A Cobb Angle Measurement Method for Scoliosis Based on Few-Sample Semi-Supervised Wavelet Adaptation

Abstract

This application provides a method for measuring the Cobb angle of scoliosis based on a few-sample semi-supervised wavelet adaptive method. Applied to the field of scoliosis angle measurement technology, this method trains a scoliosis angle prediction model using preprocessed labeled source domain images, a small number of labeled target domain images, and a large number of unlabeled target domain images, based on a constructed total target loss function. The spinal images to be processed are input into the trained scoliosis angle prediction model for analysis, obtaining the predicted vertebral body center position and vertebral body contour parameters. The dense contour of the vertebral body is reconstructed through inverse transformation of the contour parameters. Corner information is extracted from the reconstructed vertebral body contour, and the Cobb angle of scoliosis is determined based on the extracted corner information and geometric relationships. This method improves the accuracy and robustness of spinal angle measurement under whole-body images with a small number of labeled samples.

Description

Technical Field

[0001] This application relates to the field of scoliosis angle measurement technology, and in particular to a method for measuring scoliosis Cobb angle based on few-sample semi-supervised wavelet adaptation. Background Technology

[0002] Currently, automatic measurement methods for spinal curvature angle are mainly based on feature extraction and geometric relationship modeling of vertebral structures in spinal images. The spinal curvature angle is calculated by analyzing key points, contour areas, or overall morphology of the vertebral body.

[0003] Traditional keypoint detection or region segmentation methods can achieve certain results in images with clear spinal structures and simple backgrounds, but they are easily affected by non-spinal regions and have poor stability in whole-body images or complex backgrounds. While improved deep learning-based methods have improved measurement accuracy to some extent, they usually rely on manually cropped regions of interest in the spine and require a large amount of labeled data, making them difficult to adapt to real-world clinical scenarios with limited labeled samples.

[0004] In addition, existing methods mostly model at the pixel or point level, lacking explicit constraints on the geometry of the cone, which leads to unstable measurement results under complex or severely curved conditions. Summary of the Invention

[0005] This application provides a few-sample semi-supervised wavelet adaptive method for measuring the Cobb angle of scoliosis, which addresses the problems of insufficient stability in whole-body imaging, strong dependence on labeled data, and insufficient constraints on vertebral geometry in existing automatic methods for measuring spinal curvature angle. The method includes:

[0006] Construct a training dataset, which includes: labeled images in the source domain, a small number of labeled images in the target domain, and a large number of unlabeled images in the target domain;

[0007] Preprocessing is performed on source domain labeled images, target domain with a small number of labeled images, and target domain with a large number of unlabeled images;

[0008] The scoliosis angle prediction model was trained using preprocessed source domain labeled images, target domain partially labeled images, and target domain largely unlabeled images based on the constructed total target loss function.

[0009] The spinal images to be processed are input into the trained scoliosis angle prediction model for analysis and processing to obtain the predicted vertebral body center position and vertebral body contour parameter representation, and the dense contour of the vertebral body is reconstructed through the inverse transformation of the contour parameters.

[0010] Angular information is extracted from the reconstructed vertebral contour, and the Cobb angle of scoliosis is determined based on the extracted angular information and geometric relationships.

[0011] Optionally, the step of inputting the spinal image to be processed into the trained scoliosis angle prediction model for analysis and processing to obtain the predicted vertebral body center position and vertebral body contour parameters includes:

[0012] For each vertebra in the spinal image to be processed, the circumscribed quadrilateral of the vertebra is determined by arranging the four corner points in a clockwise order, and uniform sampling is performed along the circumscribed quadrilateral of the vertebra to obtain a dense contour point sequence.

[0013] The centroid of the dense contour point sequence is determined as the center position of the cone. The offset of each sampling point in the dense contour point sequence relative to the center position of the cone is calculated to obtain the centered contour offset vector.

[0014] The centered contour offset vector is normalized according to the preset size of the output feature map to obtain the normalized contour offset vector.

[0015] The normalized contour offset vector is converted into contour signals in the horizontal direction and contour signals in the vertical direction.

[0016] Discrete wavelet multiscale transform is performed on the contour signals in the lateral and longitudinal directions to obtain the contour parameter representation of the vertebral body.

[0017] Optionally, the reconstruction of the dense contour of the vertebral body through inverse transformation of contour parameters includes:

[0018] The contour parameters predicted by the scoliosis angle prediction model are represented by inverse multi-scale transformation to obtain the contour signals in the transverse and longitudinal directions after inverse transformation.

