Scoliosis screening method and system, electronic equipment and storage medium

Abstract

The invention discloses a scoliosis screening method and system, electronic equipment and a storage medium, and the method comprises the steps: obtaining a back image of a subject, and obtaining a binary contour mask and a key point set; performing region enhancement processing on the back image through learnable gating parameters according to the binary contour mask to obtain an enhanced image, dividing a key point set into a plurality of channels according to semantic categories, and performing spline curve fitting on key points of each channel to generate a multi-channel dense geometric canvas; according to the enhanced image and the multi-channel dense geometric canvas, respectively obtaining a texture feature vector and a geometric feature vector, and carrying out attention fusion on the texture feature vector and the geometric feature vector to obtain a fusion feature vector; and inputting the fusion feature vector into a conditional ordinal number regression prediction classification head, and outputting a scoliosis screening result. According to the method, through soft gating enhancement, feature fusion and ordinal regression classification, the anti-interference capability, accuracy and clinical suitability of screening are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence model collaboration technology, and in particular to a scoliosis screening method, system, electronic device and storage medium. Background Technology

[0002] Adolescent idiopathic scoliosis (AIS) is the most common type of structural scoliosis in adolescents. It refers to a complex three-dimensional deformity in which one or more segments of the spine curve laterally and are accompanied by vertebral rotation. Its etiology is not fully understood, but it may be related to a variety of factors such as genetics, skeletal development, and endocrine regulation.

[0003] The core of clinical diagnosis and treatment of AIS lies in early detection, early diagnosis, and early intervention. Screening, as a crucial link in early identification, directly impacts subsequent treatment outcomes and patient prognosis through its efficiency and accuracy. With the deep application of artificial intelligence technology in the medical field, non-invasive screening solutions based on computer vision have gradually become a research hotspot, aiming to address the pain points of traditional screening methods, such as strong subjectivity, radiation risks, and low efficiency. Existing AI-assisted screening technologies are mostly based on human back images, extracting key semantic points such as shoulders and waist through human posture detection algorithms, and then combining image features to assess scoliosis risk. However, these technologies still have significant limitations: on the one hand, the back region is not effectively separated from the background and non-back tissues (head, limbs) in the original image, and background noise easily interferes with the extraction of key features, resulting in insufficient accuracy in identifying mild scoliosis; on the other hand, the extracted key points are mostly discretely distributed, easily affected by shooting angle, posture changes, and positioning noise, making it difficult to form a continuous and stable back structural representation, and unable to accurately depict the asymmetrical morphological features of the back caused by scoliosis, thus affecting the reliability and clinical applicability of the screening results. Summary of the Invention

[0004] To address the shortcomings mentioned above, this application provides a scoliosis screening method, system, electronic device, and storage medium.

[0005] Firstly, to achieve the above objectives, this application provides a method for screening scoliosis, comprising: The back image of the subject is obtained, and the corresponding binary contour mask and the set of key points extracted from the back image are obtained. Based on the binary contour mask, the back image is enhanced by learning gating parameters to obtain the enhanced image. The key point set is divided into multiple channels according to semantic category. Spline curve fitting is performed on the key points of each channel to generate a multi-channel dense geometric canvas. Based on the enhanced image and the multi-channel dense geometry canvas, texture feature vectors and geometric feature vectors are obtained respectively, and attention is applied to fuse the texture feature vectors and geometric feature vectors to obtain the fused feature vector. The fused feature vector is input into the conditional ordinal regression prediction classifier, and the scoliosis screening results are output.

[0006] Furthermore, based on the binary contour mask, region enhancement processing is performed on the back image using learnable gating parameters to obtain the enhanced image, including: ; In the formula: ⊙ represents element-wise multiplication. For the binary contour mask of the back region, Here, α is the sigmoid function, and α is a learnable gating parameter. Image of the back; This is the enhanced image.

