Joint landmark point identification method, computer equipment and computer program product

Abstract

The embodiment of the invention is suitable for the technical field of computer-aided medical treatment, and provides a joint landmark point identification method, computer equipment and a computer program product, and the method comprises the steps: obtaining medical image data with a first resolution; determining an initial position of each joint landmark point in the medical image data; according to the initial position, carrying out registration on the initial landmark point set and a statistical shape model to obtain a target registration matrix; under the condition that missing landmark points exist in the medical image data, complementing the missing landmark points by using the target registration matrix; and projecting each complemented joint landmark point to a joint model surface corresponding to the medical image data, and outputting joint landmark point identification result data with a second resolution. By adopting the method, the joint landmark point can be efficiently and accurately identified, and the accuracy and reliability of medical image data analysis and processing are improved.

Description

Technical Field

[0001] The embodiments of the present application belong to the field of computer-aided medical technology, and particularly relate to a method for identifying joint landmark points, a computer device, and a computer program product. Background Art

[0002] The automatic recognition of human joint landmark points has important applications in fields such as medical image analysis, orthopedic surgery planning, and auxiliary diagnosis. However, existing recognition methods face problems such as insufficient accuracy, stability, and recognition rate. These problems comprehensively lead to a low success rate of one-time recognition of landmark points, seriously affecting the accuracy and efficiency of subsequent medical analysis and diagnosis. Especially in complex cases, the limitations of existing landmark point recognition technologies are more prominent and cannot meet the needs of precision medicine. Summary of the Invention

[0003] In view of this, the embodiments of the present application provide a method for identifying joint landmark points, a computer device, and a computer program product to improve the efficiency and accuracy of identifying human joint landmark points.

[0004] The first aspect of the embodiments of the present application provides a method for identifying joint landmark points, including: Obtaining medical image data with a first resolution, where the first resolution is less than the original resolution of the medical image data; Determining the initial positions of each joint landmark point in the medical image data; Performing registration on an initial landmark point set composed of each joint landmark point and a pre-constructed statistical shape model according to the initial positions to obtain a target registration matrix, where the statistical shape model is constructed based on a standard data set, and the standard data set is a data set pre-labeled with joint landmark points; When there are missing landmark points in the medical image data, using the target registration matrix to complete the missing landmark points; Projecting each completed joint landmark point onto the surface of the joint model corresponding to the medical image data, and outputting joint landmark point recognition result data with a second resolution, where the second resolution is greater than the first resolution.

[0005] The second aspect of the embodiments of the present application provides an apparatus for identifying joint landmark points, including: An obtaining module, configured to obtain medical image data with a first resolution, where the first resolution is less than the original resolution of the medical image data; A determining module, configured to determine the initial positions of each joint landmark point in the medical image data; A registration module, configured to register an initial landmark set formed by each of the joint landmarks with a pre-constructed statistical shape model according to the initial positions, to obtain a target registration matrix, where the statistical shape model is constructed based on a standard data set, and the standard data set is a data set pre-labeled with joint landmarks; A filling module, configured to fill in the missing landmarks in the medical image data by using the target registration matrix when there are missing landmarks in the medical image data; A projection module, configured to project each of the filled joint landmarks onto the surface of the joint model corresponding to the medical image data; An output module, configured to output joint landmark recognition result data with a second resolution, where the second resolution is greater than the first resolution.

[0006] A third aspect of the embodiments of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the computer device implements the method described in the first aspect above.

[0007] A fourth aspect of the embodiments of the present application provides a computer-readable storage medium, in which a computer program is stored. When the computer program is executed by a computer, the method described in the first aspect above is implemented.

[0008] A fifth aspect of the embodiments of the present application provides a computer program product, including a computer program. When the computer program runs, the method described in the first aspect above is executed.

[0009] Compared with the prior art, the embodiments of the present application have the following beneficial effects: In the embodiments of the present application, by processing medical image data into data with a lower first resolution, a computer device can use a deep neural network to determine the initial positions of each joint point. On this basis, by registering an initial landmark set formed by each joint landmark with a pre-constructed statistical shape model, a target registration matrix can be obtained, so that the missing landmarks can be filled in by using the target registration matrix. After projecting each of the filled joint landmarks onto the surface of the joint model, the computer device can output joint landmark recognition result data with a higher second resolution to complete the recognition of all landmarks. By applying the recognition method provided in the embodiments of the present application, the joint landmarks in the medical image data can be recognized efficiently and accurately, improving the accuracy and reliability of medical image data processing, and helping to provide a more solid foundation for the formulation of personalized treatment plans and the optimization of surgical planning.

[0010] The joint landmark point recognition method provided in the embodiment of the present application has broad application prospects in the medical field, and can play a key role in orthopedic clinical practice in particular. For example, in terms of surgical planning, the precise landmark point recognition technology provided in the embodiment of the present application can provide strong support for complex operations such as joint replacements. By accurately locating key anatomical structures, surgeons can develop more personalized and precise surgical plans, significantly improving the accuracy and success rate of registration during surgery.