[0019] Based on the preset size of the output feature map, the contour signals in the horizontal and vertical directions after inverse transformation are denormalized to obtain the denormalized centered contour offset vector.

[0020] The inversely normalized centered contour offset vector is added to the vertebral body center position predicted by the scoliosis angle prediction model to obtain the dense contour reconstruction result of the vertebral body.

[0021] Optionally, corner information can be extracted from the reconstructed vertebral contour, including:

[0022] Search for points that satisfy the preset linear combination minimum value condition in the reconstructed dense contour point set, and obtain the predicted coordinates of the top left corner point, the top right corner point, the bottom right corner point, and the bottom left corner point;

[0023] The predicted coordinates of the top left corner, top right corner, bottom right corner, and bottom left corner satisfy the following formula:

[0024] ;

[0025] ;

[0026] in, For extremum search operators, Predict the coordinates of the top left corner point. Predict the coordinates of the top right corner point. Predict the coordinates of the bottom right corner point. Predict the coordinates of the bottom left corner point. for The outline points of the direction, for The outline points of the direction, To reconstruct a dense contour point set.

[0027] Optionally, the Cobb angle of scoliosis can be determined based on the extracted corner point information and geometric relationships, including:

[0028] The midpoint of the left side of the vertebral body is calculated based on the predicted coordinates of the upper left corner and the lower left corner, and the midpoint of the right side of the vertebral body is calculated based on the predicted coordinates of the upper right corner and the lower right corner.

[0029] The direction vector of the vertebral body is determined based on the midpoint positions of the left and right sides of the vertebral body.

[0030] The Cobb angle of scoliosis is determined by the angle between the directional vectors of the vertebrae in the spine.

[0031] Optionally, the step of training the scoliosis angle prediction model using preprocessed source domain labeled images, target domain partially labeled images, and target domain largely unlabeled images based on the constructed total target loss function includes:

[0032] To address the large number of unlabeled images in the target domain, a collaborative mechanism between the teacher model and the student model is introduced during the training process to construct a pseudo-supervised loss for unlabeled samples. In this process, the student model updates its parameters through backpropagation, while the parameters of the teacher model are obtained by the exponential moving average of the student model's parameters.

[0033] For the labeled images in the source and target domains, the scoliosis angle prediction model is trained using a standard supervised method, and a supervised training loss for labeled samples is constructed.

[0034] The overall objective loss function is constructed based on the pseudo-supervision loss of unlabeled samples, the supervised training loss of labeled samples, and the cross-domain alignment loss.

[0035] The parameters of the scoliosis angle prediction model are updated by minimizing the total objective loss function using the stochastic gradient descent algorithm until the preset number of iterations is reached or the total objective loss function converges, thus obtaining the trained scoliosis angle prediction model.

[0036] Optionally, for a large number of unlabeled images in the target domain, a collaborative mechanism between the teacher model and the student model is introduced during the training process to construct a pseudo-supervised loss for unlabeled samples, including:

[0037] During training, the teacher model is used to predict the vertebral body center position and contour parameters for each unlabeled image in the target domain, and the dense contour of the vertebral body is reconstructed based on the prediction results.

[0038] The corner points of the vertebrae are extracted from the dense contour of the vertebrae reconstructed from the teacher model, and then uniform sampling is performed based on the corner points to obtain a regularized contour.

[0039] The contour scale is calibrated by comparing the area of ​​the original predicted contour of the teacher model with that of the regularized contour.

[0040] The scale-calibrated regularized contour is converted back into a contour parameter representation to obtain a geometrically consistent vertebral contour parameter representation.

[0041] The geometrically consistent vertebral contour parameter representation is weighted and fused with the original vertebral contour parameters predicted by the teacher model to obtain pseudo-supervised contour parameters.

[0042] A pseudo-supervised loss for unlabeled samples is constructed based on pseudo-supervised contour parameters. This pseudo-supervised loss for unlabeled samples is used to supervise the training of student models.

[0043] Optionally, the supervised training loss for the labeled samples includes: vertebral center prediction loss, center offset prediction loss, and contour parameter prediction loss; wherein, the contour parameter prediction loss includes contour parameter error and contour geometric error.

