Cervical vertebra key point detection method and device, electronic equipment and storage medium

Abstract

This application provides a method, device, electronic device, and storage medium for cervical spine key point detection, applied in the field of image processing technology. The method includes: inputting cervical spine image data into a pre-trained deep learning model; processing the cervical spine image using the encoder of the deep learning model to obtain first image features; wherein the encoder includes a dual-pooling channel attention module, the dual-pooling channel attention module includes a first branch structure, a second branch structure, and a fusion structure, the first branch structure is used to extract global semantic features, the second branch structure is used to extract local spatial features, and the fusion structure is used to fuse the global semantic features and local spatial features and perform subsequent processing to obtain second image features; the first image features are obtained based on the second image features; and the decoder of the deep learning model is used to process the first image features to obtain the positions of each cervical spine key point, which can improve the accuracy of cervical spine key point detection.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to methods, devices, electronic devices and storage media for detecting key points of the cervical spine. Background Technology

[0002] Cervical spondylosis is one of the most common degenerative spinal diseases worldwide. Detection of cervical spondylosis relies on the detection of key points on the cervical spine. Existing deep learning-based techniques for cervical key point detection use Efficient Channel Attention (ECA) for feature extraction. However, ECA heavily relies on global features, neglecting local details. This leads to inaccurate detection of key points at the affected cervical spine locations when dealing with localized lesions. Summary of the Invention

[0003] This application provides a method, device, electronic device, and storage medium for detecting key points of the cervical spine, in order to improve the accuracy of cervical spine key point detection.

[0004] This application provides a method for detecting key points in the cervical spine, including the following steps: Acquire image data of the cervical spine to be detected; The cervical spine image data is input into a pre-trained deep learning model. The encoder of the deep learning model processes the cervical spine image to obtain a first image feature. The encoder includes a dual-pooling channel attention module, which comprises a first branch structure, a second branch structure, and a fusion structure. The first branch structure extracts global semantic features, the second branch structure extracts local spatial features, and the fusion structure fuses the global semantic features and the local spatial features, and then performs subsequent processing on the fused features to obtain a second image feature. The first image feature is obtained based on the second image feature. The decoder of the deep learning model is used to process the features of the first image to obtain the positions of each key point of the cervical vertebra. According to the cervical spine key point detection method provided in this application, the encoder includes a backbone network and a cervical network, the decoder includes a posture detection head, the backbone network and / or the cervical network includes the dual-pooling channel attention module, the output of the backbone network is the input of the cervical network, and the output of the cervical network includes the first image features.

[0005] According to the cervical spine key point detection method provided in this application, the dual-pooling channel attention module includes a first dual-pooling channel attention module and a second dual-pooling channel attention module; The first dual-pooling channel attention module is located after the last feature extraction module of the backbone network, and the second dual-pooling channel attention module is located after the feature aggregation module of the P4 branch of the neck network.

[0006] According to the cervical spine key point detection method provided in this application, the data processing procedure of the dual-pooling channel attention module includes: The first branch structure is used to extract global features from the input feature data through 1×1 adaptive average pooling to obtain the first global feature; the first global feature is then subjected to a one-dimensional convolution operation to obtain the first recalibrated feature. The second branch structure is used to extract local features from the input feature data through 3×3 adaptive average pooling to obtain the first local feature; the first local feature is then subjected to a one-dimensional convolution operation to obtain the second recalibrated feature. The first recalibrated feature and the second recalibrated feature are fused using the fusion structure, and the fused features are processed using an activation function to obtain a third image feature; the third image feature is multiplied by the input feature data to obtain the output data of the dual-pooling channel attention module.

[0007] According to the cervical spine key point detection method provided in this application, the cervical network includes a dynamic sampling upsampling module, which is used to perform adaptive upsampling operation through a learnable dynamic offset to obtain upsampling features, wherein the first image features are obtained based on the second image features and the upsampling features.

[0008] According to the cervical spine key point detection method provided in this application, the data processing flow of the dynamic sampling upsampling module includes: The dynamic offset of the input feature data is predicted by the sampling point generator of the dynamic sampling upsampling module; the dynamic offset is superimposed on a standard uniform sampling grid, and the feature value is obtained at the offset coordinate position through bilinear sampling to generate the upsampled feature. The sampling point generator adopts a two-level cascaded linear projection structure.

[0009] According to the method for detecting key points of the cervical spine provided in this application, the key points of the cervical spine include the vertebral body corner point, the spinous process tip point, and the vertebral lamina edge point; The training process of the deep learning model includes: Obtain sample cervical spine images and label data of the sample cervical spine images. The label data includes the marked positions of each vertebral body corner point, the marked positions of each spinous process tip point, and the marked positions of each vertebral lamina edge point. The sample cervical spine images are input into a deep learning model to obtain the predicted positions of each vertebral body corner point, each spinous process tip point, and each vertebral lamina edge point. The model loss is calculated based on the vertebral body corner error tolerance parameter, the labeled position and predicted position of each vertebral body corner point, the spinous process tip error tolerance parameter, the labeled position and predicted position of each spinous process tip point, and the lamina edge point error tolerance parameter, the labeled position and predicted position of each lamina edge point. The vertebral body corner error tolerance parameter is less than the spinous process tip error tolerance parameter, and the spinous process tip error tolerance parameter is less than the lamina edge point error tolerance parameter. The deep learning model is tuned based on the model loss.

[0010] According to the cervical spine key point detection method provided in this application, the error tolerance parameter of the vertebral body corner point is 0.05, the error tolerance parameter of the spinous process tip point is 0.07, and the error tolerance parameter of the lamina edge point is 0.10.

[0011] According to the cervical spine key point detection method provided in this application, the vertebral body corner points include four vertebral body corner points of each of the six cervical vertebrae from C2 to C7, the spinous process tip points include the spinous process tip points of each of the six cervical vertebrae from C2 to C7, and the lamina edge points include the lamina edge points of each of the five cervical vertebrae from C3 to C7.

[0012] According to the cervical spine key point detection method provided in this application, the method further includes: generating cervical spine imaging measurement parameters based on the location of each of the cervical spine key points; wherein, the cervical spine imaging measurement parameters include at least one of alignment parameters, morphological parameters, and dynamic parameters; the alignment parameters include at least one of Cobb's angle, Ishihara index, vertical axis distance in the sagittal plane of the cervical spine, and vertebral slippage; the morphological parameters include at least one of vertebral body height-to-depth ratio, intervertebral disc height index, and sagittal diameter of the spinal canal; the dynamic parameters include at least one of segmental mobility, total cervical spine mobility, segmental translation, and interspinous distance.