[0011] In addition, the joint landmark point recognition method provided in the embodiment of the present application can also play an important role in the field of auxiliary diagnosis, especially in the early diagnosis and disease assessment of bone and joint diseases such as arthritis and osteoporosis. The application of this method can provide key reference points for the automated analysis of images such as CT and MRI, greatly improving the efficiency and accuracy of image interpretation. This is because accurate landmark point positioning helps the diagnostic system provide more standard data postures and observation surfaces, allowing doctors to more accurately analyze the patient's condition, thereby implementing timely intervention treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or prior art descriptions. Obviously, the drawings described below are only some embodiments of the present application, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0013] Figure 1 is a schematic diagram of a method for identifying joint landmark points provided in an embodiment of the present application; Figure 2 It is a schematic diagram of a deep neural network processing flow provided in an embodiment of the present application; Figure 3 It is a schematic diagram of a process of segmenting medical image data using a deep neural network provided in an embodiment of the present application; Figure 4 It is a schematic diagram of a process for constructing a shape statistical model provided in an embodiment of the present application; Figure 5 It is a flowchart of landmark point registration and gap filling provided by an embodiment of the present application; Figure 6 It is a schematic diagram of a joint landmark point projection process provided by an embodiment of the present application; Figure 7 is a schematic diagram of a joint landmark point optimization process provided in an embodiment of the present application; Figure 8 is a schematic diagram of a joint landmark point recognition process provided by an embodiment of the present application; Figure 9It is a schematic diagram of an apparatus for identifying joint landmark points provided by an embodiment of the present application; Figure 10 It is a schematic diagram of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0014] In the following description, specific details such as specific system architectures and technologies are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present application.

[0015] As mentioned above, there are various problems in joint landmark point recognition in the related art. They mainly include: 1. Difficulty in recognition caused by anatomical structure variations: Due to anatomical structure abnormalities caused by pathological features such as pelvic dysplasia and joint fusion, traditional recognition methods based on fixed templates and heatmap regression methods based on simple deep learning are difficult to adapt to clinical data, resulting in a significant decrease in recognition accuracy.

[0016] 2. Errors caused by inconsistent data acquisition postures: Patients are unable to maintain a standard scanning posture for various reasons, resulting in a deviation in the spatial position relationship of the image data from the training data, thereby affecting the performance of the recognition algorithm.

[0017] 3. Image complexity and quality degradation caused by other surgeries: Bone defects and artifacts caused by implants in other surgeries not only change the original bone anatomical structure but also introduce a large amount of metal artifacts, which greatly increases the difficulty of landmark point recognition.

[0018] 4. Insufficient ability to handle missing landmark points: Existing methods are insufficient in dealing with the situation of missing landmark points caused by various reasons and cannot effectively complement the lost key anatomical marker points.

[0019] 5. Trade-off between computational efficiency and accuracy: Improving recognition accuracy usually requires more complex algorithms and higher computational costs, which conflicts with the requirements for real-time performance in clinical practice.

[0020] In view of the above problems, the embodiments of the present application provide an intelligent method for identifying joint landmark points that can overcome data acquisition deviations, complement missing landmark points, and at the same time take into account computational efficiency. Applying this method can not only improve the accuracy and reliability of medical image analysis but also provide a more solid foundation for the formulation of personalized treatment plans and the optimization of surgical planning.

[0021] The technical solutions of the present application will be described below through specific embodiments.

[0022] Reference Figure 1 , which shows a schematic diagram of a method for identifying joint landmark points provided by an embodiment of the present application. Specifically, it may include the following steps: S101. Obtain medical image data with a first resolution.

[0023] This method can be applied to a computer device, that is, the execution subject of this method can be a computer device. By executing each step of the method provided by the embodiment of the present application, the computer device can efficiently and accurately identify the joint landmark points in the medical image data. The above computer device can be a computer-aided medical device, such as a medical device with medical image data processing functions or other types of desktop computers, cloud servers, etc. The embodiment of the present application does not limit the type of the computer device. Unless otherwise specified, the joint landmark points or landmark points mentioned in the embodiment of the present application refer to the same feature, that is, the joint landmark points.

[0024] The medical image data can be three-dimensional image data obtained by scanning a patient using an imaging device. For example, CT data obtained by scanning a patient's joint part or the whole body using a computed tomography (CT) device. Therefore, the above medical image data can be the image data of a patient's partial joint part or the whole body image data of the patient.

[0025] In a possible implementation manner of the embodiment of the present application, the computer device can be communicatively connected to the imaging device. After the imaging device such as a CT device collects the image data of the patient, it can be transmitted to the computer device for joint landmark point recognition. After receiving the original medical image data, the computer device can perform downsampling processing on it, and downsample the medical image data with the original resolution to the data with the first resolution. The above first resolution is less than the original resolution. Exemplarily, assuming that the original resolution of the medical image data collected by the CT device is 1500×1500 pixels, it can be processed into data with a lower first resolution through downsampling, for example, the first resolution is 750×750 pixels. Through downsampling processing, the number of pixels to be recognized by the subsequent computer device and the deep neural network configured in the computer device can be reduced, so that the resolution of the downsampled medical image data is as close as possible to the resolution of the data that the computer device or the deep neural network can actually process, which is convenient for the computer device or the deep neural network to understand the structural relationship of the medical image data, thereby improving the processing efficiency and recognition accuracy.

[0026] S102. Determine the initial positions of the respective joint landmark points in the medical image data.

[0027] The initial position may refer to the approximate positions of the respective joint fiducial points initially determined in the medical image data.

[0028] In a possible implementation manner of the embodiments of the present application, the computer device may use an image processing algorithm to process the medical image data, initially determine the regions where each joint fiducial point may exist in the image, and quickly locate the approximate positions of each fiducial point.

[0029] In another possible implementation manner of the embodiments of the present application, the computer device may use a deep neural network configured therein to process the medical image data with the first resolution, and output the approximate positions of each joint fiducial point, laying a foundation for subsequent fine positioning.

[0030] Specifically, a deep neural network including multiple channels may be configured in the computer device. After obtaining the medical image data, the computer device may use the deep neural network to perform segmentation processing on the image data to obtain one or more connected regions corresponding to each joint fiducial point. The computer device may calculate the initial position of each joint fiducial point based on the one or more connected regions corresponding to each joint fiducial point.