[0044] Optionally, the total objective loss function is constructed based on the pseudo-supervised loss of unlabeled samples, the supervised training loss of labeled samples, and the cross-domain alignment loss, satisfying the following formula:

[0045] ;

[0046] in, Let the total objective loss function be... This represents the mathematical expectation operation on a labeled sample set. This represents a labeled sample pair. For the input spinal imaging data, This is the actual annotation information corresponding to the input spinal imaging data. Representative input image The corresponding sample weight coefficients, This represents the supervised training loss for labeled samples. This represents the mathematical expectation operation on the input image sample set. Represents cross-domain alignment loss. Represents the unlabeled loss weight. This represents the pseudo-supervision loss of unlabeled samples.

[0047] This application provides a method for measuring the Cobb angle of scoliosis based on a few-sample semi-supervised wavelet adaptive method. The method constructs a training dataset including labeled source domain images, a small number of labeled target domain images, and a large number of unlabeled target domain images. Preprocessing is performed on these images. A scoliosis angle prediction model is trained using the preprocessed images based on a constructed total objective loss function. The spinal images to be processed are input into the trained scoliosis angle prediction model for analysis, yielding the predicted vertebral body center position and contour parameters. The dense contour of the vertebral body is reconstructed through inverse transformation of the contour parameters. Corner information is extracted from the reconstructed vertebral contour, and the Cobb angle of scoliosis is determined based on the extracted corner information and geometric relationships. This method improves the accuracy and robustness of spinal angle measurement under whole-body images with a small number of labeled samples. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0049] Figure 1 A flowchart illustrating the method for measuring the Cobb angle of scoliosis based on few-sample semi-supervised wavelet adaptation provided in this application embodiment;

[0050] Figure 2 This is a schematic diagram of the structure of the scoliosis angle prediction model provided in the embodiments of this application;

[0051] Figure 3 This is a schematic flowchart illustrating the parametric representation of the vertebral contour provided in an embodiment of this application;

[0052] Figure 4 This is a schematic diagram of the dense contour reconstruction process provided in the embodiments of this application;

[0053] Figure 5 This is a schematic diagram of the process for constructing pseudo-supervision for unlabeled target domain samples provided in an embodiment of this application;

[0054] Figure 6This is a diagram showing the comparison between the predicted scoliosis angle and the actual labeled value.

[0055] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.

[0058] In this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0059] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0060] Figure 1 This is a flowchart illustrating the Cobb angle measurement method for scoliosis based on few-sample semi-supervised wavelet adaptation provided in an embodiment of this application. Figure 1 As shown, the Cobb angle measurement method for scoliosis based on few-sample semi-supervised wavelet adaptation provided in this embodiment includes:

[0061] S1: Construct a training dataset, which includes: labeled images in the source domain, a small number of labeled images in the target domain, and a large number of unlabeled images in the target domain.

[0062] Specifically, the source domain dataset consists of a collection of spinal X-ray images with relatively narrow fields of view and sufficient vertebral annotation, used to provide stable basic geometric and anatomical priors. The target domain dataset consists of a collection of low-dose whole-body spinal images, characterized by a large field of view, complex background structure, and high manual annotation costs. For the target domain dataset, it is further divided into a small set of labeled samples and a large set of unlabeled samples to simulate the training scenario of scarce annotations in real-world applications.

[0063] S2: Preprocess the source domain labeled images, the target domain with a small number of labeled images, and the target domain with a large number of unlabeled images.

[0064] In this embodiment, all input images are uniformly adjusted to a fixed resolution, and uncropped full-body images are consistently used as network input during training and inference, thereby avoiding subjective errors caused by relying on manual cropping or predefined regions of interest. During training, labeled samples from the source domain, labeled samples from the target domain, and unlabeled samples from the target domain are interleaved and input into the network in the same training phase, and jointly optimized through a unified training objective, laying a data foundation for subsequent cross-domain learning and semi-supervised training.

[0065] S3: The scoliosis angle prediction model is trained using the preprocessed source domain labeled images, target domain with a small number of labeled images, and target domain with a large number of unlabeled images, based on the constructed total target loss function.

[0066] Figure 2 This is a schematic diagram of the scoliosis angle prediction model provided in this application embodiment. Specifically, the scoliosis angle prediction model is trained using preprocessed source domain labeled images, target domain partially labeled images, and target domain largely unlabeled images based on a constructed total target loss function, including:

[0067] S31: To address the large number of unlabeled images in the target domain, a collaborative mechanism between the teacher model and the student model is introduced during the training process to construct a pseudo-supervised loss for unlabeled samples.