[0013] This application also provides a cervical spine key point detection device, including the following modules: The image data acquisition module is used to acquire the cervical spine image data to be detected. A key point detection module is used to input the cervical spine image data into a pre-trained deep learning model, and use the encoder of the deep learning model to process the cervical spine image to obtain a first image feature. The encoder includes a dual-pooling channel attention module, which includes a first branch structure, a second branch structure, and a fusion structure. The first branch structure is used to extract global semantic features, the second branch structure is used to extract local spatial features, and the fusion structure is used to fuse the global semantic features and the local spatial features and perform subsequent processing on the fused features to obtain a second image feature. The first image feature is obtained based on the second image feature. The decoder of the deep learning model processes the first image feature to obtain the position of each cervical spine key point.

[0014] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the above-described methods for detecting key points of the cervical spine.

[0015] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the cervical spine key point detection method as described above.

[0016] This application also provides a computer program product, including a computer program that, when executed by a processor, implements any of the above-described methods for detecting key points of the cervical spine.

[0017] This application provides a method, device, electronic device, and storage medium for detecting cervical spine key points. The method includes: acquiring cervical spine image data to be detected; inputting the cervical spine image data into a pre-trained deep learning model, and processing the cervical spine image using the encoder of the deep learning model to obtain first image features; wherein the encoder includes a dual-pooling channel attention module, the dual-pooling channel attention module includes a first branch structure, a second branch structure, and a fusion structure, the first branch structure is used to extract global semantic features, the second branch structure is used to extract local spatial features, and the fusion structure is used to fuse the global semantic features and local spatial features and perform subsequent processing on the fused features to obtain second image features; the first image features are obtained based on the second image features; and the decoder of the deep learning model processes the first image features to obtain the positions of each cervical spine key point. By embedding a dual-pooling channel attention module in the encoder, extracting global semantic features through the first branch structure of the dual-pooling channel attention module, extracting local spatial features through the second branch structure, and fusing the global semantic features and local spatial features through the fusion structure, the accuracy of cervical spine key point detection can be improved by simultaneously considering global semantic features and local spatial features during the prediction process of cervical spine key points. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts of the cervical spine key point detection method provided in this application; Figure 2 This is a schematic diagram of the key points of the cervical spine provided in this application; Figure 3 This is the second flowchart of the cervical spine key point detection method provided in this application; Figure 4 This is a schematic diagram of the backbone network provided in this application; Figure 5 This is a schematic diagram of the neck network structure provided in this application; Figure 6 This is a schematic diagram of the data processing flow of the dual-pooling channel attention module provided in this application; Figure 7 This is a schematic diagram of the data processing flow of the dynamic sampling upsampling module provided in this application; Figure 8This is a flowchart illustrating the deep learning model training process provided in this application; Figure 9 This is a schematic diagram of the cervical spine key point detection device provided by the present invention; Figure 10 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0020] 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.

[0021] Existing methods for detecting cervical spine key points include two approaches: one is to rely on experienced clinicians to identify anatomical markers (key points) one by one on each image; the other is to automatically identify key points using a deep learning model based on the ECA attention mechanism.

[0022] However, the above method has some problems: Problem 1: Manual measurement is inefficient and inconsistent. Traditional methods rely on experienced clinicians to identify anatomical landmarks and perform geometric calculations on each image one by one, taking an average of about 120 seconds per image, and there is significant between-observer and within-observer variability.

[0023] Problem 2: Existing deep learning models lack generalization ability. Most models are only validated on single-center datasets, making it difficult to generalize across different medical institutions, imaging devices, and acquisition schemes. Furthermore, ECA relies solely on global average pooling for channel recalibration, which ignores local structural details for keypoint detection tasks requiring precise spatial localization.

[0024] Question 3: Incomplete anatomical coverage. Existing studies have limited the markers to the anterior column structures of the vertebral body contour or endplate, failing to assess posterior column structures such as the spinous process and lamina, which are crucial for spinal canal assessment and posterior device implantation planning.

[0025] Question 4: Separation of static and dynamic parameters. Existing automation frameworks mostly focus on the static parameters of neutral lateral radiographs or isolated dynamic parameters, while frameworks that integrate both static and dynamic measurements are very rare.

[0026] To address at least one of the aforementioned problems, embodiments of this application provide a method, apparatus, electronic device, and storage medium for detecting key points of the cervical spine. The following is a detailed description... Figures 1 to 10 Please provide a detailed explanation.

[0027] See Figure 1 , Figure 1 This is one of the flowcharts of the cervical spine key point detection method provided in this application, which includes the following steps: S101, acquire the cervical spine image data to be detected.

[0028] The size of the cervical spine image data needs to meet the input requirements of the deep learning model. Therefore, after acquiring the original cervical spine image data, the original cervical spine image data can be normalized to obtain cervical spine image data of a preset size as the cervical spine image data to be detected. The cervical spine image data can be X-ray images, MRI images, or CT images, all of which are within the scope of protection of this application.

[0029] S102, the cervical spine image data is input into a pre-trained deep learning model, and the encoder of the deep learning model is used to process the cervical spine image to obtain a first image feature; wherein, the encoder includes a dual-pooling channel attention module, the dual-pooling channel attention module includes a first branch structure, a second branch structure, and a fusion structure, the first branch structure is used to extract global semantic features, the second branch structure is used to extract local spatial features, and the fusion structure is used to fuse the global semantic features and the local spatial features and perform subsequent processing on the fused features to obtain a second image feature; the first image feature is obtained based on the second image feature.

[0030] The deep learning model here can be any neural network capable of image processing. For example, it could be You Only Look Once (YOLO), a Region-based Convolutional Neural Network (RCNN), or a Faster Region-based Convolutional Neural Network (Faster R-CNN). The deep learning model includes an encoder and a decoder; the encoder is used for feature extraction, and the decoder is used for predicting the location of key points in the cervical spine.

[0031] The encoder embeds a Dual Pooling Channel Attention (DPCA) module. This DPCA module can be located after the feature extraction module of the encoder's downsampling structure, and / or after the feature aggregation module; the specific configuration can be customized based on actual needs. The DPCA module employs a dual-branch structure with a global pooling branch (first branch structure) and a local pooling branch (second branch structure), capturing global semantic features and local spatial features respectively. It then adaptively recalibrates the channel features through element-wise fusion and activation functions.