[0031] In the embodiments of the present application, the number of channels of the deep neural network used to process the medical image data may be equal to the number of joint fiducial points to be recognized. In this way, each channel may be respectively used to perform segmentation processing on the data in the region where a joint fiducial point is located. Since the number of joint fiducial points to be recognized in each recognition process is not necessarily exactly the same, the deep neural network may include more channels. In each recognition process, the actually participating channels may be less than or equal to the actually included channels of the deep neural network.

[0032] Such as Figure 2 shown, is a schematic diagram of a deep neural network processing flow provided by the embodiments of the present application. Figure 2 The processing flow shown in may be an example of using a deep neural network to process the medical image data and output the initial positions of each joint fiducial point.

[0033] See Figure 2 , the deep neural network adopted by the embodiments of the present application may be constructed based on the U-Net structure. Figure 2 The original data in may refer to the medical image data with the original resolution, and the compressed data obtained through downsampling and other scaling processes may be the medical image data with a lower first resolution. After the medical image data with the first resolution is input into the deep neural network, the deep neural network may perform segmentation on the data and output the corresponding segmentation result, and the segmentation result may represent the specific region where the fiducial point may exist.

[0034] Combined Figure 3 As shown, it is a schematic diagram of the process of a deep neural network segmenting medical image data provided by an embodiment of the present application. Among them, Figure 3 In part (a), taking the medical image data as the CT image data X of the current case as an example, using the deep neural network to perform segmentation to obtain the CT segmentation result L representing the possible region of the fiducial points. This process is also the label region segmentation process. Among them, w, h, and d respectively represent the size of the input CT image data X.

[0035] On this basis, the computer device can further use the deep neural network to perform semantic segmentation on the above segmentation result L to locate the fiducial regions of each fiducial point, and this fiducial region is also the approximate position of each fiducial point.

[0036] Specifically, after obtaining the segmentation result L according to the segmentation process shown in part (a) of Figure 3 , the computer device can extract the maximum connected component channel by channel, and then perform semantic segmentation on the maximum connected component region according to the segmentation process shown in part (b) of Figure 3 to obtain the fiducial region H. Therefore, Figure 3 the segmentation process shown in part (b) of

[0037] is also the fiducial region segmentation process. Among them, c represents the number of fiducial points. It should be noted that the deep neural network used for further semantic segmentation of the segmentation result L obtained in the previous step can be the same neural network as the deep neural network used for segmenting the CT image data X, or a different neural network.

[0038] Exemplarily, a deep neural network based on the U-Net structure can be trained to have both Figure 3 the segmentation functions shown in part (a) and part (b) of . In this way, the deep neural network and the deep neural network Figure 3 can be the same neural network. The training data for training the function shown in part (a) of the above training Figure 3 is different from the training data for training the function shown in part (a) of

[0039] Or, two deep neural networks can be constructed based on the U-Net structure, namely the deep neural network and the deep neural network , and then they are respectively trained with the corresponding training data, so that the deep neural network hasFigure 3 The function shown in part (a) in has Figure 3 the function shown in part (b) in. By configuring the trained deep neural network and the deep neural network in a computer device, medical image data can be processed during the process of identifying joint landmark points, and the approximate positions of each landmark point can be output.

[0040] In the embodiment of the present application, based on the approximate positions of each joint landmark point output by the deep neural network, the computer device can further perform precise positioning at this position to obtain the initial positions of each joint landmark point.

[0041] Specifically, for one or more connected regions segmented from each channel of the deep neural network, the computer device can respectively determine the largest connected region among the one or more connected regions corresponding to each channel. The above-mentioned largest connected region is also the region where the corresponding joint landmark point is located. On this basis, the computer device can calculate the center point of the largest connected region corresponding to each channel as the initial position of the corresponding joint landmark point.

[0042] It should be noted that the joint landmark points obtained through the segmentation process based on the deep neural network may include all the landmark points to be identified, or may only include some of them. That is, after the medical image data is processed by the deep neural network, there may be a situation where some landmark points are not recognized. Therefore, the initial positions of the output joint landmark points do not include the positions of these unrecognized landmark points.

[0043] S103. According to the initial positions, register the initial landmark point set composed of each of the joint landmark points with a pre-constructed statistical shape model to obtain a target registration matrix.

[0044] Based on the initial positions of each recognized joint landmark point calculated in the previous step, the computer device can register the initial landmark point set with a pre-constructed statistical shape model to obtain a target registration matrix. The above-mentioned initial landmark point set is composed of those recognized joint landmark points.

[0045] In the embodiment of the present application, the statistical shape model can be a model constructed based on a standard data set. The standard data set can be a data set with pre-marked joint landmark points. Exemplarily, the standard data set can be obtained by doctors marking joint landmark points in a three-dimensional model. Based on the relevant data of each manually marked joint landmark point, the computer device can construct a statistical shape model.

[0046] Such as Figure 4As shown, it is a schematic flowchart of a process for constructing a shape statistical model provided by an embodiment of the present application. Figure 4 The landmark dataset T in [[ ]] can be a set of data points composed of multiple landmarks extracted from a standard dataset. For example, the landmark dataset T can include N landmarks. The computer device can process the landmark dataset T to calculate the mean of these N landmarks , eigenvalues and eigenvectors , and through registration within the dataset T, combined with the principal components analysis (PCA) algorithm, a statistical shape model (SSM) is constructed. This SSM model captures the spatial relationships and variation patterns between landmarks and can provide strong prior knowledge for subsequent landmark filling and optimization.

[0047] Based on the pre-constructed statistical shape model SSM, the computer device can register the identified joint landmarks into the model SSM to obtain the target registration matrix.

[0048] As Figure 5 shown, it is a schematic flowchart of a process for landmark registration and filling provided by an embodiment of the present application, Figure 5 showing the entire process of registering each landmark identified by the deep neural network with the statistical shape model SSM to obtain the target registration matrix, and completing the filling of missing landmarks based on the target registration matrix.