[0068] Understandably, given the large number of unlabeled whole-body spine image samples in the target domain, this invention introduces a teacher-student collaborative mechanism during training to fully utilize unlabeled data in model optimization. Specifically, during the training phase, a student model and a teacher model are maintained simultaneously. The student model updates its parameters through backpropagation, while the teacher model's parameters are obtained from the exponential moving average of the student model's parameters, thus ensuring that the teacher model remains relatively smooth and stable during training. The update method for the teacher model's parameters is defined as follows:

[0069]

[0070] in, For teacher model parameters, The coefficient of the exponential moving average. These are the parameters for the student model.

[0071] Figure 5 This is a schematic diagram illustrating the process of constructing pseudo-supervision for unlabeled target domain samples according to an embodiment of this application. Specifically, constructing the pseudo-supervision loss for unlabeled samples includes:

[0072] S311: During training, the teacher model is used to predict the vertebral body center position and contour parameters for each unlabeled image in the target domain, and the dense contour of the vertebral body is reconstructed based on the prediction results.

[0073] S312: Extract the vertebral corner points from the dense contour of the reconstructed vertebral body in the teacher model, and resample uniformly based on the corner points to obtain a regularized contour.

[0074] Understandably, to avoid noise accumulation caused by directly using prediction results as pseudo-labels, geometric consistency constraints are further applied to the dense contours predicted by the teacher model. Specifically, cone corner points are first extracted from the predicted contours, and then uniform sampling is performed based on these corner points to obtain a regularized cone contour representation.

[0075]

[0076] in, For a regularized representation of the vertebral body contour, For uniform resampling operators, Represents the corner extraction operator. The contour representation predicted by the teacher model.

[0077] S313: The contour scale is calibrated by comparing the area of ​​the original predicted contour of the teacher model with that of the regularized contour.

[0078] Understandably, after obtaining the regularized contour, the contour scale is calibrated by comparing the area of ​​the original predicted contour with that of the regularized contour to suppress instability caused by scale drift. The scale calibration coefficient is defined as:

[0079]

[0080] in, For scale calibration coefficients, For numerical truncation operators, This is a function for calculating the area of ​​the outline. To prevent tiny constants with a denominator of 0, This is the lower limit for scale calibration. This represents the upper limit of scale calibration.

[0081] S314: Convert the scale-calibrated regularized contour back into a contour parameter representation to obtain a geometrically consistent vertebral contour parameter representation.

[0082] S315: The geometrically consistent vertebral contour parameter representation is weighted and fused with the original vertebral contour parameters predicted by the teacher model to obtain pseudo-supervised contour parameters.

[0083] Specifically, the pseudo-supervised contour parameters satisfy the following formula:

[0084]

[0085] in, For pseudo-supervised contour parameters, For pseudo-supervision fusion weights, The original contour parameters of the teacher model. These are the parameters corresponding to the regularized contour.

[0086] S316: Construct an unlabeled sample pseudo-supervised loss based on pseudo-supervised contour parameters. The unlabeled sample pseudo-supervised loss is used to supervise the training of the student model.

[0087] Specifically, the pseudo-supervision loss for unlabeled samples satisfies the following formula:

[0088]

[0089] in, For unlabeled samples, pseudo-supervision loss, Represents mathematical expectation, Represents the predicted loss at the center of the vertebral body. Represents the center offset prediction loss. The weighting coefficients representing the contour consistency loss, This represents the loss in contour consistency.

[0090] S32: For labeled images in the source and target domains, the scoliosis angle prediction model is trained using a standard supervised method, and a supervised training loss for labeled samples is constructed.

[0091] Understandably, for labeled samples in both the source and target domains, standard supervised training is used to train the model. Specifically, for each labeled sample, the prediction of the cone center, center offset, and contour parameters are simultaneously constrained to ensure the overall performance of the model in terms of localization accuracy and geometric consistency. The corresponding supervised training loss is defined as:

[0092]

[0093] in, The contour loss term represents the supervised training loss for labeled samples. Composed of contour parameter error and contour geometric reconstruction error, it is used to simultaneously constrain the consistency of prediction results in parameter space and image space. Contour parameter error is defined as:

[0094]

[0095] in, For contour parameter error, This represents the number of visible cones in the current input image that participate in the loss calculation. Visible cones refer to the cones within the current image's field of view that can be labeled and participate in supervised training. This is the contour parameter vector of the k-th vertebra predicted by the model. This is the true contour parameter vector of the k-th vertebra. This represents the L1 norm, used to measure the difference between predicted and true parameters.