[0032] S103, the first image features are processed using the decoder of the deep learning model to obtain the positions of each key point of the cervical spine.

[0033] The structure of the decoder for a deep learning model can be found in the decoder structure of image processing neural networks in related technologies, and is not specifically limited in this application.

[0034] In this embodiment, a dual-pooling channel attention module is embedded in the encoder. Global semantic features are extracted through the first branch structure of the dual-pooling channel attention module, and local spatial features are extracted through the second branch structure. The global semantic features and local spatial features are fused through the fusion structure, so that both global semantic features and local spatial features are considered in the prediction process of cervical key points, which can improve the accuracy of cervical key point detection.

[0035] Existing cervical spine key points only cover the vertebral body outline (anterior column) and cannot assess posterior column structures such as the spinous process and lamina. To address this issue, the cervical spine key points in this application embodiment may include vertebral body corner points, spinous process tip points, and lamina edge points. In one possible implementation, see [link to relevant documentation]. Figure 2 The vertebral angle points include four vertebral angle points (anterior superior angle, anterior inferior angle, posterosuperior angle, and posteroinferior angle) for each of the six cervical vertebrae from C2 to C7, totaling 24 vertebral angle points, corresponding to... Figure 2 The red dots in the image; the spinous process tips include the spinous process tips of each of the six cervical vertebrae from C2 to C7, for a total of 6 spinous process tips, corresponding to... Figure 2 The yellow dots in the image; the lamina edge points include the lamina edge points of each of the five cervical vertebrae from C3 to C7, for a total of 5 lamina edge points, corresponding to... Figure 2 The blue dot in the diagram. Here, R represents the right view.

[0036] In this embodiment, six spinous process tip points and five lamina edge points are added to the vertebral body corner points, forming a system of 35 key cervical spine points. Based on the coordinates of the key cervical spine points, 11 cervical spine imaging measurement indicators, including alignment parameters (4 items), morphological parameters (3 items), and dynamic parameters (4 items), can be automatically calculated. Among them, 8 parameters achieved good to excellent reliability under experimental conditions, with a processing speed of 67.5 ms / image, which is about 1788 times faster than manual calculation.

[0037] exist Figure 1 or Figure 2 Based on the illustrated embodiments, in one possible implementation, see [link to relevant documentation]. Figure 3 The method shown also includes: S104, cervical spine imaging measurement parameters are generated based on the location of each of the key cervical spine points; wherein, the cervical spine imaging measurement parameters include at least one of alignment parameters, morphological parameters, and dynamic parameters.

[0038] Based on the location of each key point in the cervical spine, and according to the preset geometric definition of cervical spine imaging parameters, the system automatically calculates the cervical spine imaging measurement parameters and outputs the measurement results.

[0039] The alignment parameters and morphological parameters can be obtained from key points of the cervical spine in a single cervical spine image. The alignment parameters include at least one of the following: Cobb angle, Ishihara index, vertical axis distance in the sagittal plane of the cervical spine, and vertebral slippage. The morphological parameters include at least one of the following: vertebral body height-to-depth ratio, intervertebral disc height index, and sagittal diameter of the spinal canal.

[0040] The dynamic parameters are obtained from key cervical spine points in at least two cervical spine images acquired sequentially from the same patient (e.g., a pair of paired hyperflexion and hyperextension images). The dynamic parameters include at least one of segmental range of motion, total cervical spine range of motion, segmental translation, and interspinous distance.

[0041] The Cobb angle represents the angle between the inferior endplates of the C2 and C7 vertebral bodies. In one example, the Cobb angle can be calculated using the following formula: θ = angle(v1, v2), where v1 = C2_PI - C2_AI; v2 = C7_PI - C7_AI; angle(v1, v2) calculates the angle between vectors v1 and v2, θ is the Cobb angle, C2 represents the C2 vertebra (second cervical vertebra), C7 represents the C7 vertebra (seventh cervical vertebra), PI represents the posteroinferior angle, and AI represents the anterior inferior angle. Combining vertebral bodies with cervical key points allows for the representation of specific vertebral key points; for example, C2_PI represents the posteroinferior angle of the C2 vertebra.

[0042] The Ishihara Index represents the cervical curvature index based on the vertical distance from the line connecting C3-C6 to C2-C7. In one example, the Ishihara Index can be calculated using the following formula: CCI=(a3+a4+a5+a6) / D×100%, where CCI is the Ishihara Index; D=dist(C2_PI, C7_PI), representing the distance between the line segments from C2_PI to C7_PI; a3 represents the distance from C3_PI to the baseline (the line containing C2_PI and C7_PI); a4 represents the distance from C4_PI to the baseline; a5 represents the distance from C5_PI to the baseline; and a6 represents the distance from C6_PI to the baseline.

[0043] The sagittal vertical axis of the cervical spine represents the horizontal distance from the plumb line of the C2 centroid to the posterosuperior angle of the C7 vertebra. In one example, the sagittal vertical axis of the cervical spine can be calculated by the following formula: cSVA = x(C2_centroid) - x(C7_PS), where cSVA is the sagittal vertical axis of the cervical spine, x(C2_centroid) represents the x-coordinate of the aggregation center of the C2 vertebral body, x(C7_PS) represents the x-coordinate of the posterosuperior angle of the C7 vertebral body, and PS represents the posterosuperior angle.

[0044] Vertebral slip represents the horizontal displacement between the posterior edges of adjacent vertebrae. In one example, vertebral slip can be calculated using the following formula: Slip = x(Cn_PI) - x(Cn+1_PS), where Slip represents vertebral slip, x(Cn_PI) represents the x-coordinate of the posteroinferior angle of vertebra Cn, and x(Cn+1_PS) represents the x-coordinate of the posterosuperior angle of vertebra Cn+1.

[0045] The vertebral body height-to-depth ratio represents the ratio of the average height of the vertebral body to its anteroposterior diameter. In one example, the vertebral body height-to-depth ratio can be calculated using the following formula: H / D = (Ha + Hm + Hp) / 3 / VBd, where H / D represents the vertebral body height-to-depth ratio; Ha represents the anterior height, Ha = dist(Cn_AS, Cn_AI), representing the distance from the anterosuperior angle to the anteroinferior angle of the Cn vertebral body, where AS represents the anterosuperior angle; Hm represents the central height, Hm = dist(Cn_MS, Cn_MI), representing the distance from the midpoint of the upper-middle edge to the midpoint of the lower-middle edge of the Cn vertebral body, where MS represents the midpoint of the upper-middle edge, and MI represents the midpoint of the lower-middle edge; Hp represents the posterior height, Hp = dist(Cn_PS, Cn_PI), representing the distance from the posterosuperior angle to the posteroinferior angle of the Cn vertebral body; VBd represents the vertebral body length, VBd = dist(Cn_AI, Cn_PI) represents the distance from the anteroinferior angle of the Cn vertebral body to the posteroinferior angle of the Cn vertebral body.