[0049] Specifically, referring to Figure 5 , Figure 5 the initial point set shown in [[ ]] Ps is the entire point set to be identified, and its number is c. Therefore, this point set P s = { P s1 ……P sc}. During the process of landmark registration, the computer device can, according to the initial positions of the identified landmarks, perform null value detection on each joint landmark in the medical image data, that is, the initial point set P s to obtain the non-empty landmark index number set I 0. This non-empty landmark index number set I 0 is obtained by sorting each joint landmark identified from the medical image data.

[0050] The computer device can perform operations on the non-empty landmark index number set IPerform registration with the statistical shape model SSM to obtain the initial registration matrix M 0, and calculate the Euclidean distance between the point clouds corresponding to this initial registration matrix M 0. This distance D 0 can be used to compare with the distances corresponding to other registration matrices during the subsequent update of the initial registration matrix D 0. In M , Regi(.,.) represents point cloud registration Figure 5 .

[0051] To update the above initial registration matrix M 0, the computer device can determine multiple sets of optional point numbers based on the non-empty landmark index number set I 0. As Figure 5 shown, the set of optional point numbers can include N types, and these N types of sets of optional point numbers can be composed of arbitrarily selecting K points from the non-empty landmark index number set I 0. That is:

[0052] The above expression is the permutation and combination formula, indicating the combination of selecting K elements from elements. Among them, represents the number of landmark points included in the non-empty landmark index number set I 0. Exemplarily, if = 14 and K = 4, then the number of N types of sets of optional point numbers obtained based on the above expression is: = 1001

[0053] In a possible implementation manner of the embodiments of the present application, any set of optional point numbers may include at least 4 identified joint landmark points, that is, the value of the above K may be a natural number greater than or equal to 4. The number of joint landmark points included in each set of optional point numbers may be equal

[0054] Exemplarily, taking K = 4 as an example, the computer device can select any 4 landmark points from the non-empty landmark index number set I 0 to form a set of optional point numbers. For example, the initial point set P s includes c landmark points P s1 ，……，P sc , assuming that the identified landmark points are P s1 ，P s2 ，P s3 ，Ps4 ，P s5 ，P s7 ，……，P sc , that is, the boundary punctuation point P s6 If it is not recognized by the deep neural network, then the multiple created optional point sequence number sets may include: Optional point sequence number set 1: P s1 ，P s2 ，P s3 ，P s4 ; Optional point sequence number set 2: P s1 ，P s2 ，P s3 ，P s5 ; Optional point sequence number set 3: P s1 ，P s2 ，P s3 ，P s7 ; Optional point sequence number set 4: P s1 ，P s2 ，P s3 ，P s8 ; …… Optional point sequence number set N: P sc-3 ，P sc-2 ，P sc-1 ，P sc ; The computer device can register each optional point sequence number set with the statistical shape model. During this process, as Figure 5 shown, the computer device can obtain the initial sequence number set corresponding to the minimum Euclidean distance between point clouds through N cycles of registration I 0 and the sequence number set to be updated. The above sequence number set to be updated can be a non-empty boundary marker index sequence number set I 0 and the initial sequence number set I 1 difference. For example, the sequence number set to be updated can be expressed as I2= I 0- I 1。

[0055] Specifically, the computer device can register each set of optional point numbers with the statistical shape model SSM and calculate the Euclidean distance between the point clouds under the current registration result. In this way, after completing the registration of all N sets of optional point numbers, N results of the Euclidean distance between the point clouds can be obtained. The computer device can compare these N distances and determine the set of optional point numbers corresponding to the minimum distance as the initial number set. I 1, this initial number set I 1 is also the optimal set of optional point numbers when selecting K landmark points.

[0056] Then, the computer device can incorporate one or more joint landmark points in the sequence set to be updated I 2 into the initial number set I 0, and repeat the operation of registering the number set incorporating other joint landmark points with the statistical shape model SSM to obtain the target index number set I f , this target index number set I f The corresponding registration matrix M f is also the optimal target registration matrix.

[0057] Specifically, the computer device can, according to the index of each joint landmark point in the sequence set to be updated I 2, incorporate the joint landmark point with the smallest index value into the initial number set I 0, and register the number set after incorporating the joint landmark point with the smallest index value with the statistical shape model SSM to obtain the corresponding Euclidean distance between the point clouds.

[0058] Exemplarily, assume that the initial number set I 0 is the set of optional point numbers 1 in the foregoing example, and the landmark points it contains are P s1 ，P s2 ，P s3 and P s4 , since the landmark point not recognized by the deep neural network is the landmark point P s6 , therefore, the joint landmark point with the smallest index value in the sequence set to be updated I 2 is P s5 , and the number set obtained after incorporating this landmark point into the initial number set I 0 contains a total of 5 landmark points, namely Ps1 ，P s2 ，P s3 ，P s4 and P s5 。 The computer device can register the sequence set containing the above 5 boundary points with the statistical shape model SSM to obtain the corresponding Euclidean distance and distance difference between the point clouds. For example, the Euclidean distance between the point clouds can be the distance shown in Figure 5 D t 。 Correspondingly, the distance difference can be the Euclidean distance between the point clouds corresponding to the current sequence set D t and the Euclidean distance corresponding to the initial registration matrix M 0, that is D 0, i.e., D t - D 0.