[0096] The geometry error of the profile is defined as:

[0097]

[0098] in, For contour geometry error, The total number of sampling points on the dense contour of each vertebra. This is the robust error function, used to robustly measure the deviation between the predicted contour points and the true contour points. To predict the coordinates of the nth contour point of the reconstructed k vertebrae, The coordinates of the nth true contour point of the kth vertebra are given.

[0099] The contour consistency loss satisfies the following formula:

[0100]

[0101] in, The weighting coefficients for the contour loss.

[0102] S33: Construct the overall objective loss function based on the pseudo-supervised loss of unlabeled samples, the supervised training loss of labeled samples, and the cross-domain alignment loss.

[0103] Specifically, the overall objective loss function satisfies the following formula:

[0104] ;

[0105] in, Let the total objective loss function be... This represents the mathematical expectation operation on a labeled sample set. Describes a labeled sample pair, where For the input spinal imaging data, The actual annotation information corresponds to the input image. This actual annotation information is the vertebral structure information obtained through manual annotation. Its content may include the coordinates of the vertebral center point, the center offset, the vertebral contour parameters, and one or more of the following: a dense contour point set generated from the vertebral corner points. Representative input image The corresponding sample weight coefficients are used to weight the supervised training loss of labeled samples from different sources, in order to balance the influence of source domain samples and target domain samples in model training. This represents the supervised training loss for labeled samples. This represents the mathematical expectation operation on the input image sample set, where This represents the input spinal imaging data used in model training. It can come from one or more of the following: a set of labeled samples in the source domain, a small set of labeled samples in the target domain, or a set of unlabeled samples in the target domain. Represents cross-domain alignment loss. Represents the unlabeled loss weight. This represents the pseudo-supervision loss of unlabeled samples.

[0106] S34: The parameters of the scoliosis angle prediction model are updated by minimizing the total objective loss function using the stochastic gradient descent algorithm until the preset number of iterations is reached or the total objective loss function converges, thus obtaining the trained scoliosis angle prediction model.

[0107] S4: Input the spinal image to be processed into the trained scoliosis angle prediction model for analysis and processing to obtain the predicted vertebral body center position and vertebral body contour parameter representation, and reconstruct the dense contour of the vertebral body through inverse transformation of the contour parameters.

[0108] Figure 3 This is a flowchart illustrating the parametric representation of the vertebral body contour provided in this application embodiment. Specifically, the spinal image to be processed is input into a trained scoliosis angle prediction model for analysis and processing to obtain the predicted vertebral body center position and the vertebral body contour parameter representation, including:

[0109] S411: For each vertebra in the spinal image to be processed, the circumscribed quadrilateral of the vertebra is determined by arranging the four corner points in a clockwise order, and uniform sampling is performed along the circumscribed quadrilateral of the vertebra to obtain a dense contour point sequence.

[0110] Understandably, for each cone, its circumscribed quadrilateral is described using four corner points, arranged in clockwise order. To obtain a continuous and structurally consistent geometric description, uniform sampling is performed along each edge of the cone quadrilateral, transforming sparse corner point information into an ordered, dense sequence of contour points. Let the number of sampling points on each edge be... Then each vertebra receives a total of A contour point.

[0111] Specifically, for any line from endpoints and A defined boundary segment, its first Each sampling point is defined as:

[0112]

[0113] By sequentially piecing together the sampling points of the four sides, the cone body is obtained. Dense outline representation :

[0114]

[0115] in, For the first Coordinates of dense contour points for vertebral reconstruction.

[0116] S412: Determine the centroid of the dense contour point sequence as the center position of the cone, calculate the offset of each sampling point in the dense contour point sequence relative to the center position of the cone, and obtain the centered contour offset vector.

[0117] Understandably, after obtaining the dense contour, to eliminate the influence of the cone's absolute position in the image, the centroid of the contour point set is first calculated as the center position of the cone, and all contour points are converted into offsets relative to this center. The contour centroid is defined as:

[0118]

[0119] in, No. The center position of each vertebra.

[0120] The corresponding centered contour offset is:

[0121]

[0122] in, This is the centered contour offset vector.

[0123] S413: Normalize the centered contour offset vector according to the preset size of the output feature map to obtain the normalized contour offset vector.