[0046] The intervertebral disc height index (DHI) represents the ratio of the average intervertebral disc height to the anteroposterior diameter of the superior vertebral body. In one example, the DHI can be calculated using the following formula: DHI = (Da + Dp) / 2 / VBd, where DHI represents the intervertebral disc height index; Da represents the anterior edge height of the intervertebral disc, Da = dist(Cn_AI, Cn+1_AS), which represents the distance from the anteroinferior angle of vertebral body Cn to the anterosuperior angle of vertebral body Cn+1; and Dp represents the posterior edge height of the intervertebral disc, Dp = dist(Cn_PI, Cn+1_PS), which represents the distance from the posteroinferior angle of vertebral body Cn to the posterosuperior angle of vertebral body Cn+1.

[0047] The sagittal diameter of the spinal canal represents the anterior-posterior distance from the posterior surface of the vertebral body to the spinous process lamina line. In one example, the sagittal diameter of the spinal canal can be calculated using the following formula: SCD=dist(Cn_PM,Cn_LP), where SCD represents the sagittal diameter of the spinal canal, dist(Cn_PM,Cn_LP) represents the distance from the midpoint of the posterior margin of the Cn vertebral body to the anterior point of the vertebral arch of the Cn vertebral body, PM represents the midpoint of the posterior margin, and LP represents the anterior margin of the lamina.

[0048] Segmental range of motion represents the angular change of the segment between flexion and extension. In one example, segmental range of motion can be calculated by the following formula: ROM1=|θf - θe|, where ROM1 represents segmental range of motion, θf represents the segmental angle in flexion, and θe represents the segmental angle in hyperextension.

[0049] Total cervical range of motion (ROM2) represents the change in the C2-C7 Cobb angle between maximum flexion and hyperextension. In one example, ROM2 can be calculated using the following formula: ROM2 = |αf - αe|, where ROM2 represents the total cervical range of motion, αf represents the Cobb angle in flexion, and αe represents the Cobb angle in hyperextension.

[0050] Segmental translation represents the change in horizontal displacement between adjacent vertebrae during flexion and extension. In one example, segmental translation can be calculated using the following formula: T = |df - de|, where T represents segmental translation, df represents loudness displacement in flexion, df = x(Cn_PI) - x(Cn+1_PS), x(Cn_PI) represents the x-coordinate of the posteroinferior angle of vertebra Cn, x(Cn+1_PS) represents the x-coordinate of the posterosuperior angle of vertebra Cn+1, and de represents loudness displacement in hyperextension.

[0051] The spinous process spacing represents the linear distance between the tips of adjacent spinous processes. In one example, the spinous process spacing can be calculated by the following formula: ISD=dist(SPn, SPn+1), where ISD represents the spinous process spacing, SPn represents the tip of the spinous process of vertebra Cn, SPn+1 represents the tip of the spinous process of vertebra Cn+1, and dist(SPn, SPn+1) represents the distance between SPn and SPn+1.

[0052] In this embodiment, after obtaining the location of each key point of the cervical spine, the cervical spine imaging measurement parameters can be automatically calculated, providing key indicators for subsequent assessment of the biomechanical state of the spine; no manual calculation is required, which can reduce the workload of manual calculation and improve the calculation efficiency of cervical spine imaging measurement parameters.

[0053] exist Figure 1 or Figure 3 Based on the illustrated embodiment, in one possible implementation, taking the YOLO deep learning model as an example, the encoder includes a backbone network and a neck network, the decoder includes a pose head, the backbone network and / or the neck network includes the dual-pooling channel attention module, the output of the backbone network is the input of the neck network, and the output of the neck network includes the first image features. In one example, the dual-pooling channel attention module includes a first dual-pooling channel attention module and a second dual-pooling channel attention module; the first dual-pooling channel attention module is located after the last feature extraction module of the backbone network, and the second dual-pooling channel attention module is located after the feature aggregation module of the P4 branch of the neck network.

[0054] YOLO's backbone network typically includes five feature extraction modules, see [link to YOLO backbone network]. Figure 4 In this case, the last feature extraction module is the fifth feature extraction module. The output of the last feature extraction module of the backbone network is used as the input of the first dual-pooling channel attention module, and the output of the first dual-pooling channel attention module is used as Backbone-P5. YOLO's neck network can employ a bidirectional feature fusion structure of Feature Pyramid Network (FPN) + Path Aggregation Network (PANet). See [link to relevant documentation] Figure 5 The output of the feature aggregation module of the P4 branch of the neck network is used as the input of the second dual-pooling channel attention module, and the output of the second dual-pooling channel attention module is used as Neck-P4. It is understood that... Figure 4 and Figure 5 The network structure shown is for illustrative purposes only. Except for the dual-pooling channel attention module, the specific structure of the backbone network and neck network can be found in the network structure of the relevant YOLO model.

[0055] In the backbone network and neck network, except for the dual-pooling channel attention module, the other structures can follow the corresponding structures in YOLO. The structure of the pose detection head can also follow the corresponding structures in YOLO.

[0056] In this embodiment, the first dual-pooling channel attention module is embedded at position P5 after the last feature extraction module of the backbone network, and the second dual-pooling channel attention module is embedded after the feature aggregation module of the P4 branch of the neck network, forming a dual-position attention enhancement configuration. Its optimal performance can reach an average precision (mAP)@50:95=90.05%, and it only adds a small amount of computational overhead. While greatly improving the prediction accuracy, it basically does not increase the amount of computation.

[0057] The data processing procedures for other structures in the encoder besides the dual-pooling channel attention module can be found in related technologies, and are not specifically limited in this application. The following describes the data processing procedure of the dual-pooling channel attention module. Figure 1 or Figure 3 Based on the illustrated embodiments, in one possible implementation, see [link to relevant documentation]. Figure 6 The data processing procedure of the dual-pooling channel attention module includes: S601, using the first branch structure, global feature extraction is performed on the input feature data through 1×1 adaptive average pooling to obtain the first global feature; the first global feature is then subjected to a one-dimensional convolution operation to obtain the first recalibrated feature.