[0007] Furthermore, spline curve fitting is performed on the key points of each channel to generate a multi-channel dense geometric canvas, including: ; In the formula: To obtain spline fitting curves by performing spline curve fitting on the key points of each channel; For semantic categories; The space broadening used to control the Gaussian response is a constant; For pixels To the fitted curve The shortest Euclidean distance.

[0008] Furthermore, based on the enhanced image and the multi-channel dense geometric canvas, texture feature vectors and geometric feature vectors are obtained, including: Multi-scale feature extraction is performed on the enhanced image using a deep learning network to obtain a first feature map and its corresponding first feature, and a second feature map and its corresponding second feature. The first feature includes local fine-grained information, and the second feature includes global semantic information. Based on the first feature, a first weight vector is constructed along the height and width dimensions of the first feature map, and the first weight vector is fused with the first feature map to obtain the first enhanced feature; Based on the second feature, a second weight vector is constructed along the height and width dimensions of the second feature map, and the second weight vector is fused with the second feature map to obtain the second enhanced feature; The first and second enhancement features are aggregated to obtain the texture feature vector; A multi-channel dense geometric canvas is input into a convolutional neural network, and the number of input channels of the convolutional neural network is equal to the number of channels of the multi-channel dense geometric canvas. The output is an implicit feature vector. Input the set of key points into a convolutional neural network and output an explicit feature vector; The implicit eigenvectors and explicit eigenvectors are concatenated to obtain the geometric eigenvectors.

[0009] Furthermore, the explicit feature vectors include a 24-dimensional clinical geometric descriptor related to horizontal distance statistics, curve length asymmetry, local curvature, and midline deviation.

[0010] Furthermore, the texture feature vector and the geometric feature vector are fused by attention to obtain a fused feature vector, including: By projecting the feature modality, the texture feature vector and the geometric feature vector are aligned, and then concatenated according to the channel dimension to obtain the modal feature vector; The modal feature vectors are input into the attention-aware network to obtain the dynamic weight coefficients of the texture feature vectors and the dynamic weight coefficients of the geometric feature vectors. The aligned texture feature vector and geometric feature vector are fused by attention according to the corresponding dynamic weight coefficients to obtain the fused feature vector.

[0011] Furthermore, the fused feature vectors are input into the conditional ordinal regression prediction classifier head, and the output scoliosis screening results include: The conditional ordinal regression prediction classifier divides scoliosis into multiple tasks according to severity and generates a corresponding number of sub-tasks. The fused feature vector is input into the conditional ordinal regression prediction classifier, and the output is the conditional probability corresponding to the sub-task. Based on conditional probability, scoliosis screening results including multi-level tasks are obtained.

[0012] Secondly, this application also provides a scoliosis screening system, comprising: The acquisition module is used to acquire images of the subject's back and obtain a binary contour mask corresponding to the back image, as well as a set of key points extracted from the back image; The processing module is used to perform region enhancement processing on the back image based on the binary contour mask and learnable gating parameters to obtain the enhanced image. The key point set is divided into multiple channels according to semantic categories, and spline curve fitting is performed on the key points of each channel to generate a multi-channel dense geometric canvas. Based on the enhanced image and the multi-channel dense geometric canvas, texture feature vectors and geometric feature vectors are obtained respectively, and attention is used to fuse the texture feature vectors and geometric feature vectors to obtain the fused feature vector. The execution module is used to input the fused feature vector into the conditional ordinal regression prediction classification head and output the scoliosis screening results.

[0013] Thirdly, this application also provides an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the above-described method.

