Acetabulum index key point detection method fusing edge features and detail sensing network

Abstract

The invention provides an acetabulum index key point detection method fusing edge features and a detail perception network, and the method comprises the steps: obtaining a to-be-detected image which comprises an X-ray normal position image of a hip joint; based on a YOLOv8s algorithm, in combination with edge feature enhancement and a detail perception network, constructing a hip joint acetabulum index key point detection model, and training to obtain a target hip joint acetabulum index key point detection model; and inputting the to-be-detected image into the target hip joint acetabulum index key point detection model, and outputting through the target hip joint acetabulum index key point detection model to obtain a corresponding hip joint acetabulum index key point. According to the invention, the method can achieve the recognition of the edge region features of the key points and the perception of the fine structure features, and improves the detection precision of the hip joint acetabulum index key points.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular to a method for detecting key points of acetabulum index by integrating edge features with a detail perception network. Background Art

[0002] There are various clinical diagnostic methods for developmental dysplasia of the hip (DDH). Usually, special points in the pelvic structure of the hip X-ray image are marked, and the corresponding angle is calculated based on the position information of these points. The value of the angle is the acetabular index (AI). The specific diagnosis of hip dysplasia is as follows: Figure 1 As shown in the figure, the key points include: 0) Right Acetabulum Superolateral Margin (RASM), 1) Right Tri-radiate Cartilage Center (RTCC), 2) Left Tri-radiate Cartilage Center (LTCC), 3) Left Acetabulum Superolateral Margin (LASM). The value of the acetabular index provides a key basis for evaluating the development of the hip joint, so accurately identifying and correctly marking the key points is a necessary prerequisite for diagnosing DDH.

[0003] In clinical diagnosis, the key points of the hip acetabulum index are usually manually marked, but this process is time-consuming and easily affected by subjective bias. At the same time, due to the low degree of bone modification in children and the poor quality of hip X-ray images, the edges of the bone area where the key points are located are blurred, and the key structures in the local neighborhood of the key points are easily disturbed by complex backgrounds, which limits the precise positioning of the key points. There are also studies in the prior art that use edge detectors and template matching technology to identify key points of the hip acetabulum index, but this method lacks consideration of global context information. When faced with challenges such as illumination changes, occlusions, and image rotation, its robustness and generalization capabilities are insufficient.

[0004] In addition, with the development of deep learning, the use of deep learning models has significantly improved the accuracy and generalization ability of image recognition by automatically learning and extracting complex visual features from large amounts of medical image data. However, it is costly to construct accurately annotated datasets in the field of medical images.

[0005] The deep learning-based DDH early diagnosis algorithm transforms the recognition target from the key points of the hip joint into the local neighborhood of the key points, locates the local neighborhood by learning the local features in these areas and the global features of the image, and then takes the center point of the local neighborhood as the key point of the hip acetabular index. Although this method can improve the positioning accuracy of the key points and the robustness of the algorithm, it has problems such as noise interference and insufficient use of contextual information. To this end, some studies use pyramid non-local modules, local-non-local modules, dual-level routing attention and other methods to reduce background interference by gradually narrowing the focus area, but these methods ignore the detailed edge features in the local neighborhood because they rely on deep features for attention interaction. Other studies use end-to-end detection networks or hourglass networks to mine contextual information from the refined feature level. However, these methods aggravate the dilution of edge information in the refined features due to frequent up- and down-sampling, while bone edges are the basic features that assist in identifying the key points of the hip acetabular index. In addition, in medical images, many key structures occupy a very small pixel area in the image. Although the attention mechanism performs well in deep network features, it is mostly difficult to apply to shallow network features with rich detailed information due to the large amount of computation, which limits its ability to capture subtle structures.

[0006] In summary, existing work has not fully considered the information of the bone edge area where the key points are located, and has deficiencies in identifying key structural features within the local neighborhood, making it difficult to effectively improve the positioning accuracy of key points. Summary of the invention

[0007] Based on this, it is necessary to provide a acetabular index key point detection method that integrates edge features and detail perception network to address the above technical problems.