[0059] The computer device can judge the above distance difference to determine its magnitude relationship with the distance threshold T. As shown in Figure 5 , if the distance difference D t - D 0 is greater than or equal to the distance threshold T, it means that compared with the initial sequence set I 0, after incorporating the joint boundary point with the smallest current index value (i.e., the boundary point in the previous example P s5 ), a better sequence set is not obtained. At this time, the computer device can discard the joint boundary point with the smallest current index value from the sequence set to be updated. If the distance difference D t - D 0 is less than the distance threshold T, it means that compared with the initial sequence set I 0, a better sequence set can be obtained after incorporating the joint boundary point with the smallest current index value. At this time, the computer device can retain the joint boundary point with the smallest current index value (i.e., the boundary point in the previous example P s5 ) in the initial sequence set I 0. Taking the previous example as an example, the sequence set at this time is updated to a set containing P s1 ，P s2 ，P s3 ，P s4 and P s5 a total of 5 boundary points.​

[0060] The computer device can repeatedly execute the foregoing steps of taking the joint boundary point with the smallest index value into the initial sequence set and performing registration according to the index of each joint boundary point in the sequence set to be updated, and finally obtain the target index sequence set I f and the target registration matrix M f .

[0061] Exemplarily, in the foregoing example, after obtaining a set containing a total of P s1 ，P s2 ，P s3 ，P s4 and P s5 a total of 5 boundary points, since the boundary point P s6 is not recognized, the boundary point with the smallest index value contained in the sequence set to be updated I 2 is the boundary point P s7 . After this boundary point is incorporated into the sequence set, a sequence set containing P s1 ，P s2 ，P s3 ，P s4 ，P s5 and P s7 a total of 6 boundary points can be obtained. The computer device can repeatedly execute the registration process, register this sequence set with the statistical shape model SSM, and determine whether to retain or discard this boundary point in the sequence set by comparing the Euclidean distance between point clouds P s7 or discard.

[0062] The above registration operation can be repeated multiple times until the Euclidean distance D t between the point clouds and the Euclidean distance M 0 corresponding to the initial registration matrix D 0 has a distance difference less than the distance threshold T and contains the largest number of boundary points. At this time, the sequence set can be used as the final index sequence I f , and this sequence is the target index sequence set. It can be considered that the boundary points included in the final index sequence I f are all accurately recognized boundary points and are not included in the above final index sequenceI f The boundary punctuation points in it need to be completed in subsequent steps. After obtaining the final index sequence I f correspondingly, the target registration matrix is obtained M f .

[0063] S104. In the case where there are missing boundary punctuation points in the medical image data, use the target registration matrix to complete the missing boundary punctuation points.

[0064] In the embodiment of the present application, if all the boundary punctuation points in the medical image data are not recognized after the segmentation process of the deep neural network, that is, there are missing boundary punctuation points, the computer device can use the previously obtained target registration matrix to complete the missing boundary punctuation points and complete the process of filling in the missing boundary punctuation points.

[0065] As Figure 5 shown, when using the target registration matrix M f to complete the missing boundary punctuation points, the empty boundary marker index number set can be determined first based on the target index number set I f The empty boundary marker index number set I e represents the set of each boundary punctuation point that needs to be filled in. I e In the embodiment of the present application, the empty boundary marker index number set

[0066] represents those boundary punctuation points that were not accurately recognized before performing the boundary punctuation point filling operation compared to all the boundary punctuation points to be recognized. Since the accurately recognized boundary punctuation points are the boundary punctuation points included in the final index sequence I e , the above empty boundary marker index number set I f can be expressed as the difference between the original number set I e and the final index sequence I That is I f = I e = I - I f .

[0067] Then, the computer device can use the target registration matrix M f , and in the statistical shape model SSM, the ones corresponding to the empty boundary marker index number set I eThe corresponding standard fiducial points are transformed into the medical image data to complete the missing fiducial points. Specifically, as Figure 5 shown, this process can be expressed as:

[0068] In the above expression, represents the set of points obtained after completing the filling of the empty fiducial index number set I e After completion, represents the standard fiducial points corresponding to the empty fiducial index number set I e in the statistical shape model SSM. The above expression can obtain the completed set of points by transforming the standard fiducial points using the inverse matrix of the target registration matrix M f .

[0069] After completing the operation of filling the missing fiducial points, a set containing all the fiducial points to be recognized can be obtained, that is, Figure 5 the repaired fiducial point set shown in P f .

[0070] S105. Project each of the completed joint fiducial points onto the surface of the joint model corresponding to the medical image data.

[0071] In the embodiment of the present application, since the position of the missing fiducial points filled in the above is calculated using the target registration matrix, there may be a situation where the calculated position is not on the surface of the joint model corresponding to the medical image data. In response to this situation, the computer device can project each of the completed joint fiducial points, so as to project each joint fiducial point onto the surface of the joint model.

[0072] As Figure 6 shown, it is a schematic diagram of a joint fiducial point projection process provided by the embodiment of the present application. As Figure 6 shown, before projecting the fiducial points, the computer device can determine the surface of the joint model corresponding to the medical image data by performing morphological erosion on the medical image data. Figure 6 In , it represents morphological erosion, that is, eroding the medical image data L, K represents the convolution kernel, and d(.,.) represents calculating the Euclidean distance of spatial points. After morphological erosion, the original medical image data L can be shrunk by one circle, so as to retain the surface area of the model.

[0073] Specifically, as Figure 6 shown, the medical image data L can be eroded first using the method of morphological erosion, and its expression is , the result obtained after erosion is the data after the original medical image data L is reduced by one circle (the surface of the medical image data L is eroded). Then, subtracting the data of from the original medical image data L can obtain the surface area of the model .

[0074] On this basis, the repaired boundary point set obtained after completion P f can have each boundary point projected onto the model surface.

[0075] Specifically, the computer device can, for any joint boundary point after completion p , calculate the distances between each point s on the joint model surface and other joint boundary points except the current joint boundary point , and use the point corresponding to the minimum value in the distances as the projection point of the current joint boundary point on the joint model surface, obtaining a set of projection points, that is, Figure 6 the surface projection landmarks in . Each of the above points s belongs to the points on the joint model surface S, that is, . Figure 6 In the expression, argmin represents the set of parameter values that make the objective function obtain the minimum value, and it is used as a whole to represent obtaining the corresponding points p of each joint boundary point on the joint model surface S. The set composed of the points is the boundary points obtained after projection .

