Automatic segmentation method of shoulder joint glenoid cavity based on artificial intelligence
By using an automatic segmentation method based on a 3D U-Net network and the angle relationship between key point normal vectors, the problem of inaccurate measurement of the glenoid cavity in existing technologies is solved, achieving efficient and accurate extraction of the glenoid surface and improving segmentation accuracy and consistency.
Patent Information
- Application Number
- CN202511090386.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2025-11-18
AI Technical Summary
Existing methods for measuring glenoid bone defects in the shoulder joint lack accuracy and consistency, and neglect the influence of glenoid depth and indentation on shoulder joint stability, resulting in insufficient reliability and accuracy of measurement results.
A glenoid segmentation model based on a 3D U-Net network was constructed. By combining the angle relationship between the normal vectors of key points and sphere fitting, the glenoid surface was automatically segmented, the extraction range was optimized, and image preprocessing and segmentation models were used for training and segmentation.
It improves the segmentation accuracy and efficiency of glenoid segmentation, reduces the influence of operator experience, provides consistent segmentation results, and can quickly process large amounts of image data.
Smart Images
Figure CN120976241A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of medical image processing, and particularly relates to a shoulder joint glenoid automatic segmentation method based on artificial intelligence. BACKGROUND
[0002] Shoulder dislocation is a common injury, especially anterior dislocation, which can cause shoulder bone defects during the injury process, especially the recurrence rate of which is as high as 90%. In recent years, studies have shown that in addition to the degree of shoulder glenoid bone defect, the depth or degree of glenoid concave is also very important for evaluating shoulder dislocation.
[0003] There are many methods for measuring bone defects at present, including linear measurement and area measurement based on CT and MR, most of which are based on the "best fitting circle method", but these methods do not consider the influence of the degree of concave or depth on the stability of the shoulder joint. In addition, studies have shown that different angles or directions measured by the "best fitting circle method" will have significant differences. Due to the differences in these measurement methods, the results of calculating the degree of bone defect are different, and the reliability and accuracy are also different. At present, there is still a lack of a universally accepted measurement method.
[0004] At present, artificial intelligence (AI) has been widely applied in the field of medical image processing, but its application in the shoulder joint is relatively less. Therefore, it is urgent to propose a shoulder joint glenoid automatic segmentation method based on artificial intelligence. SUMMARY
[0005] To solve the above technical problems, the application provides a shoulder joint glenoid automatic segmentation method based on artificial intelligence to solve the problems of inaccurate measurement, great influence of operator experience and ignoring the depth and concave degree of the glenoid in the prior art.
[0006] To achieve the above purpose, the application provides a shoulder joint glenoid automatic segmentation method based on artificial intelligence, which comprises the following steps:
[0007] Obtaining a shoulder joint scan image, performing resampling and standardization processing;
[0008] Constructing a shoulder joint glenoid segmentation model based on a 3D U-Net network, and training the shoulder joint glenoid segmentation model based on the shoulder joint scan image after standardization processing;
[0009] Segmenting the shoulder joint scan image to be processed based on the trained shoulder joint glenoid segmentation model;
[0010] Selecting key points on the segmented glenoid surface, screening potential points by using the normal vector angle relationship of the key points, and combining sphere fitting to optimize the extraction range of the glenoid surface to finally obtain the extracted glenoid surface.
[0011] Optionally, the process of obtaining the shoulder joint scan image, resampling and standardizing includes:
[0012] The shoulder joint scan image is resampled, the z-axis interval is adjusted to match the x-axis and y-axis intervals, and the excess slices in the z-axis direction are cropped to obtain a cropped shoulder joint scan image; and then the cropped shoulder joint scan image is standardized by using a window width of 1800 Hu and a window level of 400 Hu.
[0013] Optionally, the 3D U-Net network adopts an encoder-decoder architecture, extracts high-level semantic features through the encoder part, combines the extracted high-level semantic features with spatial detail information of the decoder part through a jump connection, and then completes image segmentation.
[0014] Optionally, each convolutional block of the 3D U-Net network is composed of a 3x3x3 convolutional layer followed by a batch normalization layer and a ReLU activation layer.
[0015] Optionally, the downsampling operation of the image segmentation based on the 3D U-Net network is performed using maximum pooling, and the upsampling operation is realized by trilinear interpolation.