[0008] A method for detecting acetabular index key points by integrating edge features and detail perception network comprises the following steps: obtaining an image to be detected, wherein the image to be detected includes an X-ray anteroposterior image of a hip joint; constructing a hip acetabular index key point detection model based on a YOLOv8s algorithm, combining edge feature enhancement and a detail perception network, and training a target hip acetabular index key point detection model; inputting the image to be detected into the target hip acetabular index key point detection model, and obtaining corresponding hip acetabular index key points through the output of the target hip acetabular index key point detection model.

[0009] In one of the embodiments, the hip acetabulum index key point detection model is constructed based on the YOLOv8s algorithm, combined with edge feature enhancement and detail perception network, including: taking YOLOv8s as the benchmark network model, constructing the hip acetabulum index key point detection model, including a backbone network, a neck network and a detection head; wherein the backbone network is used to extract basic feature information of the image, and an edge feature enhancement module is embedded in the cross-stage fusion module in the backbone network, which is used to capture the neighborhood information of the key points in the backbone network and enhance the edge area information of the bones where they are located, so as to obtain enhanced bone edge features; the neck network adopts a feature pyramid network and a path aggregation network to fuse multi-scale features, and is provided with a detail perception network to integrate feature information of different scales and capture subtle structural features in the local neighborhood of the key points; the detection head adopts a decoupled design to perform classification and regression tasks respectively to obtain prediction results.

[0010] In one embodiment, the training obtains a target hip acetabular index key point detection model, including: obtaining training images with marked hip acetabular index measurement key points, the training images including images of normal development of hip acetabular index and images of abnormal development of hip acetabular index; constructing a training data set based on all the acquired training images; using the training data set to train the hip acetabular index key point detection model, and based on the classification loss function and the regression loss function, judging whether the model converges, and obtaining the target hip acetabular index key point detection model after the training is completed.

[0011] In one of the embodiments, the edge feature enhancement module includes an edge enhancement branch, a convolution branch and a residual branch; the edge enhancement branch performs a convolution operation on the initial feature map through a sobel operator, extracts edge features in the horizontal and vertical directions respectively, and adds the extracted edge features to obtain enhanced bone edge features; the convolution branch uses a 3×3 convolution to capture the local features and basic structural information of the initial feature map to obtain basic features, splices the bone edge features and basic features, and performs batch normalization processing, uses a 1×1 convolution to process the spliced ​​features and reduces the number of feature channels to 1 / 2 to obtain a convolution splicing feature; the residual branch adds the initial feature map to the convolution splicing feature, and obtains an edge output feature through a 1×1 convolution.

[0012] In one embodiment, the neck network also includes a dual-scale fusion module, which performs 1×1 convolution and downsampling operations on the first input feature and the second input feature, respectively, and splices the two output features in the channel dimension to obtain feature one; inputs the feature one into three depth-separable convolution branches, respectively, and residually connects feature one to obtain feature two; processes feature two using 1×1 convolution and residually connects it with feature one to obtain a fused output feature.

[0013] In one embodiment, the neck network also includes a full-core module, the input feature of the full-core module is the spliced ​​feature obtained by splicing the upsampled feature of the fused output feature and the edge output feature, and the full-core module is used to integrate the shallow detail information and deep semantic information of the input feature; the full-core module includes a local branch, a large-scale branch and a global branch, and the input feature is adjusted by 1×1 convolution and input to the local branch, large-scale branch and global branch respectively; wherein the local branch obtains local information through 1×1 depthwise separable convolution; the large-scale branch uses convolution kernels of different sizes to capture contextual information; the global branch includes a dual-domain channel attention module and a frequency-based spatial attention module, and uses an attention mechanism to obtain a global receptive field; the output features of the three branches are added, and the full-core output features are obtained after adjustment by 1×1 convolution.