[0076] S106. Output the joint boundary point recognition result data with the second resolution.

[0077] The aforementioned processing of the medical image data is all carried out after downsampling and other scaling of the original image data to the first resolution. The final processed joint boundary points located on the joint model surface are also presented in the medical image data with the first resolution, and its resolution is relatively low. Therefore, after projecting the completed boundary points onto the joint model surface, the computer device can also perform optimization processing on them and output the joint boundary point recognition result data with the second resolution. The above second resolution is greater than the first resolution. For example, the second resolution can be the same as the original resolution of the medical image data.

[0078] As Figure 7 shown, it is a schematic diagram of an optimization process of joint boundary points provided by an embodiment of the present application. According to the Figure 7 shown optimization process, the computer device can use a deep neural network, such as the deep neural network Optimize and fine-tune each joint landmark point on the model surface to obtain the final set of each landmark point The above-mentioned deep neural network can be the same neural network as the aforementioned deep neural network and the deep neural network with the same structure. For example, the deep neural network can also be a neural network constructed based on the U-Net structure.

[0079] In the embodiment of the present application, by processing medical image data into data with a lower first resolution, the computer device can use a deep neural network to determine the initial positions of each joint point. On this basis, by registering the initial landmark point set composed of each joint landmark point with a pre-constructed statistical shape model, a target registration matrix can be obtained, and thus the missing landmark points can be complemented using the target registration matrix. After projecting the complemented joint landmark points onto the surface of the joint model, the computer device can output joint landmark point recognition result data with a higher second resolution to complete the recognition of all landmark points. Applying the recognition method provided by the embodiment of the present application can efficiently and accurately identify the joint landmark points in medical image data, improving the accuracy and reliability of medical image data processing. This method not only improves the recognition accuracy of human joint landmark points under complex pathology and non-standard postures, but also solves the problem of missing landmark points caused by lesions, surgeries, or acquisition postures. It also reduces the spatial misalignment caused by false detections by optimizing the landmark point positioning accuracy, ensuring the accuracy of the recognition result.

[0080] It should be noted that the magnitudes of the sequence numbers of the above steps in the above embodiments do not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0081] For the convenience of understanding, the following introduces a complete example of the joint landmark point recognition method provided by the embodiment of the present application.

[0082] In the embodiment of the present application, the computer-aided recognition of joint landmark points is mainly divided into a model training stage and an inference recognition stage. Among them, in the model training stage, a CT data set and a landmark point cloud data set can be used to train a deep neural network and a shape statistical model. The data volume of the training set can account for 80% of the data set, and a data augmentation method can be used to augment the data set. The optimal model in the test set is taken as the final model. When constructing the shape statistical model, the point clouds in the data set can be registered in the same coordinate system, and the PCA algorithm can be used to calculate the mean, eigenvalues, and eigenvectors of the data set. This process can be as Figure 4 shown.

[0083] In the inference and recognition stage, the computer device can downsample the CT image data with the original resolution in the joint area to 1 / 2 of the original size, and then input it into the neural network for segmentation to obtain the segmented joint labels . By inputting the label data into the neural network , the landmark area can be obtained . This process can be as shown Figure 3 . Using connected component extraction to process the landmark area can obtain the initial positions of each landmark point. After registering and inputting each initially recognized landmark point into the statistical shape model for completion, a set of repaired landmark points can be obtained . This process can be as shown Figure 5 . When using the morphological method to calculate the edge of the label and project the completed landmark points onto the edge of the label , all landmark points can be projected onto the surface of the joint model corresponding to the medical image data. This process can be as shown Figure 6 . Finally, inputting each edge point and the CT with the original resolution of the corresponding block into the neural network can obtain the fine-tuned landmarks . This process can be as shown Figure 7 .

[0084] Based on the above introduction, using computer assistance to identify joint landmark points mainly includes steps such as landmark center extraction based on a deep neural network, point cloud registration based on a greedy algorithm, missing point reconstruction based on a shape statistical model, and surface refinement and high-resolution optimization of landmark points. Combining Figure 8 , it is a schematic diagram of an identification process of joint landmark points provided by an embodiment of the present application. According to the process shown Figure 8 , the specific introduction of each step is as follows: (1) Landmark center extraction based on a deep neural network This step mainly includes low-resolution neural network segmentation and extraction of the initial landmark point set

[0085] 1. Low-resolution neural network segmentation: First, the input low-resolution medical image data can be segmented by a neural network and the maximum connected component is extracted channel by channel. Then, using a deep neural network to perform further semantic segmentation on the segmentation result, relevant information about the possible areas of each landmark point can be obtained. This step can quickly locate the approximate positions of the landmark points and lay a foundation for subsequent fine positioning

[0086] As shown Figure 8As shown, the current case CT image data can be processed low-resolution medical image data. Using a neural network After segmentation, a CT segmentation result can be obtained. This process is the label area segmentation process introduced in the foregoing embodiments. Figure 8 The landmark area segmentation neural network shown in can be a deep neural network After further processing the CT segmentation result using the deep neural network

[0087] The input and output data and processing flow involved in the above steps can be expressed as: - Input: Joint label data

[0088] - Processing: Use a deep neural network for semantic segmentation to obtain the areas where landmark points may exist - Output: Landmark area

[0089] 2. Initial landmark point set extraction: Based on the segmentation result of the previous step, by extracting the largest connected region from the probability map corresponding to each landmark point, its center can be calculated as the initial position of the landmark point.

[0090] Based on the landmark points with the calculated initial positions, an initial landmark point set can be formed P s ={ P s1 ……P sc}.