[0058] The first branch structure may include a 1×1 adaptive average pooling structure and a one-dimensional convolution structure. The 1×1 adaptive average pooling structure is used to perform 1×1 adaptive average pooling on the feature data input by the dual-pooling channel attention module to extract the global context and obtain the first global feature. Then, the one-dimensional convolution structure of the first branch structure is used to perform one-dimensional convolution channel recalibration on the first global feature to obtain the first recalibrated feature.

[0059] S602, using the second branch structure, local features are extracted from the input feature data through 3×3 adaptive average pooling to obtain the first local features; the first local features are then subjected to a one-dimensional convolution operation to obtain the second recalibrated features.

[0060] The second branch structure can include a 3×3 adaptive average pooling structure and a one-dimensional convolution structure. The 3×3 adaptive average pooling structure is used to perform 3×3 adaptive average pooling on the feature data input to the dual-pooling channel attention module to retain fine-grained spatial structure and obtain the first local feature. Then, the one-dimensional convolution structure of the second branch structure is used to perform one-dimensional convolution channel recalibration on the first local feature to obtain the second recalibrated feature.

[0061] S603, the first recalibrated feature and the second recalibrated feature are fused using the fusion structure, and the fused features are processed using an activation function to obtain a third image feature; the third image feature is multiplied by the input feature data to obtain a second image feature.

[0062] The first and second labeling features are added together at the same positions to obtain a fused feature. This fused feature is then processed using an activation function to obtain a third image feature. The activation function here can be a Sigmoid activation function or a Tanh activation function, etc. The third image feature is then multiplied by the same positions of the feature data input to the dual-pooling channel attention module to obtain the second image feature, which is the output data of the dual-pooling channel attention module. This second image feature is then output to subsequent modules for further processing.

[0063] The above process can be expressed by the following calculation formula: y=σ(Conv1d(GAP(x))+Conv1d(LAP(x))) x; Where y is the output data of the dual-pooling channel attention module; x is the feature data input to the dual-pooling channel attention module; GAP() represents global average pooling, i.e., 1×1 adaptive average pooling; LAP() represents local average pooling, i.e., 3×3 adaptive average pooling; Conv1d() represents one-dimensional convolution; σ() represents the activation function. This indicates that elements at the same position are multiplied.

[0064] In this embodiment, the data processing procedure of the dual-pooling channel attention module is given. By fusing global semantic features and local spatial features, the accuracy of cervical key point detection can be improved by simultaneously considering global semantic features and local spatial features in the prediction process of cervical key points.

[0065] The upsampling module in the feature pyramid network structure of the YOLO neck network uses standard bilinear interpolation with a fixed kernel function, which cannot adapt to feature variations in different anatomical regions. To improve the adaptability of the neck network, in one example, a dynamic upsampling module can be used instead of the traditional upsampling module. The neck network may include a dynamic upsampling module, which is used to perform adaptive upsampling operations using learnable dynamic offsets to obtain upsampled features, where the first image feature is obtained based on the second image feature and the upsampled feature. Figure 1 or Figure 3 Based on the illustrated embodiments, in one possible implementation, see [link to relevant documentation]. Figure 7 The data processing flow of the dynamic sampling upsampling module includes: S701, the sampling point generator of the dynamic sampling upsampling module is used to predict the dynamic offset of the input feature data, wherein the sampling point generator adopts a two-level cascaded linear projection structure.

[0066] In one example, the dynamic sampling upsampling module can specifically be a DySample module. For the input feature map F∈R... (C ×H×W) The sampling point generator employs a variant of Linear Projection (LP) to generate a dynamic offset ΔG∈R through two levels of cascaded linear projection. (2×sH×sW) In one example, the dynamic offset can also be normalized through range constraints, such as using pixel shuffle and the sigmoid function to normalize the offset and control the maximum displacement range of the sampling points. The dimension is transformed from (C,H,W) to (2,sH,sW) to achieve upsampling of spatial resolution.

[0067] S702, the dynamic offset is superimposed on the standard uniform sampling grid, and the feature value is obtained at the offset coordinate position through bilinear sampling to generate the upsampled feature.

[0068] The dynamic offset is superimposed on the standard uniform sampling grid G0, and feature values ​​are obtained at the offset coordinate positions through bilinear sampling to generate upsampled feature data: F'∈R (C×sH×sW) .

[0069] In this embodiment, the dynamic sampling upsampling module is integrated into the upsampling operation of the cervical network, and the upsampling is reconstructed into a dynamic point sampling problem. Experiments show that the dynamic sampling upsampling module contributes the largest single improvement (+0.88mAP@50:95), indicating that the use of the dynamic sampling upsampling module improves the accuracy of cervical key point localization and detection.

[0070] The following describes the training process of deep learning models. (See also...) Figure 8 , Figure 8 This is a schematic diagram of a deep learning model training method in an embodiment of this application, including: S801, acquire sample cervical spine images and label data of the sample cervical spine images, wherein the label data includes the labeled positions of each vertebral body corner point, the labeled positions of each spinous process tip point, and the labeled positions of each vertebral lamina edge point.

[0071] The size of the sample cervical spine images needs to meet the input requirements of the deep learning model. The sample cervical spine images can be X-ray images, MRI images, or CT images, all of which are within the scope of this application. The label data of the sample cervical spine images includes the location of each key point of the cervical spine in the sample cervical spine images. Specifically, it can include the labeled locations of each vertebral body corner point, each spinous process tip point, and each lamina edge point in the sample cervical spine images. The label data of the sample cervical spine images can be obtained through manual labeling.

[0072] S802, the sample cervical spine image is input into the deep learning model to obtain the predicted position of each vertebral body corner point, the predicted position of each spinous process tip point, and the predicted position of each vertebral lamina edge point.

[0073] The deep learning model includes an encoder and a decoder. The encoder of the deep learning model includes a dual-pooling channel attention module. Sample cervical spine images are input into the deep learning model to obtain the prediction results. The prediction results include the predicted positions of each vertebral angle point, the predicted position of each spinous process tip point, and the predicted position of each lamina edge point in the sample cervical spine image.