[0014] Compared with the prior art, the advantages and beneficial effects of the present invention are: by acquiring an X-ray anteroposterior image containing a hip joint, based on the YOLOv8s algorithm, combined with edge feature enhancement and detail perception network, a hip joint acetabular index key point detection model is constructed, and a target hip joint acetabular index key point detection model is trained. The model can capture the spatial features around the key points and enhance the edge features of the bones where they are located through edge feature enhancement, and the detail perception network can enhance the network's ability to capture subtle structural features. When performing detection, the image to be detected is input into the target hip joint acetabular index key point detection model, and the corresponding hip joint acetabular index key points are output. The detection accuracy of the hip joint acetabular index key points is improved by identifying the edge area features of the key points and perceiving the subtle structural features. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of key points for reference in the diagnosis of developmental dysplasia of the hip;

[0016] Figure 2 It is a flowchart of a method for detecting key points of acetabular index by fusing edge features with a detail perception network in one embodiment;

[0017] Figure 3The overall network architecture of EDA-YOLOv8s in one embodiment;

[0018] Figure 4 Schematic diagrams of three neck network architectures in YOLOv8s in one embodiment, wherein (a) is the neck network architecture of YOLOv8s, (b) is the neck network architecture for small target detection, and (c) is the neck network architecture including the DAN network;

[0019] Figure 5 is a schematic diagram of the architecture of an edge feature enhancement module in one embodiment;

[0020] Figure 6 is a schematic diagram of the architecture of a dual-scale fusion module in one embodiment;

[0021] Figure 7 Schematic diagram of the architecture of a full-core module in one embodiment;

[0022] Figure 8 A schematic diagram of the effect of key point marking in one embodiment;

[0023] Fig. 9 This is a schematic diagram of the accuracy comparison of different detection algorithms based on YOLOv8s in an embodiment. DETAILED DESCRIPTION

[0024] Before describing the specific embodiments of the present invention, the overall concept of the present invention is described as follows:

[0025] The present invention is mainly developed for the detection process of the key points of the hip acetabulum index. Currently, due to the low degree of bone modification in children and the poor quality of hip joint X-ray images, the edges of the bone area where the key points are located are blurred, and the key structures within the local neighborhood of the key points are easily affected by complex background interference, making it difficult to improve the positioning accuracy of the key points.

[0026] Therefore, the present invention proposes a acetabular index key point detection method that integrates edge features and detail perception networks. By acquiring an X-ray anteroposterior image containing a hip joint, a hip acetabular index key point detection model is constructed based on the YOLOv8s algorithm, combined with edge feature enhancement and detail perception networks, and a target hip acetabular index key point detection model is trained. The model can capture the spatial features around the key points and enhance the edge features of the bones where the key points are located through edge feature enhancement, and the detail perception network can enhance the network's ability to capture subtle structural features. When performing detection, the image to be detected is input into the target hip acetabular index key point detection model, and the corresponding hip acetabular index key points are output. The detection accuracy of the hip acetabular index key points is improved by recognizing the edge area features of the key points and perceiving the subtle structural features.

[0027] After introducing the overall concept of the present invention, in order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail by specific implementation methods in combination with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0028] In one embodiment, Figure 2 As shown, a method for detecting key points of acetabular index by integrating edge features and detail perception network is provided, comprising the following steps:

[0029] Step S110, obtaining an image to be detected, wherein the image to be detected includes an X-ray AP image of the hip joint.

[0030] Specifically, when hip joint development detection is required, an image to be detected is obtained, and the image to be detected includes an X-ray anteroposterior image of the hip joint, so as to facilitate the detection of key points of the hip joint acetabulum index.

[0031] Step S120, based on the YOLOv8s algorithm, combined with edge feature enhancement and detail perception network, a hip joint acetabular index key point detection model is constructed, and a target hip joint acetabular index key point detection model is trained.