[0124] Understandably, to unify the scale of different cones in the network output space, the aforementioned offsets are normalized according to the scale of the output feature map. Let the size of the output feature map be... Define the scale factor The normalized contour offset is then expressed as:

[0125]

[0126] in, This is the normalized contour offset vector. This is the contour normalization scaling factor.

[0127] S414: Converts the normalized contour offset vector into contour signals in the horizontal and vertical directions.

[0128] Understandably, the two-dimensional normalized contour offset sequence is converted into two one-dimensional signal sequences, corresponding to the horizontal and vertical directions respectively. For the vertebral body... Its one-dimensional contour signal is represented as:

[0129]

[0130] in, It is a one-dimensional contour signal in the horizontal direction. It is a vertical one-dimensional contour signal.

[0131] S415: Perform discrete wavelet multi-scale transformation on the contour signals in the lateral and longitudinal directions to obtain the contour parameter representation of the vertebral body.

[0132] Understandably, applying a depth of [depth value] to the two one-dimensional signals mentioned above is [description value]. The discrete wavelet multi-scale transform is used to obtain approximation coefficients and detail coefficients at different scales, and the coefficients at each scale are concatenated to form a fixed-length vector representation. Finally, the... and The vertebral body is obtained by stitching together the multi-scale coefficients from two directions. The contour parameters are represented as follows:

[0133]

[0134] in, For the first A vector of contour parameters for each vertebra.

[0135] The wavelet basis functions for the discrete wavelet multiscale transform can be, for example, one of the Haar wavelet, Daubechies wavelet, Symlets wavelet, or other wavelet basis functions. In a preferred embodiment, the discrete wavelet transform uses the Daubechies 4 wavelet (db4) as the wavelet basis function, with a decomposition level of 4.

[0136] The purpose of steps S411-S415 is to transform the discrete corner point annotations of the original vertebrae into a parametric contour representation suitable for network regression and geometric constraints.

[0137] Figure 4 This is a schematic flowchart of dense contour reconstruction provided in an embodiment of this application. Specifically, the dense contour of the vertebral body is reconstructed through inverse transformation of contour parameters, including:

[0138] S421: The contour parameters predicted by the scoliosis angle prediction model are represented by inverse multi-scale transformation to obtain the contour signals in the transverse and longitudinal directions after inverse transformation.

[0139] Specifically, for the predicted contour parameters Performing the inverse transform yields:

[0140]

[0141] in, Representing depth Discrete multiscale transformation operator.

[0142] S422: Based on the preset size of the output feature map, the contour signals in the horizontal and vertical directions after inverse transformation are denormalized to obtain the denormalized centered contour offset vector.

[0143] Specifically, the output feature map is inversely normalized according to its scale to recover a centered offset representation in the output grid coordinate system:

[0144]

[0145] in, The inversely normalized centered contour offset vector The scale parameter represents the output feature map and is used to restore the normalized contour signal to the grid coordinate system scale corresponding to the output feature map.

[0146] S423: Add the inversely normalized centered contour offset vector to the vertebral body center position predicted by the scoliosis angle prediction model to obtain the dense contour reconstruction result of the vertebral body.

[0147] Specifically, this offset is compared with the vertebral body center position predicted by the network. Adding them together yields the dense contour reconstruction result of the vertebral body:

[0148]

[0149] in, Representing the Coordinates of dense contour points for vertebral reconstruction.

[0150] S5: Extract corner information from the reconstructed vertebral contour, and determine the Cobb angle of scoliosis based on the extracted corner information and geometric relationships.

[0151] In this embodiment, the Cobb angle of scoliosis is the spinal curvature angle calculated based on the angle between the vertebral direction vectors.

[0152] After obtaining the dense contour reconstruction results, the four corner points are extracted by applying an extreme value criterion to the contour point set. Specifically, points satisfying different linear combination minimum conditions are searched in the contour point set and designated as the top left, top right, bottom right, and bottom left corner points, as defined below:

[0153] ;

[0154] ;

[0155] in, For extremum search operators, Predict the coordinates of the top left corner point. Predict the coordinates of the top right corner point. Predict the coordinates of the bottom right corner point. Predict the coordinates of the bottom left corner point. for The outline points of the direction, for The outline points of the direction, To reconstruct a dense contour point set.