[0016] Optionally, the process of selecting key points on the segmented glenoid surface, screening potential points using the normal vector angle relationship of the key points, and combining the sphere fitting to optimize the extraction range of the glenoid surface to finally obtain the extracted glenoid surface includes:
[0017] Two key points are selected from the segmented glenoid surface, and a first included angle between the normal vectors of the two key points is obtained; wherein the first point is located near the bottom of the glenoid, and the second point is located inside the glenoid;
[0018] The second included angle between the normal vector of each point on the surface of the scapula and the normal vector of the first point is obtained, and points satisfying the second included angle less than the first included angle are screened out;
[0019] A third point is introduced inside the glenoid, a third included angle between the normal vector of the third point and the normal vector of the first point is obtained, and points satisfying the third included angle less than the first included angle are screened out;
[0020] The points screened out twice are fitted into a sphere, and the extracted glenoid surface is finally obtained.
[0021] The application also provides an artificial intelligence-based automatic glenoid segmentation system for the shoulder joint, which is used to implement the method, and includes an image preprocessing module, a segmentation model construction module, a glenoid surface segmentation module, and a glenoid surface extraction module.
[0022] The image preprocessing module is configured to acquire a shoulder joint scan image, and perform resampling and standardization processing on the shoulder joint scan image.
[0023] The segmentation model construction module is configured to construct a shoulder joint glenoid segmentation model based on a 3D U-Net network, and train the shoulder joint glenoid segmentation model based on the shoulder joint scan image after standardization processing.
[0024] The glenoid surface segmentation module is configured to segment the shoulder joint scan image to be processed based on the trained shoulder joint glenoid segmentation model.
[0025] The glenoid surface extraction module is configured to select key points on the segmented glenoid surface, filter potential points by using the normal vector angle relationship of the key points, and combine a sphere fitting to optimize the extraction range of the glenoid surface, so as to finally obtain the extracted glenoid surface.
[0026] The application further provides a computer device, including a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the method.
[0027] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the method.
[0028] The application further provides a computer program product, which includes a computer program, and the computer program is executed by a processor to realize the steps of the method.
[0029] Compared with the prior art, the application has the following advantages and technical effects:
[0030] The application constructs a shoulder joint glenoid segmentation model based on a 3D U-Net network, and trains the model; segments the shoulder joint scan image to be processed based on the trained shoulder joint glenoid segmentation model; selects key points on the segmented glenoid surface, filters potential points by using the normal vector angle relationship of the key points, and combines a sphere fitting to optimize the extraction range of the glenoid surface, so as to finally obtain the extracted glenoid surface. Compared with the traditional manual method, the automatic segmentation method proposed by the application can quickly process a large amount of image data, improve the segmentation efficiency, and is not affected by the experience of operators, can provide consistent segmentation results based on the 3D U-Net network, and greatly improves the segmentation accuracy. BRIEF DESCRIPTION OF DRAWINGS
[0031] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application and are incorporated herein for illustrative purposes. The schematic embodiments of the present application and their descriptions are used to explain the present application, and do not constitute improper limitations on the present application. In the drawings:
[0032] Figure 1A schematic diagram of a shoulder glenoid segmentation model based on a 3D U-Net network according to an embodiment of the present application;
[0033] Figure 2 A schematic diagram of shoulder glenoid surface extraction according to an embodiment of the present application. DETAILED DESCRIPTION
[0034] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.
[0035] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.
[0036] Embodiment one
[0037] Regarding the measurement of the shoulder joint, most of the previous studies are based on the "best fitting circle method" measurement or other manual measurement methods, and the consistency coefficient (ICC) value of these methods is usually low. Although the best fitting circle method is widely used, its results are often affected by the measurement angle, the selection of the center point and the experience of the delineator. In addition, the method based on the "best fitting circle method" measurement is actually to analyze the glenoid as a plane, ignoring the existence of depth. The glenoid is a curved surface structure and has a certain depth. Studies have shown that the change of depth is also an important factor affecting the stability of the shoulder joint. In view of the above problems, the present embodiment provides an artificial intelligence-based automatic segmentation method for the glenoid of the shoulder joint, comprising the following steps:
[0038] Obtaining a shoulder joint scan image, performing resampling and standardization processing;
[0039] Building a shoulder glenoid segmentation model based on a 3D U-Net network, and training the shoulder glenoid segmentation model based on the shoulder joint scan image after standardization processing;
[0040] Segmenting the shoulder joint scan image to be processed based on the trained shoulder glenoid segmentation model;
[0041] Selecting key points on the segmented glenoid surface, screening potential points using the normal vector angle relationship of the key points, and combining sphere fitting to optimize the extraction range of the glenoid surface, to finally obtain the extracted glenoid surface.