[0014] Fourthly, this application also provides a storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: This application divides key points into multiple channels according to semantic categories. It generates a multi-channel dense geometric canvas through spline curve fitting and Gaussian kernel rendering, transforming discrete semantic key points into a continuous and smooth structural feature field. This overcomes the shortcomings of discrete key points being susceptible to changes in shooting angle, pose, and positioning noise, effectively offsetting feature fluctuations caused by local key point jitter. This allows for a more accurate depiction of the asymmetrical back morphology caused by scoliosis, improving the stability and reliability of structural representation. This application extracts texture features from the enhanced RGB image and geometric features from the dense geometric canvas, respectively, and achieves complementary advantages between the two types of features through attention fusion. Compared to screening methods that rely solely on texture or geometric features, this application's dual-stream feature fusion strategy can comprehensively capture the pathological manifestations of scoliosis, preserving subtle differences in surface texture while also considering morphological distortions in deep structures, significantly improving the accuracy and generalization ability of screening results. Attached Figure Description

[0016] Figure 1 This is a flowchart of the scoliosis screening method of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments 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.

[0018] Reference Figure 1 This application provides a method for screening scoliosis, comprising: S101: Obtain the back image of the subject and obtain the binary contour mask corresponding to the back image, as well as the set of key points extracted from the back image; As one implementation method, images of the subject's back are acquired using terminals such as mobile phones, tablets, and computers. Semantic segmentation of the human back region is performed on these images, automatically distinguishing the target back region from the background region to obtain pixel-level binary segmentation results. The pixel value of the target back region is set to 1, and the background region is set to 0. A pre-trained keypoint detection network is then used to detect semantic keypoints in the enhanced image, forming a keypoint set.

[0019] For example, RGB images of the subject's back were acquired using mobile phones, tablets, computers, or other terminals. During acquisition, the subject was required to stand naturally with their back exposed or wearing tight-fitting, plain clothing to avoid obscuring the spine-related areas. The acquired raw RGB images were standardized to 224×224 pixels, and color gamut calibration and brightness normalization were performed to eliminate image differences caused by different shooting devices and lighting conditions. A pre-trained SAM segmentation model was used to perform semantic segmentation of the human back region on the standardized RGB images. The SAM segmentation model learns from the back skin... The system automatically distinguishes the target back region from the background region (clothing, sheets, furniture, etc.) based on texture and human contour features, outputting pixel-level binary segmentation results. The target back region pixel values ​​are set to 1, and the background region to 0, forming an initial binary contour mask with the same size as the input RGB image. Morphological optimization is performed on the initial binary contour mask, filling small holes inside the mask through dilation operations using 3×3 structuring elements, removing isolated noise points at the edges through erosion operations, and finally eliminating jagged artifacts at the segmentation edges using a contour smoothing algorithm, resulting in a final binary contour mask with smooth edges and complete regions. A pre-trained YOLOv8 detection network is used to detect back images, extracting anatomical semantic key points such as the left shoulder, right shoulder, left waist, and right waist, forming a key point set.

[0020] With the above settings, the binary contour mask generated by semantic segmentation clarifies the effective range of subsequent feature extraction, avoids background clutter and non-back tissues interfering with the extraction of spine-related features, and effectively separates the target area from the background.

[0021] S102: Based on the binary contour mask, the back image is enhanced by learning gating parameters to obtain the enhanced image. The key point set is divided into multiple channels according to semantic categories. Spline curve fitting is performed on the key points of each channel to generate a multi-channel dense geometric canvas. As one implementation method, learnable gating parameters are pre-set, and the binary contour mask is broadcast in the channel dimension to match the back image dimension. A smooth gain field is generated according to the learnable gating parameters, and the back image is fused with the smooth gain field to achieve soft amplification of pixel values ​​in the human body area and relative suppression of the background area, resulting in an enhanced image.

[0022] For example, the enhanced image can be obtained by selecting the following formula: ; In the formula: ⊙ represents element-wise multiplication. For the binary contour mask of the back region, Here, α is the sigmoid function, and α is a learnable gating parameter. Image of the back; This is the enhanced image.

[0023] As one implementation method, the key point set is divided into corresponding channels according to the extracted semantic key point categories. Spline curve fitting is performed on the key points of each channel to generate continuous spline fitting curves, and a multi-channel dense geometric canvas is generated based on the Gaussian kernel function.