[0032] Specifically, based on the YOLOv8s algorithm, the edge feature enhancement and detail perception network are integrated to obtain the key point detection algorithm for hip dysplasia (Enhanced Edge Feature and Detail-Aware Pyramid Network for YOLOv8s, EDA-YOLOv8s). The algorithm architecture is as follows Figure 3 As shown in the figure, a hip joint acetabular index key point detection model is constructed, and the hip joint acetabular index key point detection model is trained with a training data set to obtain a target acetabular angle key point detection module to improve the detection accuracy of key points. The edge feature enhancement captures the neighborhood information of key points in the backbone network while enhancing the information of the edge area of ​​the bone. The algorithm improves the original neck network and integrates feature information of different scales through detail perception to enhance the network's ability to capture important subtle structures in the local neighborhood of key points.

[0033] Among them, step S120 includes: using YOLOv8s as a benchmark network model to construct a hip acetabulum index key point detection model, including a backbone network, a neck network and a detection head; wherein the backbone network is used to extract basic feature information of the image, and an edge feature enhancement module is embedded in the cross-stage fusion module in the backbone network, which is used to capture the neighborhood information of the key points in the backbone network and enhance the edge area information of the bones where they are located, so as to obtain enhanced bone edge features; the neck network adopts a feature pyramid network and a path aggregation network to fuse multi-scale features, and is provided with a detail perception network to integrate feature information of different scales and capture subtle structural features in the local neighborhood of the key points; the detection head adopts a decoupled design to perform classification and regression tasks respectively to obtain prediction results.

[0034] Specifically, as a one-stage target detector, YOLO directly predicts the position of the bounding box and the category of the target by converting the target detection task into a regression problem, using the entire image as the network input, and has obvious advantages in processing speed. Through comparative experiments on YOLOv8, YOLOv9, YOLOv10 and their detection networks of different scales, we finally used YOLOv8s, which performed best on the dataset, as the benchmark network model to build a hip acetabular index key point detection model. YOLOv8 achieves a balance between high efficiency and high precision in target detection.

[0035] The hip acetabulum index key point detection model consists of three parts: Backbone network, Neck network and Head detection head. The Backbone is used to extract the basic feature information of the image, and the cross-stage fusion module in the backbone feature network is embedded with an edge feature enhancement module, which is used to capture the neighborhood information of the key points in the backbone network and enhance the edge area information of the bones where they are located to obtain the enhanced bone edge features; the neck network adopts Feature Pyramid Networks (FPN) and Path Aggregation Network (PANet) to effectively fuse multi-scale features to enhance the network's recognition sensitivity to the target object, and is equipped with a detail perception network to integrate feature information of different scales and capture the subtle structural features in the local neighborhood of the key points; the detection head adopts a decoupled design to perform classification and regression tasks respectively to obtain prediction results.

[0036] The target hip joint acetabular index key point detection model converts key point detection into local neighborhood detection of key points. The network provides more contextual information about key points. The designed EFEM (Edge feature enhancement module) is embedded in the four C2f (Crossstage feature fusion) modules in the backbone network to enhance the edge information of key points lost in the convolution operation process in the backbone network.

[0037] In hip joint X-ray images, subtle key structures in the local neighborhood of key points can assist the network in identifying key points. In the YOLO detection network, by optimizing the feature fusion strategy of the Neck network or incorporating high-resolution features, the network's ability to perceive subtle structures can be enhanced, such as Figure 4 As shown, three different Neck network architectures based on YOLOv8s are shown.

[0038] In the Neck network of YOLOv8s, the size of the highest resolution input feature P3 is only 80 pixels × 80 pixels, such as Figure 4 As shown in (a) in the figure, each pixel of this layer is responsible for compressing and expressing the 64-pixel area information of the initial input image. As the network level goes deeper, the area that each pixel needs to represent increases significantly. Although this high-load information compression helps integrate features, it also restricts the network's extraction and expression of subtle structures in the image.

[0039] In order to improve the network's ability to express subtle structures and recognize small targets, some existing technologies use P2 features with rich detail information in YOLOv8s to build a small target detection head to assist the network in extracting detail features in the image to detect small targets that have been missed, such as Figure 4 As shown in (b) in , this type of method increases the amount of computation and does not optimize the high-detail information provided by the P2 feature.