[0074] S803, calculate the model loss based on the vertebral body corner error tolerance parameter, the labeled position and predicted position of each vertebral body corner point, the spinous process tip error tolerance parameter, the labeled position and predicted position of each spinous process tip point, and the lamina edge point error tolerance parameter, the labeled position and predicted position of each lamina edge point, wherein the vertebral body corner error tolerance parameter is less than the spinous process tip point error tolerance parameter, and the spinous process tip point error tolerance parameter is less than the lamina edge point error tolerance parameter.

[0075] The key points of the cervical spine are divided into three morphological categories based on their anatomical morphological characteristics: the first category is the vertebral body corner point, which has a clear cortical bone boundary and is assigned a small error tolerance parameter, namely the vertebral body corner point error tolerance parameter σ1; the second category is the spinous process tip point, which has a clear but occasionally obscured boundary and is assigned a medium error tolerance parameter, namely the spinous process tip point error tolerance parameter σ2; the third category is the lamina edge point, which has an inherent boundary ambiguity and is assigned a large error tolerance parameter, namely the lamina edge point error tolerance parameter σ3; where σ1 < σ2 < σ3.

[0076] The error tolerance parameters can be set according to the actual situation, but it is necessary to ensure that σ1 < σ2 < σ3. In one example, the error tolerance parameter for the vertebral body corner can be in the range of 0.04-0.06, the error tolerance parameter for the spinous process tip can be in the range of 0.06-0.08, and the error tolerance parameter for the lamina edge can be in the range of 0.09-0.11.

[0077] In one example, the model loss can be calculated using the following formula: L=1-Σ i exp(-d i 2 / (2s 2 σ i 2 ))×δ(v i >0) / Σ i δ(v i >0); Where di is the distance between the labeled location and the predicted location of the i-th cervical vertebra key point; s is the target scale factor; σ i Let δ(v) be the error tolerance parameter for the i-th cervical vertebra key point; i >0) is a visibility mask, indicating whether the i-th cervical vertebra key point is visible. When the i-th cervical vertebra key point is visible, δ(v) = 0. i >0)=1, otherwise δ(v) i >0)=0.

[0078] The model loss can also be calculated using a weighted approach. This involves calculating the losses at the vertebral body corners, the spinous process tip, and the lamina edge separately, and then weighting these three losses together to obtain the model loss. A larger error tolerance parameter corresponds to a smaller weight value. In one example, the model loss can be calculated using the following formula: ; Where L is the model loss and a is the number of cone corner points. This is the labeled position of the i-th vertebral angle point. Let be the predicted position of the i-th vertebral angle; b is the number of spinous process tip points. This is the labeled position of the tip of the j-th spinous process. Let be the predicted location of the tip of the j-th spinous process; c is the number of points on the edge of the lamina. This is the location of the m-th lamina edge point. Let σ1 be the predicted position of the m-th lamina edge point, where σ1 < σ2 < σ3. It is understood that the formula here is merely illustrative, and any equivalent substitutions made within the spirit and scope of the technical solution of this application are within the protection scope of this application.

[0079] S804, The deep learning model is tuned according to the model loss.

[0080] For methods of tuning the parameters of deep learning models, please refer to relevant techniques. In one example, backpropagation can be used to adjust the parameters of a deep learning model. New sample cervical spine images are selected to continue training the deep learning model until the model loss converges or the preset number of training iterations is reached, thus obtaining the trained deep learning model.

[0081] In this embodiment, a dual-pooling channel attention module is embedded in the encoder. Global semantic features are extracted through the first branch structure of the dual-pooling channel attention module, and local spatial features are extracted through the second branch structure. The global semantic features and local spatial features are fused through the fusion structure, so that both global semantic features and local spatial features are considered in the prediction process of cervical key points, which can improve the accuracy of cervical key point detection.

[0082] exist Figure 8 Based on the illustrated embodiment, in one possible implementation, the error tolerance parameter for the vertebral body corner point is 0.05, the error tolerance parameter for the spinous process tip point is 0.07, and the error tolerance parameter for the lamina edge point is 0.10. The magnitude of the error tolerance parameter reflects the inherent variability of the annotation—stricter for clear boundaries and more lenient for blurred boundaries. When the error tolerance parameter for the vertebral body corner point is 0.05, the error tolerance parameter for the spinous process tip point is 0.07, and the error tolerance parameter for the lamina edge point is 0.10, taking the YOLO deep learning model as an example, its individual contribution is +0.09 mAP@50:95. After collaboration with the dual-pooling channel attention module and the dynamic sampling upsampling module, the excess gain produced is +1.11 mAP@50:95, which is greater than the sum of their individual gains, indicating a positive synergy between cervical key point classification supervision and spatial feature enhancement.

[0083] The following example uses the YOLO deep learning model to illustrate the model training process. The YOLO26s-Pose model is selected as the main architecture. A dual-pooling channel attention module is embedded after the last feature extraction module of the YOLO26s-Pose backbone network, and another dual-pooling channel attention module is embedded after the feature aggregation module of the P4 branch of the neck network. Simultaneously, a dynamic sampling upsampling module replaces the standard bilinear interpolation upsampling module in the neck network. Furthermore, the input to the deep learning model is configured as a set of (x, y) coordinates of 35 cervical spine key points. The input to the deep learning model is a 1280×1280 pixel sample image. The key parameters of the deep learning model are shown in Table 1. Table 1 The training platform consisted of a server equipped with four NVIDIA A800 GPUs (each with 80GB of VRAM) and used Compute Unified Device Architecture (CUDA) 13.0. The training cycle consisted of 600 epochs with a batch size of 16. The AdamW optimizer was used, with an initial learning rate of 5×10⁻⁶. -4 The weight is reduced to 0.5% through cosine annealing, with a weight reduction of 5 × 10⁻⁶. -4 Data augmentation included ±5° rotation and ±30% scale variation, with horizontal flipping disabled to maintain anatomical lateral consistency. The model checkpoint with the lowest validation loss was selected as the final evaluation model.

[0084] The sample data came from three clinical centers and one public dataset, totaling 3310 cervical spine X-ray images. The public dataset was VinDr-SpineXR. Lateral cervical spine X-ray images (including neutral, hyperflexion, and hyperextension views) in Digital Imaging and Communications in Medicine (DICOM) format were de-identified and then scaled to 1280×1280 pixels as sample cervical spine images for the deep learning model. Data were randomly allocated as a 70% training set, 10% validation set, and 20% test set, using stratified sampling to maintain a balanced distribution of image type and data source.