[0024] For example, a pre-trained YOLOv8 detection network is used to detect back images, extracting anatomical semantic key points such as the left shoulder, right shoulder, left waist, and right waist to form a key point set. The key point set is divided into four independent channels according to semantic categories, corresponding to the left shoulder channel, right shoulder channel, left waist channel, and right waist channel, respectively. For the key points in each channel, a cubic spline curve fitting algorithm is used to construct a continuous spline fitting curve, realizing the continuous representation of discrete key points. Dense response maps of the four channels are generated based on a Gaussian kernel function, which can be selected using the following formula: ; In the formula: To obtain spline fitting curves by performing spline curve fitting on the key points of each channel; For semantic categories; The space broadening used to control the Gaussian response is a constant; For pixels To the fitted curve The shortest Euclidean distance.

[0025] Through the above settings, the soft gating mechanism maintains the gradient continuity of the human body contour edges, eliminates unnatural high-frequency hard edge artifacts, avoids interference with the spectral response of the convolutional neural network, and improves the accuracy of feature extraction. Separating channels according to semantic categories allows for the separate encoding of structural features of the shoulders and waist, avoiding interference between different semantic features and more accurately depicting the local asymmetry caused by scoliosis. The dense geometric canvas possesses spatial continuity and noise resistance, providing stable input for feature extraction of geometrically symmetrical flow, compensating for the lack of spatial context in key points, and improving the reliability of geometric feature modeling.

[0026] S103: Based on the enhanced image and the multi-channel dense geometry canvas, obtain the texture feature vector and the geometric feature vector respectively, and then perform attention fusion on the texture feature vector and the geometric feature vector to obtain the fused feature vector; As one implementation, the enhanced image is input into a deep learning network to obtain a first feature map and its corresponding first feature, and a second feature map and its corresponding second feature. The first feature includes local fine-grained information, and the second feature includes global semantic information. Based on the first feature, a first weight vector is constructed along the height and width dimensions of the first feature map, and the first weight vector is fused with the first feature map to obtain a first enhanced feature. Based on the second feature, a second weight vector is constructed along the height and width dimensions of the second feature map, and the second weight vector is fused with the second feature map to obtain a second enhanced feature. The first enhanced feature and the second enhanced feature are aggregated to obtain a texture feature vector.

[0027] A multi-channel dense geometric canvas is input into a convolutional neural network, with the input channels of the convolutional neural network being equal to the channels of the multi-channel dense geometric canvas, and the output is an implicit feature vector; a set of key points is input into the convolutional neural network, and the output is an explicit feature vector; the implicit feature vector and the explicit feature vector are concatenated to obtain a geometric feature vector.

[0028] By projecting the feature modality, the texture feature vector and the geometric feature vector are aligned and then concatenated according to the channel dimension to obtain the modal feature vector. The modal feature vector is then input into an attention-based perceptual network to obtain the dynamic weight coefficients of the texture feature vector and the geometric feature vector. The aligned texture feature vector and the geometric feature vector are then fused by attention according to the corresponding dynamic weight coefficients to obtain the fused feature vector.

[0029] For example, the enhanced image is input into the ConvNeXt-Tiny backbone network, and shallow edge features, mid-level fine-grained features (Stage 3 output), and deep semantic features (Stage 4 output) are extracted sequentially. The mid-level features contain local fine-grained information, and the deep features contain global semantic information. Coordinate attention modules are introduced onto the feature maps output from Stage 3 and Stage 4, respectively, decomposing the two-dimensional global pooling into one-dimensional feature encodings along the height and width directions to generate a direction-aware attention weight tensor A. h and A w The enhanced features are obtained by fusing them with the original feature map through element-wise multiplication. Global average pooling is then performed on the attention-enhanced Stage3 and Stage4 feature maps to compress the high-dimensional feature maps into low-dimensional vectors, which are then concatenated along the channel dimension to obtain the texture feature vector.