[0091] The input and output data and processing flow involved in the above steps can be expressed as: - Input: Segmented landmark area

[0092] - Processing: Extract the largest connected region from each segmentation channel and calculate its center as the initial landmark point - Output: Initial landmark point set

[0093] (2) Point cloud registration based on the greedy algorithm This step mainly includes the construction of a statistical shape model (SSM) and data registration optimization.

[0094] 3. Statistical Shape Model (SSM) Construction: In this step, a shape statistical model can be constructed by using the landmark points in the standard dataset through within-dataset registration and principal component analysis methods. This model captures the spatial relationships and variation patterns among the landmark points, providing strong prior knowledge for subsequent landmark point completion and optimization.

[0095] The input and output data and processing flow involved in the above steps can be expressed as follows: - Input: Landmark point annotations in the standard dataset - Processing: Register the data within the dataset to the same space and apply PCA analysis - Output: SSM model 4. Data Registration Optimization: In this method, an innovative iterative registration strategy is adopted when registering the initial set of landmark points P s with the SSM model. K (K >= 4) points are selected from the optional point set for point-to-point registration. After repeating multiple times, the optimal registration matrix, i.e., the target registration matrix introduced in the foregoing embodiments, can be obtained through screening. M f This step effectively reduces the impact of misdetected landmark points on the overall registration and improves the stability of the system.

[0096] The input and output data and processing flow involved in the above steps can be expressed as follows: - Input: Initial set of landmark points, SSM model - Processing: a) Traverse the optional point set to complete the registration b) Use the greedy strategy to screen the optimal registration matrix c) Iteratively incorporate other points and distinguish successful points from misdetected points - Output: Optimized registration matrix (III) Missing Point Reconstruction Based on the Shape Statistical Model This step mainly includes using the SSM to complete the missing landmark points.

[0097] 5. SSM Completion of Missing Landmark Points: The optimized registration matrix, i.e., the target registration matrix M f can be used to transform the current set of landmark points into the SSM mean space, and then the SSM model is used to estimate and complete the missing landmark points. This step fills in the missing landmark points and ensures the complete output of the landmark points.

[0098] The input and output data and processing flow involved in the above steps can be expressed as follows: - Input: Optimized registration matrix, SSM model - Processing: Transform the standard points through the registration matrix and use the SSM to complete the missing points - Output: The completed landmark point set (4) and surface refinement and high-resolution optimization of the landmarks This step mainly includes surface projection correction and high-resolution refinement.

[0099] 6. Surface projection correction: By projecting the completed landmark point set onto the surface of the segmented 3D data, it can be ensured that all landmark points are located on the actual surface of the anatomical structure, such as the joint surface. This step corrects the deviation that may be brought about by the SSM completion and improves the anatomical accuracy of the landmark points.

[0100] The input and output data and processing flow involved in the above steps can be expressed as: - Input: The completed landmark point set, joint label data

[0101] - Processing: Project the completed points onto the nearest model surface - Output: The roughly corrected landmark point set 7. High-resolution refinement: Finally, input the medical image data and landmarks at the original resolution into the trained deep neural network , and local optimization can be performed on each landmark point. The deep neural network can perform precise positioning only within a small range around the landmark point, ensuring the image quality of the output landmark recognition result and further improving the accuracy of landmark positioning.

[0102] The input and output data and processing flow involved in the above steps can be expressed as: - Input: The roughly corrected landmark point set, the original high-resolution image - Processing: Optimize each landmark point using the neural network trained based on the original resolution dataset - Output: The finally refined landmark point set The method for identifying joint landmarks provided by the embodiments of the present application fully considers the balance between efficiency and accuracy. This method quickly locates the approximate position of the landmark points from the low-resolution preliminary screening, uses the shape statistical model to supplement prior knowledge and process the missing landmark points, and finally uses high-resolution refinement to ensure the high accuracy of the final result. Each step provides an optimized basis for the next step, forming a closed-loop and self-improving recognition system. Applying this multi-stage and adaptive landmark recognition method provided by the embodiments of the present application can effectively handle the landmark recognition problem in complex pathological conditions and provide high-accuracy and high-reliability landmark recognition results.

[0103] Refer toFigure 9 , which shows a schematic diagram of a device for identifying joint fiducial points provided by an embodiment of the present application. Specifically, it may include an acquisition module 901, a determination module 902, a registration module 903, a filling module 904, a projection module 905, and an output module 906, where: The acquisition module 901 is configured to acquire medical image data with a first resolution, where the first resolution is less than the original resolution of the medical image data; The determination module 902 is configured to determine the initial positions of each joint fiducial point in the medical image data; The registration module 903 is configured to register an initial fiducial point set formed by each joint fiducial point with a pre-constructed statistical shape model according to the initial positions to obtain a target registration matrix. The statistical shape model is constructed based on a standard data set, and the standard data set is a data set pre-labeled with joint fiducial points; The filling module 904 is configured to, when there are missing fiducial points in the medical image data, use the target registration matrix to complete the missing fiducial points; The projection module 905 is configured to project each of the completed joint fiducial points onto the surface of the joint model corresponding to the medical image data; The output module 906 is configured to output joint fiducial point recognition result data with a second resolution, where the second resolution is greater than the first resolution.

[0104] In a possible implementation manner of the embodiment of the present application, the determination module 902 may specifically be configured to: Segment the medical image data by using a deep neural network with multiple channels to obtain one or more connected regions corresponding to each joint fiducial point. The number of channels of the deep neural network is equal to the number of joint fiducial points to be recognized, and any one of the channels is used to segment the data in the region where one joint fiducial point is located; Calculate the initial positions of each joint fiducial point based on one or more connected regions corresponding to each joint fiducial point.