[0085] A deep learning model (an improved YOLO26s-Pose model) was trained using training and validation sets, and then tested using a test set. Test results showed that the deep learning model achieved an mAP@50:95 of 90.05%, an improvement of +1.11 percentage points compared to the conventional YOLO26s-Pose (88.94%). In cross-dataset evaluation, the mAP@50 of the deep learning model trained in this application remained above 98.7% across all configurations. Using 35 cervical spine key points predicted by the deep learning model trained in this application, 11 cervical spine imaging measurement parameters were automatically calculated according to predefined geometric formulas. Among these, 3 parameters achieved excellent reliability (ICC ≥ 0.90), 5 achieved good reliability (ICC ≥ 0.776), and 1 achieved moderate reliability, significantly improving the reliability of cervical spine imaging measurement parameters. ICC stands for Intraclass Correlation Coefficient.

[0086] The cervical spine key point detection device provided in this application is described below. The cervical spine key point detection device described below can be referred to in correspondence with the cervical spine key point detection method described above.

[0087] See Figure 9 This application provides a cervical spine key point detection device, including: Image data acquisition module 901 is used to acquire cervical spine image data to be detected; The key point detection module 902 is used to input the cervical spine image data into a pre-trained deep learning model, and use the encoder of the deep learning model to process the cervical spine image to obtain a first image feature. The encoder includes a dual-pooling channel attention module, which includes a first branch structure, a second branch structure, and a fusion structure. The first branch structure is used to extract global semantic features, the second branch structure is used to extract local spatial features, and the fusion structure is used to fuse the global semantic features and the local spatial features and perform subsequent processing on the fused features to obtain a second image feature. The first image feature is obtained based on the second image feature. The decoder of the deep learning model processes the first image feature to obtain the position of each cervical spine key point.

[0088] In one possible implementation, the encoder includes a backbone network and a neck network, the decoder includes a pose detection head, the backbone network and / or the neck network includes the dual-pooling channel attention module, the output of the backbone network is the input of the neck network, and the output of the neck network includes the first image features.

[0089] In one possible implementation, the dual-pooling channel attention module includes a first dual-pooling channel attention module and a second dual-pooling channel attention module; the first dual-pooling channel attention module is located after the last feature extraction module of the backbone network, and the second dual-pooling channel attention module is located after the feature aggregation module of the P4 branch of the neck network.

[0090] In one possible implementation, the keypoint detection module 902 is specifically used for: extracting global features from the input feature data using 1×1 adaptive average pooling through the first branch structure to obtain a first global feature; performing a one-dimensional convolution operation on the first global feature to obtain a first recalibrated feature; extracting local features from the input feature data using 3×3 adaptive average pooling through the second branch structure to obtain a first local feature; performing a one-dimensional convolution operation on the first local feature to obtain a second recalibrated feature; fusing the first recalibrated feature and the second recalibrated feature using the fusion structure, and processing the fused feature using an activation function to obtain a third image feature; and multiplying the third image feature with the input feature data to obtain the output data of the dual-pooling channel attention module.

[0091] In one possible implementation, the neck network includes a dynamic sampling upsampling module for obtaining upsampled features through adaptive upsampling operations using learnable dynamic offsets, wherein the first image features are obtained based on the second image features and the upsampled features.

[0092] In one possible implementation, the key point detection module 902 is specifically used to: predict the dynamic offset of the input feature data using the sampling point generator of the dynamic sampling upsampling module; superimpose the dynamic offset onto a standard uniform sampling grid, obtain feature values ​​at the offset coordinate positions through bilinear sampling, and generate upsampled features, wherein the sampling point generator adopts a two-level cascaded linear projection structure.

[0093] In one possible implementation, the key points of the cervical spine include vertebral body corner points, spinous process tip points, and lamina edge points; the device further includes: a model training module, used to acquire sample cervical spine images and label data of the sample cervical spine images, the label data including the labeled positions of each vertebral body corner point, each spinous process tip point, and each lamina edge point; inputting the sample cervical spine images into a deep learning model to obtain the predicted positions of each vertebral body corner point, each spinous process tip point, and each lamina edge point; calculating the model loss based on the vertebral body corner point error tolerance parameter, the labeled and predicted positions of each vertebral body corner point, the spinous process tip point error tolerance parameter, the labeled and predicted positions of each spinous process tip point, and the lamina edge point error tolerance parameter, the labeled and predicted positions of each lamina edge point, wherein the vertebral body corner point error tolerance parameter is less than the spinous process tip point error tolerance parameter, and the spinous process tip point error tolerance parameter is less than the lamina edge point error tolerance parameter; and tuning the deep learning model based on the model loss.

[0094] In one possible implementation, the vertebral body corner error tolerance parameter is 0.05, the spinous process tip error tolerance parameter is 0.07, and the lamina edge error tolerance parameter is 0.10.

[0095] In one possible implementation, the vertebral body corner points include four vertebral body corner points for each of the six cervical vertebrae from C2 to C7, the spinous process tip points include the spinous process tip points for each of the six cervical vertebrae from C2 to C7, and the lamina edge points include the lamina edge points for each of the five cervical vertebrae from C3 to C7.

[0096] In one possible implementation, the device further includes: a parametric geometry calculation module, used to generate cervical spine imaging measurement parameters based on the positions of each of the key cervical vertebrae; wherein the cervical spine imaging measurement parameters include at least one of alignment parameters, morphological parameters, and dynamic parameters; the alignment parameters include at least one of Cobb's angle, Ishihara index, vertical axis distance in the sagittal plane of the cervical vertebrae, and vertebral slippage; the morphological parameters include at least one of vertebral body height-to-depth ratio, intervertebral disc height index, and sagittal diameter of the spinal canal; the dynamic parameters include at least one of segmental mobility, total cervical vertebral mobility, segmental translation, and interspinous distance.

[0097] Figure 10 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 10 As shown, the electronic device may include: a processor 1010, a communications interface 1020, a memory 1030, and a communications bus 1040, wherein the processor 1010, the communications interface 1020, and the memory 1030 communicate with each other through the communications bus 1040. The processor 1010 can call logical instructions in the memory 1030 to execute any of the cervical spine key point detection methods described in this application. The method includes: acquiring cervical spine image data to be detected; inputting the cervical spine image data into a pre-trained deep learning model, and processing the cervical spine image using the encoder of the deep learning model to obtain a first image feature; wherein the encoder includes a dual-pooling channel attention module, the dual-pooling channel attention module includes a first branch structure, a second branch structure, and a fusion structure, the first branch structure is used to extract global semantic features, the second branch structure is used to extract local spatial features, and the fusion structure is used to fuse the global semantic features and the local spatial features and perform subsequent processing on the fused features to obtain a second image feature; the first image feature is obtained based on the second image feature; and the decoder of the deep learning model is used to process the first image feature to obtain the position of each cervical spine key point.