[0156] Furthermore, the Cobb angle of scoliosis is calculated based on geometric relationships. For each vertebra, the midpoint positions on the left and right sides of the vertebra are first calculated based on the predicted upper, lower, left, and right corner positions:

[0157]

[0158] in, The location is the midpoint of the left side of the vertebral body. This is the midpoint of the right side of the vertebral body.

[0159] The directional vector of the vertebral body is formed by the left and right midpoints:

[0160]

[0161] in, The direction vector of the vertebral body.

[0162] The angle between any two vertebrae is calculated using the angle between their direction vectors, specifically expressed as:

[0163]

[0164] in, Vertebral body directional vector, Vertebral body directional vector, Let be the Euclidean distance between vectors.

[0165] Within the entire spine, by analyzing the changes in the angles between the directional vectors of each vertebra and the overall offset of the spinal midline, the location of the main curve and the corresponding zone curvature angle are determined, thereby completing the automatic curvature angle measurement of the whole-body spinal image.

[0166] Figure 6 This diagram illustrates the comparison between the predicted scoliosis angle and the actual labeled value. It shows the correspondence between the predicted scoliosis angle obtained by different methods and the manually labeled actual angle on the test dataset. The reference line represents the ideal state where the predicted value and the actual value are completely consistent. The prediction results of the method of this invention are generally closer to the reference line and have less dispersion, indicating that it has higher measurement accuracy and stability in the automatic measurement of scoliosis angle.

[0167] This application provides a method for measuring the Cobb angle of scoliosis based on few-sample semi-supervised wavelet adaptation. It models the vertebral structure in full-body medical images as a joint optimization problem of the vertebral center position and contour shape. First, each vertebral contour is centered and scale-normalized, converting the two-dimensional discrete contour point sequence into a one-dimensional signal representation. The corresponding contour parameter vector is then obtained through multi-scale orthogonal transformation. Based on this, the vertebral contour reconstruction process is represented as a constrained optimization problem, and a depth-unfolded structure is used to parameterize the optimization process. Each iteration corresponds to a network module, and the proximal operator in the iteration is replaced by a multi-layer convolutional structure. The iterative step size and related control parameters are adaptively learned through training to achieve gradual optimization and reconstruction of the vertebral contour shape. To address the problem of insufficient labeled samples, a pseudo-supervision mechanism based on prediction consistency is introduced. The predicted vertebral contours in unlabeled images are geometrically regularized, and the regularized contours are back-mapped to the contour parameter space to construct pseudo-supervision signals for constraining model training. Finally, the direction of the vertebral endplate is extracted based on the reconstructed vertebral contour, and the angle between adjacent vertebrae is calculated, thereby realizing the automatic measurement of spinal curvature angle. This invention can still work stably under conditions of a small number of labeled samples, significantly improving the accuracy and robustness of spinal angle measurement under whole-body images.