[0030] A 4-channel dense geometric canvas is input into a modified ResNet-18 network (the number of input channels for the first convolutional layer is adjusted from 3 to 4 to adapt to multi-channel input). Global topological features of the shoulder and waist contours are extracted through residual block groups of the network, and a 512-dimensional implicit topological feature vector is output after global average pooling. Based on the keypoint set, four sets of 24-dimensional clinical geometric descriptors are calculated, including horizontal distance statistics, curve length asymmetry, local curvature, and midline deviation. Layer normalization is performed on the descriptors to eliminate dimensional differences. The normalized 24-dimensional clinical descriptors are input into a multilayer perceptron and mapped to a 64-dimensional explicit expert feature vector. The implicit topological feature vector and the explicit expert feature vector are then concatenated along the channel dimension to obtain the geometric feature vector.

[0031] A dedicated linear projection head for texture features and a dedicated linear projection head for geometric features are constructed. Each projection head contains a linear layer, a batch normalization layer, and a non-linear activation function. The texture and geometric feature vectors are mapped to a 512-dimensional shared semantic space, resulting in aligned feature vectors. The aligned features are concatenated along the channel dimension and input into an attention-based perceptual network (consisting of two fully connected layers, a GELU activation function, and layer normalization). The network learns the modal reliability association of the current sample and outputs the original weight scores. After Softmax normalization, the weights of the texture and geometric features are summed to equal 1. A weighted summation operation is performed on the aligned feature vectors to obtain the fused feature vector.

[0032] Through the above settings, multi-scale feature encoding and aggregation preserve fine-grained texture details such as skin wrinkles and local shadows while capturing global semantic patterns such as scapular shadow distribution, overcoming the shortcomings of traditional methods that rely solely on sparse points and lack texture information. The continuous nature of the dense geometric canvas reduces the dependence on keypoint localization accuracy, effectively offsetting noise caused by keypoint jitter, and is more robust than traditional multinomial fitting. Cross-modal projection alignment resolves the differences in dimensionality and semantic distribution between texture and geometric features, providing a unified semantic basis for attention fusion and avoiding feature conflicts.

[0033] S104: Input the fused feature vector into the conditional ordinal regression prediction classifier and output the scoliosis screening results.

[0034] As one implementation method, the conditional ordinal regression prediction classifier divides scoliosis into multi-level tasks according to severity (e.g., healthy, mild, moderate, severe) and generates a corresponding number of sub-tasks. The fused feature vector is input into the conditional ordinal regression prediction classifier, and the conditional probability corresponding to the sub-task is output. Based on the conditional probability, the scoliosis screening results including multi-level tasks are obtained.

[0035] For example, the fused feature vector is input into the prediction head of the conditional ordinal regression model. This prediction head transforms the 4-level scoliosis grading task into 3 ordered binary subtasks, outputting 3 conditional logic values, corresponding to the judgment logic of whether the sample level is greater than healthy, greater than mild, and greater than moderate, respectively. Each conditional logic value is converted into a conditional probability using the Sigmoid function. The level is determined based on the set of conditional probabilities: if the first conditional probability ≥ 0.5, it is judged as severe; if the second conditional probability ≥ 0.5 and the first conditional probability < 0.5, it is judged as moderate; if the third conditional probability ≥ 0.5 and the second conditional probability < 0.5, it is judged as mild; if the third conditional probability < 0.5, it is judged as healthy. The final output is a screening result conforming to the ordinal relationship of healthy, mild, moderate, and severe.