[0040] This embodiment designs a DAN network based on the original Neck network. Figure 4 As shown in (c) in the figure, DAN inputs the high-resolution P2 and P3 features into the Dual-Scale Fusion Module (DSFM) to perform feature focusing, and then concatenates them with the features sampled from the deep network. By using the lightweight Omni-Kernel Module (OKM), the refined shallow detail information is fully integrated with the deep semantic information to capture the key subtle structures.

[0041] This embodiment adopts DAN (Detail aware network) in the neck network. By focusing on features of different high-resolution levels and fusing them with deep semantic features, it realizes the comprehensive integration of information from coarse-grained to fine-grained, enhances the model's ability to extract important subtle structures around key points, and thus improves the positioning accuracy of key points.

[0042] Among them, the training obtains the target hip joint acetabular index key point detection model, including: obtaining training images with marked hip joint acetabular index measurement key points, the training images include images of normal hip joint development and images of abnormal hip joint development; constructing a training data set based on all the acquired training images; using the training data set to train the hip joint acetabular index key point detection model, and based on the classification loss function and the regression loss function, judging whether the model converges, and the target hip joint acetabular index key point detection model is obtained after the training is completed.

[0043] Specifically, when training the model, a training data set is constructed by acquiring existing training images with marked acetabular angle measurement key points, wherein the training images include images of normal hip joint development and images of abnormal hip joint development, so as to facilitate the training model to detect the development of the hip joint; at the same time, a classification loss function and a regression loss function are used to balance positive and negative samples through a dynamic strategy to determine whether the model converges. After the training is completed, a target hip joint acetabular index key point detection model is obtained, thereby improving the consistency and accuracy of the detection network output results.

[0044] Among them, the edge feature enhancement module includes an edge enhancement branch, a convolution branch and a residual branch; the edge enhancement branch performs a convolution operation on the initial feature map through the sobel operator, extracts edge features in the horizontal and vertical directions respectively, and adds the extracted edge features to obtain the enhanced bone edge features; the convolution branch uses 3×3 convolution to capture the local features and basic structural information of the initial feature map to obtain the basic features, splices the bone edge features and the basic features, and performs batch normalization, and uses 1×1 convolution to process the spliced ​​features and reduce the number of feature channels to 1 / 2 to obtain the convolution splicing features; the residual branch adds the initial feature map and the convolution splicing features, and obtains the edge output features through 1×1 convolution.

[0045] Specifically, during the network convolution process, the bone edge information where the key points are located will gradually be diluted, which will affect the accurate recognition of the bone contour, thereby limiting the network's ability to accurately locate the key points. Therefore, this embodiment uses an edge feature enhancement module to enable the network to enhance its ability to capture bone edge features while ensuring the capture of surrounding spatial information.

[0046] The structure of the edge feature enhancement module is as follows: Figure 5 As shown in Figure 1, it includes edge enhancement branch, ordinary convolution branch and residual branch. in When the initial feature map enters the edge enhancement branch, the Sobel operator is used for convolution operation to extract edge features in the horizontal and vertical directions respectively, and the extraction results in the two directions are added to obtain the enhanced bone edge feature f1. in The image is sent to the ordinary convolution branch, and a 3×3 ordinary convolution is used to capture the local features and basic structural information of the image to obtain the basic feature f2. Considering that the outputs of the two branches may differ in numerical scale, the output features f1 and f2 are concatenated in the channel dimension and batch normalized (BatchNorm) is performed to ensure the consistency of the numerical scales of different types of features. Subsequently, a 1×1 ordinary convolution is used to further refine the concatenated features and reduce the number of feature channels to 1 / 2 to obtain the convolution concatenated feature f3. In addition, the edge feature enhancement module also supplements the input information through the residual branch to convert the initial feature map f in Directly add the feature f3 after the fusion of the edge enhancement branch and the ordinary convolution branch, and adjust it through 1×1 ordinary convolution to generate a finer feature representation to obtain the edge output feature f out The specific calculation process is as follows:

[0047] f1=Sobelx(fin)+Sobely(fin) (1)

[0048] f2=Conv3×3(fin) (2)

[0049] f3=Conv1×1(BatchNorm(Concat(f1,f2))) (3)

[0050] fout=Conv1×1(fin+f3) (4)

[0051] By fusing the features extracted by the ordinary convolution branch and the edge enhancement branch, the network can enhance its sensitivity to edge features while extracting local features of the image, thereby providing the network with richer and more accurate feature information.