[0042] As a specific implementation, the process of obtaining a shoulder joint scan image, performing resampling and standardization processing includes:
[0043] The shoulder joint scan image is acquired, and due to the anisotropic spacing of the original shoulder joint scan image in different directions, the spacing of the z-axis is unified to match the spacing of the x-axis and the y-axis. The resampled data has extra slices in the z-axis direction, so it needs to be cropped to a size of (512, 512, 128). Then a standardization operation is performed on it with a window width of 1800 Hu and a window level of 400 Hu.
[0044] As a specific embodiment, for the bones with clear three-dimensional continuous structure, the deep learning-based method is very effective in achieving accurate segmentation results. Therefore, the segmentation of the present embodiment is realized by constructing a shoulder glenoid segmentation model based on a 3D-UNet network. The 3D U-Net network adopts an encoder-decoder architecture, extracts high-level semantic features through the encoder part, and combines the extracted high-level semantic features with the spatial detail information of the decoder part through the jump connection to retain important spatial details. These jump connections allow the network to capture richer context information and improve the ability to process complex 3D data.
[0045] Further, each convolutional block of the 3D U-Net network is composed of a 3x3x3 convolutional layer followed by a batch normalization layer and a ReLU activation layer.
[0046] Further, the down-sampling operation of the image segmentation based on the 3D U-Net network is performed using maximum pooling, and the up-sampling operation is realized by trilinear interpolation. The detailed network structure is shown in Figure 1 .
[0047] As a specific embodiment, the process of training the shoulder glenoid segmentation model includes: randomly selecting 20% from the data set as the test set, and the rest as the training set to train the shoulder glenoid segmentation model. In the training stage, the present embodiment adopts a block training strategy, which can solve the problem of insufficient GPU memory. At the same time, there is no need to crop data when inferring new data. The trained network can effectively segment the glenoid region to further extract the glenoid surface.
[0048] As a specific embodiment, the process of selecting key points on the segmented glenoid surface, screening potential points using the normal vector angle relationship of the key points, and combining sphere fitting to optimize the extraction range of the glenoid surface to finally obtain the extracted glenoid surface includes:
[0049] Two key points are selected from the segmented glenoid surface, and the first angle between the normal vectors of these two key points is obtained. The first point is located near the bottom of the glenoid, and the second point is located inside the glenoid. The second angle between the normal vector of each point on the scapular surface and the normal vector of the first point is obtained, and points that satisfy the condition that the second angle is less than the first angle are selected. A third point is introduced located inside the glenoid, and the third angle between the normal vector of the third point and the normal vector of the first point is obtained. Points that satisfy the condition that the third angle is less than the first angle are selected. The points selected in the two steps are fitted into a sphere to finally obtain the extracted glenoid surface.
[0050] As an alternative implementation, the glenoid cavity can be simply considered as a bowl-shaped structure. Therefore, in this embodiment, two sample points are manually selected: point A at the bottom of the bowl and point B on the inner side of the bowl. As long as the angle β between the normal vector of a point on the scapular surface and the normal vector of point A is less than the angle θ formed by the normal vectors of points A and B, and greater than 0 (in other words, 0 < β < θ), then this point can be classified as part of the glenoid cavity. However, the shape of the glenoid cavity is not always so simple. Some points may satisfy the condition 0 < β < θ, but are located at the boundary between the glenoid cavity and other areas, causing the extraction results to extend beyond the glenoid cavity region. To alleviate this problem, such as... Figure 2 As shown, the purple area represents the glenoid surface of the shoulder joint, the green area represents the scapula, and the arrows of different colors represent the surface normal vectors at positions A, B, and C. First, points A and C are determined, where point A is located near the bottom of the glenoid fossa, and point C is located inside the glenoid fossa, forming an angle θ1 less than θ with point A. θ1 is used to identify points definitely within the glenoid fossa, and then a sphere is fitted using these points. In this way, the glenoid fossa can be stably extracted when the condition 0 < β < θ is applied again. Using the extracted glenoid fossa, calculating numerical parameters becomes simple. The reciprocal of the radius of the sphere fitted using the extracted points represents the curvature of the glenoid fossa. As for the area of the glenoid fossa, this embodiment directly uses the "mesh_surface_area" function from the "scikit-image" library in Python to calculate it, and the depth is determined by the shortest side of the minimum bounding box surrounding the glenoid fossa, providing more comprehensive information for clinical evaluation.
[0051] As a specific implementation method, this embodiment also optimizes the segmentation results, including but not limited to noise removal and hole filling, to ensure the accuracy and robustness of the segmentation.
[0052] Example 2
[0053] This embodiment also provides an artificial intelligence-based automatic glenoid segmentation system for implementing the method, including: an image preprocessing module, a segmentation model construction module, a glenoid surface segmentation module, and a glenoid surface extraction module;
[0054] The image preprocessing module is configured to acquire a shoulder joint scan image, perform resampling and standardization processing on the shoulder joint scan image.