[0036] For example, the predictor first maps the fused feature vector into three ordered conditional logic values, corresponding to binary judgment logics for whether the sample level is greater than healthy, whether the sample level is greater than mild, and whether the sample level is greater than moderate; then, the Sigmoid function is used to convert each conditional logic value into a conditional probability in the interval [0,1], denoted as: P > Healthy: The probability that the sample level is greater than healthy (i.e., the probability of being mild, moderate, or severe). P > Mild: The probability that the sample level is greater than mild (i.e., the probability of being moderate or severe). P > Moderate: The probability that the sample grade is greater than moderate (i.e., the probability of belonging to severe).

[0037] By multiplying and complementing conditional probabilities, the absolute probability of a sample belonging to one of the four levels is derived. The core principle is that the probability of the current level = the probability of not being greater than the current level × the probability of being greater than the previous level (for the health level, which has no previous level, the probability of not being greater than the health level is taken directly). The specific formula is as follows: Absolute probability of health level: P(healthy) = 1 - P(>healthy); (the probability that the sample is no larger than the health level, i.e., only belongs to the healthy category). Absolute probability of mild severity: P(mild) = P(>healthy) × [1-P(>mild)]; (the probability that the sample size is greater than healthy but not greater than mild, i.e., only mild). Absolute probability of moderate severity: P(moderate) = P(>healthy) × P(>mild) × [1-P(>moderate)]; (the probability that the sample is greater than healthy and greater than mild, but not greater than moderate, i.e., only moderate). Absolute probability of severe level: P(severe) = P(>healthy) × P(>mild) × P(>moderate); (The probability that the sample is greater than healthy, mild and moderate, i.e., only severe).

[0038] The four absolute probabilities of healthy, mild, moderate, and severe are compared, and the level with the highest probability value is selected as the final scoliosis screening result to ensure that the judgment result conforms to the orderly logic of clinical levels (healthy < mild < moderate < severe) and avoids logical contradictions across levels.

[0039] With the above settings, the conditional ordinal regression loss imposes a greater penalty on cross-level misjudgments, enabling the model to learn level boundaries more accurately, resulting in a better Quadratic Weighted Kappa (QWK) metric and higher consistency with clinical diagnosis.

[0040] This application also provides a scoliosis screening system, including a data acquisition module, a processing module, and an execution module. The acquisition module is used to acquire the back image of the subject, obtain the binary contour mask corresponding to the back image, and extract the set of key points from the back image. The processing module is used to perform region enhancement processing on the back image based on the binary contour mask and learnable gating parameters to obtain the enhanced image. The key point set is divided into multiple channels according to semantic categories, and spline curve fitting is performed on the key points of each channel to generate a multi-channel dense geometric canvas. Based on the enhanced image and the multi-channel dense geometric canvas, texture feature vectors and geometric feature vectors are obtained respectively, and attention is used to fuse the texture feature vectors and geometric feature vectors to obtain the fused feature vector. The execution module is used to input the fused feature vector into the conditional ordinal regression prediction classification head and output the scoliosis screening results.

[0041] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for screening scoliosis, characterized in that, include: Acquire a back image of the subject and obtain a binary contour mask corresponding to the back image, as well as a set of key points extracted from the back image; Based on the binary contour mask, the back image is enhanced by learning gating parameters to obtain the enhanced image. The key point set is divided into multiple channels according to semantic categories. Spline curve fitting is performed on the key points of each channel to generate a multi-channel dense geometric canvas. Based on the enhanced image and the multi-channel dense geometric canvas, texture feature vectors and geometric feature vectors are obtained respectively, and attention fusion is performed on the texture feature vectors and geometric feature vectors to obtain fused feature vectors; The fused feature vector is input into the conditional ordinal regression prediction classifier to output the scoliosis screening results.

2. The scoliosis screening method according to claim 1, characterized in that, The step of performing region enhancement processing on the back image based on the binary contour mask and using learnable gating parameters to obtain the enhanced image includes: ； In the formula: ⊙ represents element-wise multiplication. This is a binary contour mask for the back region. Here, α is the sigmoid function, and α is a learnable gating parameter. The image is of the back. The enhanced image.