[0052] Among them, the neck network also includes a two-scale fusion module, which performs 1×1 convolution and downsampling operations on the first input feature and the second input feature respectively, and splices the two output features in the channel dimension to obtain feature one; feature one is input into three depth-separable convolution branches respectively, and feature one is residually connected to obtain feature two; feature two is processed by 1×1 convolution and residually connected with feature one to obtain a fused output feature.

[0053] Specifically, although incorporating higher-resolution features helps improve the network's ability to recognize detail features, directly using existing methods to use these features to build a small target detection head not only increases the computational burden, but may also introduce irrelevant details. In order to reduce the amount of computation and optimize the input detail features, this application designs a DSFM (Dual-scale fusion module), such as Figure 6 As shown in (a) of Figure 2, the input of this module consists of two features: P2 and P3. Compared with the conventional downsampling operation, Figure 6 (b) Adown (downsampling) module in the downsampling, Figure 6 In (b), s is the step size, and P represents padding, which means adding zero values ​​to the edge of the input feature map to maintain or change the size of the output feature map. Padding can prevent information loss during the convolution process. The Adown module can reduce the amount of calculation and maximize the retention of key information of the image. Therefore, the DSFM module performs a P2 feature f with an input size of C×H×W. in1 ADown convolution is performed to downsample, so that the feature map size is reduced by 1 / 2 while the number of channels remains unchanged. At the same time, the P3 feature f with an input size of 2C×H / 2×W / 2 is in2 Then, a 1×1 normal convolution is used to adjust the number of channels to halve. Then, the output is summed with f in the channel dimension. in1 The corresponding output features are concatenated to obtain feature 1 f1, so that features at different levels can complement each other. Therefore, the present application inputs the concatenated feature f1 into three depth-separable convolution branches to enhance the perception of subtle structures, and adds residual connections to supplement the original information to obtain feature 2 f2. Finally, the f2 feature is adjusted using a 1×1 ordinary convolution, and then added to the residual connection to obtain the fused output feature f out The specific operation process is as follows:

[0054] f1=Concat(ADown(fin1),Conv1×1(fin2))(5)

[0055]

[0056] fout=Conv1×1(f2)+f1(7)

[0057] Where n is the number of depth-wise separable convolution branches, which is set to 3 in this embodiment, k is the size of the convolution kernel, which is set to 3, 5, and 7 respectively, and f in1 is the P2 feature, f in2 is the P3 feature, f1 and f2 are both intermediate layer features within the module, and f out is the fusion output feature.

[0058] In this embodiment, the dual-scale fusion module introduces high-resolution hierarchical features and performs feature fusion and focus optimization, which not only enhances the network's understanding of local image features but also improves its sensitivity to subtle structures in the image, ultimately improving the detection accuracy of the model.

[0059] Among them, the neck network also includes a full-core module. The input features of the full-core module are the concatenated features obtained by concatenating the upsampling features of the fused output features and the edge output features. The full-core module is used to integrate the shallow detail information and deep semantic information of the input features; the full-core module includes local branches, large-scale branches and global branches. The input features are adjusted by 1×1 convolution and input to the local branches, large-scale branches and global branches respectively; the local branches obtain local information through 1×1 depthwise separable convolution; the large-scale branches use convolution kernels of different sizes to capture contextual information; the global branch includes a dual-domain channel attention module and a frequency-based spatial attention module, and uses the attention mechanism to obtain the global receptive field; the output features of the three branches are added and adjusted by 1×1 convolution to obtain the full-core output features.