[0055] The segmentation model construction module is configured to construct a shoulder joint glenoid segmentation model based on a 3D U-Net network, and train the shoulder joint glenoid segmentation model based on the shoulder joint scan image after the standardization processing.
[0056] The glenoid surface segmentation module is configured to segment the shoulder joint scan image to be processed based on the trained shoulder joint glenoid segmentation model.
[0057] The glenoid surface extraction module is configured to select key points on the segmented glenoid surface, filter potential points by using a normal vector angle relationship of the key points, and combine a sphere fitting to optimize an extraction range of the glenoid surface, so as to finally obtain the extracted glenoid surface.
[0058] Embodiment three
[0059] The embodiment also provides a computer device, including a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the method.
[0060] Embodiment four
[0061] The embodiment also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the method.
[0062] Embodiment five
[0063] The embodiment also provides a computer program product, which includes a computer program, and the computer program is executed by a processor to realize the steps of the method.
[0064] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims
1. An automatic segmentation method for the glenoid cavity of the shoulder joint based on artificial intelligence, characterized in that, Includes the following steps: Acquire shoulder joint scan images, and perform resampling and standardization processing; A glenoid segmentation model for the shoulder joint is constructed based on a 3D U-Net network, and the glenoid segmentation model is trained based on standardized shoulder joint scan images. The trained glenoid segmentation model is used to segment the shoulder joint scan image to be processed. Key points are selected on the segmented glenoid surface. Potential points are screened by using the angle relationship between the normal vectors of the key points. The extraction range of the glenoid surface is optimized by combining sphere fitting, and finally the extracted glenoid surface is obtained.
2. The method according to claim 1, characterized in that, The process of acquiring shoulder joint scan images and performing resampling and standardization includes: The shoulder joint scan image was resampled, the z-axis spacing was adjusted to match the x-axis and y-axis spacing, and excess slices in the z-axis direction were cropped to obtain the cropped shoulder joint scan image; then the cropped shoulder joint scan image was standardized using a window width of 1800 HU and a window level of 400 HU.
3. The method according to claim 1, characterized in that, The 3D U-Net network adopts an encoder-decoder architecture. It extracts high-level semantic features through the encoder part and combines the extracted high-level semantic features with the spatial detail information of the decoder part through skip connections, thereby completing image segmentation.
4. The method according to claim 3, characterized in that, Each convolutional block of the 3D U-Net network consists of a 3x3x3 convolutional layer, followed by a batch normalization layer and a ReLU activation layer.
5. The method according to claim 3, characterized in that, Image segmentation based on the 3D U-Net network uses max pooling for downsampling and trilinear interpolation for upsampling.
6. The method according to claim 1, characterized in that, The process of selecting key points on the segmented glenoid surface, filtering potential points using the angle relationship between the normal vectors of the key points, and optimizing the extraction range of the glenoid surface by combining sphere fitting, ultimately obtaining the extracted glenoid surface includes: Two key points are selected from the segmented glenoid surface, and the first angle between the normal vectors of the two key points is obtained; the first point is located near the bottom of the glenoid, and the second point is located inside the glenoid. Obtain the second angle between the normal vector of each point on the surface of the scapula and the normal vector of the first point, and filter out the points that satisfy the condition that the second angle is less than the first angle; Introduce a third point located inside the glenoid cavity, obtain the third included angle between the normal vector of the third point and the first point, and filter out points that satisfy the condition that the third included angle is less than the first included angle; The points selected in the two screenings are fitted into a sphere to obtain the extracted glenoid surface.
7. An automatic glenoid segmentation system for the shoulder joint based on artificial intelligence, characterized in that, The method for implementing any one of claims 1-6 includes: an image preprocessing module, a segmentation model construction module, a glenoid surface segmentation module, and a glenoid surface extraction module; The image preprocessing module is used to acquire shoulder joint scan images and perform resampling and standardization processing. The segmentation model construction module is used to construct a glenoid segmentation model based on a 3D U-Net network, and to train the glenoid segmentation model based on a standardized shoulder joint scan image. The glenoid surface segmentation module is used to segment the shoulder joint scan image to be processed based on the trained shoulder glenoid segmentation model. The glenoid surface extraction module is used to select key points on the segmented glenoid surface, filter potential points by using the angle relationship between the normal vectors of the key points, and optimize the extraction range of the glenoid surface by combining sphere fitting, and finally obtain the extracted glenoid surface.
8. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.
10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.