3. The scoliosis screening method according to claim 1, characterized in that, The step of performing spline curve fitting on the key points of each channel to generate a multi-channel dense geometric canvas includes: ； In the formula: To obtain a spline fitting curve by performing spline curve fitting on the key points of each channel; For semantic categories; The space broadening used to control the Gaussian response is a constant; For pixels To the fitted curve The shortest Euclidean distance.

4. The scoliosis screening method according to claim 1, characterized in that, The step of obtaining texture feature vectors and geometric feature vectors based on the enhanced image and the multi-channel dense geometric canvas includes: The enhanced image is subjected to multi-scale feature extraction by a deep learning network to obtain a first feature map and a corresponding first feature, and a second feature map and a corresponding second feature. The first feature includes local fine-grained information, and the second feature includes global semantic information. Based on the first feature, a first weight vector is constructed along the height and width dimensions of the first feature map, and the first weight vector is fused with the first feature map to obtain a first enhanced feature; Based on the second feature, a second weight vector is constructed along the height and width dimensions of the second feature map, and the second weight vector is fused with the second feature map to obtain the second enhanced feature. The first enhancement feature and the second enhancement feature are aggregated to obtain the texture feature vector; The multi-channel dense geometric canvas is input into a convolutional neural network, and the input channels of the convolutional neural network are equal to the channels of the multi-channel dense geometric canvas, and an implicit feature vector is output. The set of key points is input into the convolutional neural network, and an explicit feature vector is output. The implicit feature vector and the explicit feature vector are concatenated to obtain the geometric feature vector.

5. The scoliosis screening method according to claim 4, characterized in that, The explicit feature vectors include a 24-dimensional clinical geometric descriptor related to horizontal distance statistics, curve length asymmetry, local curvature, and midline deviation.

6. The scoliosis screening method according to claim 1, characterized in that, The step of performing attention fusion on the texture feature vector and the geometric feature vector to obtain a fused feature vector includes: By projecting the feature modality, the texture feature vector and the geometric feature vector are aligned, and then concatenated according to the channel dimension to obtain the modal feature vector; The modal feature vector is input into an attention-aware network to obtain the dynamic weight coefficients of the texture feature vector and the dynamic weight coefficients of the geometric feature vector; The aligned texture feature vector and the geometric feature vector are fused by attention according to the corresponding dynamic weight coefficients to obtain the fused feature vector.

7. The scoliosis screening method according to claim 1, characterized in that, The step of inputting the fused feature vector into the conditional ordinal regression prediction classifier and outputting the scoliosis screening result includes: The conditional ordinal regression prediction classification head divides scoliosis into multi-level tasks according to severity and generates a corresponding number of sub-tasks. The fused feature vector is input into the conditional ordinal regression prediction classification head, and the conditional probability corresponding to the sub-task is output. Based on the conditional probability, the scoliosis screening results, including the multi-level tasks, are obtained.

8. A scoliosis screening system, characterized in that, include: The acquisition module is used to acquire the back image of the subject, and obtain the binary contour mask corresponding to the back image, as well as the set of key points extracted from the back image; The processing module is used to perform region enhancement processing on the back image based on the binary contour mask and learnable gating parameters to obtain an enhanced image; divide the key point set into multiple channels according to semantic categories; perform spline curve fitting on the key points of each channel to generate a multi-channel dense geometric canvas; obtain texture feature vectors and geometric feature vectors based on the enhanced image and the multi-channel dense geometric canvas; and perform attention fusion on the texture feature vectors and geometric feature vectors to obtain a fused feature vector. The execution module is used to input the fused feature vector into the conditional ordinal regression prediction classification head and output the scoliosis screening results.

9. An electronic device, characterized in that, The method includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the method described in any one of claims 1 to 7.

10. A storage medium, characterized in that, It stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the method described in any one of claims 1 to 7.

Images