[0060] Specifically, in the Neck network, the DSFM module provides refined detail information by performing dual-scale feature fusion on the P2 and P3 features, while the features sampled on the deep network provide stronger semantic information. In order to effectively fuse the two features after splicing, this embodiment introduces the OKM module to optimize the input features. In addition, in order to reduce the amount of calculation, the module is embedded in the CSP (Cross Stage Partial) structure. In the convolutional neural network, the CSP structure is mainly used to improve the performance and feature representation ability of the model, thereby achieving lightweight, such as Figure 7 As shown in (a) in .

[0061] The OKM module consists of three branches: local branch, large-scale branch and global branch, such as Figure 7 In the feature fusion process, first, the input features are adjusted through 1×1 ordinary convolution; secondly, the adjusted features are input into the three branches for their respective operations. Specifically, the local branch supplements the fine local information through 1×1 depthwise separable convolution, the large-scale branch uses large convolution kernels of different sizes to capture a wider range of contextual information, and the global branch uses the attention mechanism to obtain the global receptive field, as shown in (d). Figure 7As shown in (b) and (c) in the figure, the dual-domain channel attention module (DCAM) and the frequency-based spatial attention module (FSAM) are used to enhance the attention to the subtle structure in the neighborhood of the key point in different dimensions, so as to help the network understand the location and feature distribution of the key points globally; finally, the output features of the three branches are added and adjusted using a 1×1 ordinary convolution to obtain the full-core output features.

[0062] In this embodiment, the OKM module can sense and extract key structural information near key points through a multi-branch architecture, and this information helps the auxiliary network to further accurately identify key points.

[0063] Step S130, inputting the image to be detected into the target hip joint acetabular index key point detection model, and obtaining the corresponding hip joint acetabular index key point through the output of the target hip joint acetabular index key point detection model.

[0064] Specifically, after the model is constructed and trained, the image to be detected is input into the target hip joint acetabular index key point detection model, and the marked hip joint acetabular index key points are output. The marking effect is as follows: Figure 8 As shown, the target acetabular angle can be calculated based on the marked hip acetabular index key points, thereby realizing the judgment of the development of the hip acetabular index. The accuracy of the judgment result is significantly better than the existing method, and can effectively improve the detection accuracy of the hip joint key points.

[0065] In this embodiment, by acquiring an X-ray anteroposterior image containing a hip joint, a hip acetabulum index key point detection model is constructed based on the YOLOv8s algorithm, combined with edge feature enhancement and detail perception network, and a target hip acetabulum index key point detection model is trained. The model can capture the spatial features around the key points and enhance the edge features of the bones where they are located through edge feature enhancement, and the detail perception network can enhance the network's ability to capture subtle structural features. When performing detection, the image to be detected is input into the target hip acetabulum index key point detection model, and the corresponding hip acetabulum index key points are output. The detection accuracy of the hip acetabulum index key points is improved by recognizing the edge area features of the key points and perceiving the subtle structural features.

[0066] In one embodiment, the effectiveness of key point detection is verified on a hip joint X-ray image dataset. The detection accuracy of YOLOv8s, YOLOv8s+EFEM, YOLOv8s+DAN and the EDA-YOLOv8s algorithm of this application is compared. Fig. 9 As shown in the figure, the experimental results show that the average positioning error and average angle error of the key points are reduced to 4.2090 pixels and 1.4872°, respectively, which are 6.8% and 9.9% lower than those of YOLOv8s, significantly improving the detection accuracy of the key points of the hip acetabulum index, providing a more credible scientific basis for clinical applications.

[0067] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0068] Obviously, those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a computer storage medium (ROM / RAM, magnetic disk, optical disk) and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than that here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Therefore, the present invention is not limited to any specific combination of hardware and software.