[0098] Furthermore, the logical instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the cervical spine key point detection method provided by the above methods. The method includes: acquiring cervical spine image data to be detected; inputting the cervical spine image data into a pre-trained deep learning model, and processing the cervical spine image using the encoder of the deep learning model to obtain a first image feature; wherein the encoder includes a dual-pooling channel attention module, the dual-pooling channel attention module includes a first branch structure, a second branch structure, and a fusion structure, the first branch structure is used to extract global semantic features, the second branch structure is used to extract local spatial features, and the fusion structure is used to fuse the global semantic features and the local spatial features and perform subsequent processing on the fused features to obtain a second image feature; the first image feature is obtained based on the second image feature; and the first image feature is processed using the decoder of the deep learning model to obtain the position of each cervical spine key point.

[0100] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program performs the cervical spine key point detection method provided by the above methods. The method includes: acquiring cervical spine image data to be detected; inputting the cervical spine image data into a pre-trained deep learning model, and processing the cervical spine image using the encoder of the deep learning model to obtain a first image feature; wherein the encoder includes a dual-pooling channel attention module, the dual-pooling channel attention module includes a first branch structure, a second branch structure, and a fusion structure, the first branch structure is used to extract global semantic features, the second branch structure is used to extract local spatial features, and the fusion structure is used to fuse the global semantic features and the local spatial features and perform subsequent processing on the fused features to obtain a second image feature; the first image feature is obtained based on the second image feature; and the first image feature is processed using the decoder of the deep learning model to obtain the position of each cervical spine key point.

[0101] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0102] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting key points of the cervical spine, characterized in that, include: Acquire image data of the cervical spine to be detected; The cervical spine image data is input into a pre-trained deep learning model. The encoder of the deep learning model processes the cervical spine image to obtain a first image feature. The encoder includes a dual-pooling channel attention module, which comprises a first branch structure, a second branch structure, and a fusion structure. The first branch structure extracts global semantic features, the second branch structure extracts local spatial features, and the fusion structure fuses the global semantic features and the local spatial features, and then performs subsequent processing on the fused features to obtain a second image feature. The first image feature is obtained based on the second image feature. The first image features are processed using the decoder of the deep learning model to obtain the positions of each key point of the cervical vertebra.

2. The method according to claim 1, characterized in that, The encoder includes a backbone network and a neck network, the decoder includes a pose detection head, the backbone network and / or the neck network includes the dual-pooling channel attention module, the output of the backbone network is the input of the neck network, and the output of the neck network includes the first image features.

3. The method according to claim 2, characterized in that, The dual-pooling channel attention module includes a first dual-pooling channel attention module and a second dual-pooling channel attention module; The first dual-pooling channel attention module is located after the last feature extraction module of the backbone network, and the second dual-pooling channel attention module is located after the feature aggregation module of the P4 branch of the neck network.

4. The method according to claim 1, characterized in that, The data processing procedure of the dual-pooling channel attention module includes: The first branch structure is used to extract global features from the input feature data through 1×1 adaptive average pooling to obtain the first global feature; the first global feature is then subjected to a one-dimensional convolution operation to obtain the first recalibrated feature. The second branch structure is used to extract local features from the input feature data through 3×3 adaptive average pooling to obtain the first local feature; the first local feature is then subjected to a one-dimensional convolution operation to obtain the second recalibrated feature. The first recalibrated feature and the second recalibrated feature are fused using the fusion structure, and the fused features are processed using an activation function to obtain a third image feature; the third image feature is multiplied by the input feature data to obtain a second image feature.

5. The method according to claim 2, characterized in that, The neck network includes a dynamic sampling upsampling module, which is used to perform adaptive upsampling operations through learnable dynamic offsets to obtain upsampling features. The first image feature is obtained based on the second image feature and the upsampling feature.

6. The method according to claim 5, characterized in that, The data processing flow of the dynamic sampling upsampling module includes: The dynamic offset of the input feature data is predicted by the sampling point generator of the dynamic sampling upsampling module; the dynamic offset is superimposed on a standard uniform sampling grid, and the feature value is obtained at the offset coordinate position through bilinear sampling to generate the upsampled feature. The sampling point generator adopts a two-level cascaded linear projection structure.

7. The method according to claim 1, characterized in that, The key points of the cervical spine include the vertebral body angle, the tip of the spinous process, and the edge of the lamina. The training process of the deep learning model includes: Obtain sample cervical spine images and label data of the sample cervical spine images. The label data includes the marked positions of each vertebral body corner point, the marked positions of each spinous process tip point, and the marked positions of each vertebral lamina edge point. The sample cervical spine images are input into a deep learning model to obtain the predicted positions of each vertebral body corner point, each spinous process tip point, and each vertebral lamina edge point. The model loss is calculated based on the vertebral body corner error tolerance parameter, the labeled position and predicted position of each vertebral body corner point, the spinous process tip error tolerance parameter, the labeled position and predicted position of each spinous process tip point, and the lamina edge point error tolerance parameter, the labeled position and predicted position of each lamina edge point. The vertebral body corner error tolerance parameter is less than the spinous process tip error tolerance parameter, and the spinous process tip error tolerance parameter is less than the lamina edge point error tolerance parameter. The deep learning model is tuned based on the model loss.

8. A cervical spine key point detection device, characterized in that, include: The image data acquisition module is used to acquire the cervical spine image data to be detected. A key point detection module is used to input the cervical spine image data into a pre-trained deep learning model, and use the encoder of the deep learning model to process the cervical spine image to obtain a first image feature. The encoder includes a dual-pooling channel attention module, which includes a first branch structure, a second branch structure, and a fusion structure. The first branch structure is used to extract global semantic features, the second branch structure is used to extract local spatial features, and the fusion structure is used to fuse the global semantic features and the local spatial features and perform subsequent processing on the fused features to obtain a second image feature. The first image feature is obtained based on the second image feature. The decoder of the deep learning model processes the first image feature to obtain the position of each cervical spine key point.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the cervical spine key point detection method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the cervical spine key point detection method as described in any one of claims 1 to 7.

Images