[0168] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0169] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for measuring the Cobb angle of scoliosis based on few-sample semi-supervised wavelet adaptation, characterized in that, The method includes: Construct a training dataset, which includes: labeled images in the source domain, a small number of labeled images in the target domain, and a large number of unlabeled images in the target domain; Preprocessing is performed on source domain labeled images, target domain with a small number of labeled images, and target domain with a large number of unlabeled images; The scoliosis angle prediction model is trained using preprocessed source domain labeled images, target domain with a small number of labeled images, and target domain with a large number of unlabeled images, based on a constructed overall target loss function. This training includes: During training, the teacher model is used to predict the vertebral body center position and contour parameters for each unlabeled image in the target domain, and the dense contour of the vertebral body is reconstructed based on the prediction results. The corner points of the vertebrae are extracted from the dense contour of the vertebrae reconstructed from the teacher model, and then uniform sampling is performed based on the corner points to obtain a regularized contour. The contour scale is calibrated by comparing the area of ​​the original predicted contour of the teacher model with that of the regularized contour. The scale-calibrated regularized contour is converted back into a contour parameter representation to obtain a geometrically consistent vertebral contour parameter representation. The geometrically consistent vertebral contour parameter representation is weighted and fused with the original vertebral contour parameters predicted by the teacher model to obtain pseudo-supervised contour parameters. A pseudo-supervised loss for unlabeled samples is constructed based on pseudo-supervised contour parameters. This pseudo-supervised loss is used to supervise the training of the student model. The student model updates its parameters through backpropagation, and the parameters of the teacher model are obtained by the exponential moving average of the student model parameters. For the labeled images in the source and target domains, the scoliosis angle prediction model is trained using a standard supervised method, and a supervised training loss for labeled samples is constructed. The overall objective loss function is constructed based on the pseudo-supervision loss of unlabeled samples, the supervised training loss of labeled samples, and the cross-domain alignment loss. The parameters of the scoliosis angle prediction model are updated by minimizing the total objective loss function using the stochastic gradient descent algorithm until the preset number of iterations is reached or the total objective loss function converges, thus obtaining the trained scoliosis angle prediction model. The process involves inputting the spinal image to be processed into a trained scoliosis angle prediction model for analysis and processing to obtain the predicted vertebral body center position and vertebral body contour parameter representation. Then, a dense vertebral body contour is reconstructed through inverse transformation of the contour parameters. This process includes: For each vertebra in the spinal image to be processed, the circumscribed quadrilateral of the vertebra is determined by arranging the four corner points in a clockwise order, and uniform sampling is performed along the circumscribed quadrilateral of the vertebra to obtain a dense contour point sequence. The centroid of the dense contour point sequence is determined as the center position of the cone. The offset of each sampling point in the dense contour point sequence relative to the center position of the cone is calculated to obtain the centered contour offset vector. The centered contour offset vector is normalized according to the preset size of the output feature map to obtain the normalized contour offset vector. The normalized contour offset vector is converted into contour signals in the horizontal direction and contour signals in the vertical direction. Discrete wavelet multiscale transform is performed on the contour signals in the lateral and longitudinal directions to obtain the contour parameter representation of the vertebral body. The contour parameters predicted by the scoliosis angle prediction model are represented by inverse multi-scale transformation to obtain the contour signals in the transverse and longitudinal directions after inverse transformation. Based on the preset size of the output feature map, the contour signals in the horizontal and vertical directions after inverse transformation are denormalized to obtain the denormalized centered contour offset vector. The inversely normalized centered contour offset vector is added to the vertebral body center position predicted by the scoliosis angle prediction model to obtain the dense contour reconstruction result of the vertebral body. Angular information is extracted from the reconstructed vertebral contour, and the Cobb angle of scoliosis is determined based on the extracted angular information and geometric relationships.

2. The method according to claim 1, characterized in that, Corner information is extracted from the reconstructed vertebral contour, including: Search for points that satisfy the preset linear combination minimum value condition in the reconstructed dense contour point set, and obtain the predicted coordinates of the top left corner point, the top right corner point, the bottom right corner point, and the bottom left corner point; The predicted coordinates of the top left corner, top right corner, bottom right corner, and bottom left corner satisfy the following formula: ； ； in, For extremum search operators, Predict the coordinates of the top left corner point. Predict the coordinates of the top right corner point. Predict the coordinates of the bottom right corner point. Predict the coordinates of the bottom left corner point. for The outline points of the direction, for The outline points of the direction, To reconstruct a dense contour point set.

3. The method according to claim 2, characterized in that, The Cobb angle of scoliosis is determined based on the extracted corner point information and geometric relationships, including: The midpoint of the left side of the vertebral body is calculated based on the predicted coordinates of the upper left corner and the lower left corner, and the midpoint of the right side of the vertebral body is calculated based on the predicted coordinates of the upper right corner and the lower right corner. The direction vector of the vertebral body is determined based on the midpoint positions of the left and right sides of the vertebral body. The Cobb angle of scoliosis is determined by the angle between the directional vectors of the vertebrae in the spine.

4. The method according to claim 1, characterized in that, The supervised training loss for labeled samples includes: vertebral body center prediction loss, center offset prediction loss, and contour parameter prediction loss; wherein, the contour parameter prediction loss includes contour parameter error and contour geometric error.

5. The method according to claim 1, characterized in that, The overall objective loss function is constructed based on the pseudo-supervised loss of unlabeled samples, the supervised training loss of labeled samples, and the cross-domain alignment loss, satisfying the following formula: ； in, Let the total objective loss function be... This represents the mathematical expectation operation on a labeled sample set. This represents a labeled sample pair. For the input spinal imaging data, This is the actual annotation information corresponding to the input spinal imaging data. Representative input image The corresponding sample weight coefficients, This represents the supervised training loss for labeled samples. This represents the mathematical expectation operation on the input image sample set. Represents cross-domain alignment loss. Represents the unlabeled loss weight. This represents the pseudo-supervision loss of unlabeled samples.

Images