[0069] The above contents are further detailed descriptions of the present invention in combination with specific implementation methods, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A method for detecting key points of acetabular index by integrating edge features and detail perception network, characterized in that: The following steps are involved: Acquire an image to be detected, wherein the image to be detected includes an X-ray anteroposterior image of the hip joint; Based on the YOLOv8s algorithm, combined with edge feature enhancement and detail perception network, a hip joint acetabular index key point detection model was constructed, and the target hip joint acetabular index key point detection model was trained; The image to be detected is input into the target hip joint acetabular index key point detection model, and the corresponding hip joint acetabular index key point is obtained through the output of the target hip joint acetabular index key point detection model.

2. The acetabular index key point detection method of integrating edge features and detail perception network according to claim 1 is characterized in that: The hip joint acetabulum index key point detection model is constructed based on the YOLOv8s algorithm, combined with edge feature enhancement and detail perception network, including: Using YOLOv8s as the benchmark network model, a hip acetabulum index key point detection model was constructed, including the backbone network, neck network and detection head; The backbone network is used to extract basic feature information of the image, and an edge feature enhancement module is embedded in the cross-stage fusion module in the backbone network, which is used to capture the neighborhood information of key points in the backbone network and enhance the edge area information of the bones where they are located, so as to obtain enhanced bone edge features; The neck network adopts a feature pyramid network and a path aggregation network to fuse multi-scale features, and is provided with a detail perception network to integrate feature information of different scales and capture subtle structural features in the local neighborhood of key points; The detection head adopts a decoupled design to perform classification and regression tasks separately to obtain prediction results.

3. The acetabular index key point detection method of integrating edge features and detail perception network according to claim 1 is characterized in that: The training obtains a target hip joint acetabulum index key point detection model, including: Acquire training images with marked hip joint measurement key points, wherein the training images include images of normal hip joint development and images of abnormal hip joint development; Construct a training data set based on all the acquired training images; The training data set is used to train the hip joint acetabulum index key point detection model, and based on the classification loss function and the regression loss function, it is judged whether the model converges, and the target hip joint key point detection model is obtained after the training is completed.

4. The acetabular index key point detection method of integrating edge features and detail perception network according to claim 2 is characterized in that: The edge feature enhancement module includes an edge enhancement branch, a convolution branch and a residual branch; The edge enhancement branch performs a convolution operation on the initial feature map through a sobel operator, extracts edge features in the horizontal direction and the vertical direction respectively, and adds the extracted edge features to obtain enhanced bone edge features; The convolution branch uses 3×3 convolution to capture the local features and basic structural information of the initial feature map to obtain basic features, splices the bone edge features and basic features, and performs batch normalization processing, and uses 1×1 convolution to process the spliced ​​features and reduce the number of feature channels to 1 / 2 to obtain convolution splicing features; The residual branch adds the initial feature map to the convolutional splicing feature and obtains the edge output feature through a 1×1 convolution.

5. The acetabular index key point detection method of integrating edge features and detail perception network according to claim 4 is characterized in that: The neck network further includes a dual-scale fusion module, which performs 1×1 convolution and downsampling operations on the first input feature and the second input feature respectively, and concatenates the two output features in the channel dimension to obtain feature one; Input the feature one into three depth-separable convolution branches respectively, and residually connect the feature one to obtain the feature two; The feature 2 is processed by 1×1 convolution and is residually connected with the feature 1 to obtain a fused output feature.

6. The acetabular index key point detection method of integrating edge features and detail perception network according to claim 5 is characterized in that: The neck network also includes a full-core module, the input feature of the full-core module is a spliced ​​feature obtained by splicing the up-sampled feature of the fused output feature and the edge output feature, and the full-core module is used to integrate shallow detail information and deep semantic information of the input feature; The full-core module includes a local branch, a large-scale branch and a global branch. After the input features are adjusted by 1×1 convolution, they are input to the local branch, the large-scale branch and the global branch respectively. The local branch obtains local information through 1×1 depth-wise separable convolution; the large-scale branch uses convolution kernels of different sizes to capture contextual information; the global branch includes a dual-domain channel attention module and a frequency-based spatial attention module, and uses an attention mechanism to obtain the global receptive field; The output features of the three branches are added together and adjusted using 1×1 convolution to obtain the full-core output features.

Images