point about the midline of the spine. It is the Frobenius norm. These are hyperparameters used to balance various constraints. It is a mirror coordinate mapping about the midline of the spine.

7. The method for screening the risk of congenital scoliosis according to claim 6, characterized in that, The congenital scoliosis risk screening model is trained using a two-layer optimization framework, wherein... Inner layer optimization is used to update model parameters. The inner layer optimization satisfies: ； Outer layer optimization is used to update the hyperparameter set. The outer layer optimization satisfies: ； In the inner layer iterative update process, the parameter update rule obtained by discretizing the stochastic differential equation is adopted, satisfying... , in, In Dimensions and same It is a set of hyperparameters. It is the step size parameter. It's a temperature parameter. It is a Gaussian random perturbation term. Is with parameters An identity matrix with consistent dimensions.

8. The method for screening the risk of congenital scoliosis according to any one of claims 1 to 4, characterized in that, After completing the forward inference of the pre-trained congenital scoliosis risk screening model, a model defined on the back surface domain is constructed based on the model's response to input body surface features. On the continuous risk scoring field The risk scoring field is derived from the intermediate latent variables output by the model via a mapping function. The transformation yields the result and satisfies the following conditions: ,in The latent variable representation of the model at the corresponding location on the dorsal body surface region; The risk scoring field is located in the dorsal body surface area. Perform spatial integration and normalization to obtain the global risk scalar. Its satisfaction ,in This represents the measurement of the dorsal body surface region; Based on the global risk scalar And combined with a pre-set set of risk assessment rules The risk screening result of congenital scoliosis for the subject is determined by comparing the global risk scalar with the threshold in the set of judgment rules. The risk scoring field Considered as defined in the dorsal body surface region A random field is used to construct a risk probability measure based on the random field. ; where, for any measurable sub-region The risk probability measure satisfies ; Based on the aforementioned risk probability measure, the expected value of risk is defined as follows: Furthermore, based on the aforementioned risk probability measure, the risk variance term is defined as follows: Based on the aforementioned risk expectation Risk variance and the preset uncertainty constraint function Construct a joint decision function in: This represents the first spatial gradient of the risk scoring field. It is the second-order Wasserstein distance. It is a predefined normal body surface risk distribution. and Represents the weight parameters used in the decision phase. By applying the joint decision function The results of the congenital scoliosis risk screening for the tested subject are compared with the preset judgment range and output.

9. The method for screening the risk of congenital scoliosis according to claim 8, characterized in that, After constructing the joint decision function Ψ, the process further includes: introducing a probabilistically constrained stochastic optimization process in the risk assessment stage to perform a secondary discrimination on the joint decision function, wherein the secondary discrimination process satisfies a constrained optimization model of the following form: in, The joint decision function is constructed according to claim 9; The risk offset control variable is introduced to describe the adjustment range of the judgment threshold under random disturbances; This is the upper bound for determining the risk level; This is the preset risk tolerance probability; Furthermore, the constraint optimization process solves the problem by transforming the probabilistic constraints into an equivalent Lagrangian form, whose Lagrangian function is expressed as: in For Lagrange multipliers, The indicator function is used to determine the final judgment interval corresponding to the risk screening result by satisfying the corresponding KKT conditions.

10. A congenital scoliosis risk screening device, characterized in that, include: The acquisition module is used to acquire images of the back surface of the subject, which are used to reflect the morphological distribution of the back region. The preprocessing module is used to preprocess the back surface image to obtain a standardized surface image, so as to reduce the influence of differences in acquisition posture, imaging conditions or object body shape. The construction module is used to construct surface morphological feature information for risk screening of congenital scoliosis based on the standardized surface images. The surface morphological feature information includes at least the left and right symmetry of the back, the state of midline deviation of the spine, or the trend of surface morphological changes. The generation module is used to input the surface morphological feature information into a pre-trained congenital scoliosis risk screening model and generate congenital scoliosis risk screening results.

Images