[0105] In the embodiment of the present application, the determination module 902 may also be configured to: For one or more connected regions segmented by each channel of the deep neural network, respectively determine the largest connected region among one or more connected regions corresponding to each channel. The largest connected region is the region where the corresponding joint fiducial point is located; Calculate the center point of the largest connected region corresponding to each channel as the initial position of the corresponding joint fiducial point.

[0106] In a possible implementation of the embodiment of the present application, the registration module 903 may be specifically used to: According to the initial position, performing null value detection on each of the joint landmark points in the medical image data to obtain a non-null landmark index sequence number set, wherein the non-null landmark index sequence number set is obtained by sorting each of the joint landmark points identified from the medical image data; Determine a plurality of optional point sequence number sets based on the non-empty landmark index sequence number set; By aligning each of the optional point sequence number sets with the statistical shape model, an initial sequence number set and a sequence number set to be updated are obtained, wherein the sequence number set to be updated is the difference between the non-empty landmark index sequence number set and the initial sequence number set; After incorporating one or more joint landmark points in the to-be-updated sequence number set into the initial sequence number set, repeatedly aligning the sequence number set incorporating other joint landmark points with the statistical shape model to obtain a target index sequence number set, and the alignment matrix corresponding to the target index sequence number set is the target alignment matrix.

[0107] In the embodiment of the present application, the registration module 903 may also be used for: According to the index of each joint landmark point in the to-be-updated sequence number set, the joint landmark point with the smallest index value is included in the initial sequence number set, and the sequence number set after the joint landmark point with the smallest index value is included is aligned with the statistical shape model to obtain the corresponding Euclidean distance and distance difference between point clouds; If the distance difference is greater than or equal to the distance threshold, the joint landmark point with the smallest current index value is discarded from the set of sequence numbers to be updated; If the distance difference is less than the distance threshold, retaining the joint landmark point with the smallest current index value in the initial sequence number set; Repeat the steps of including the joint landmark point with the smallest index value into the initial sequence number set and performing registration according to the index of each joint landmark point in the sequence number set to be updated, to obtain a target index sequence number set.

[0108] In a possible implementation manner of the embodiment of the present application, any of the optional point number sets includes at least 4 identified joint landmark points, and the number of the joint landmark points included in each of the optional point number sets is equal.

[0109] In a possible implementation of the embodiment of the present application, the gap filling module 904 may be specifically used to: Determine an empty landmark index number set based on the target index number set; Using the target registration matrix, each standard landmark point corresponding to the empty landmark index number set in the statistical shape model is transformed into the medical image data to complete the missing landmark points.

[0110] In a possible implementation manner of the embodiment of the present application, the projection module 905 may specifically be used for: Determine the surface of the joint model corresponding to the medical image data by performing morphological erosion on the medical image data; For any one of the completed joint landmark points, calculate the distances between each point on the surface of the joint model and other joint landmark points except the current joint landmark point respectively; Take the point corresponding to the minimum distance as the projection point of the current joint landmark point on the surface of the joint model.

[0111] An identification device for joint landmark points provided by an embodiment of the present application may be the computer device introduced in the foregoing method embodiment, or a processing unit or module capable of implementing corresponding functions in the computer device. By applying this device, each step in the foregoing method embodiments can be implemented, and joint landmark points can be identified efficiently and accurately, improving the accuracy and reliability of medical image data analysis and processing.

[0112] For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For related parts, refer to the description in the method embodiment part.

[0113] Refer to Figure 10 , which shows a schematic diagram of a computer device provided by an embodiment of the present application. As Figure 10 shown, the computer device 1000 in the embodiment of the present application includes: a processor 1010, a memory 1020, and a computer program 1021 stored in the memory 1020 and executable on the processor 1010. When the processor 1010 executes the computer program 1021, the steps in each embodiment of the above-mentioned method for identifying joint landmark points are implemented, such as Figure 1 the steps S101 to S106 shown. Alternatively, when the processor 1010 executes the computer program 1021, the functions of each module / unit in the above-mentioned device embodiments are implemented, such as Figure 9 the functions of the modules 901 to 906 shown.

[0114] Exemplarily, the computer program 1021 can be divided into one or more modules / units, which are stored in the memory 1020 and executed by the processor 1010 to complete this application. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and these instruction segments can be used to describe the execution process of the computer program 1021 in the computer device 1000. For example, the computer program 1021 can be divided into an acquisition module, a determination module, a registration module, a filling module, a projection module, and an output module. The specific functions of each module are as follows: The acquisition module is configured to acquire medical image data with a first resolution, where the first resolution is less than the original resolution of the medical image data; The determination module is configured to determine the initial positions of respective joint landmark points in the medical image data; The registration module is configured to register an initial landmark point set formed by the respective joint landmark points with a pre-constructed statistical shape model according to the initial positions to obtain a target registration matrix. The statistical shape model is constructed based on a standard data set, and the standard data set is a data set pre-labeled with joint landmark points; The filling module is configured to, when there are missing landmark points in the medical image data, fill in the missing landmark points by using the target registration matrix; The projection module is configured to project the filled respective joint landmark points onto the surface of the joint model corresponding to the medical image data; The output module is configured to output joint landmark point recognition result data with a second resolution, where the second resolution is greater than the first resolution.

[0115] The computer device 1000 can be an electronic device capable of implementing each step or related function in the foregoing various method embodiments. The computer device 1000 can be a desktop computer, a cloud server, or other devices. Exemplarily, the computer device 1000 can be a computer-aided medical device. The computer device 1000 may include, but is not limited to, a processor 1010 and a memory 1020. Those skilled in the art can understand that Figure 10 This is only an example of the computer device 1000 and does not constitute a limitation on the computer device 1000. It may include more or fewer components than shown in the figure, or combine certain components, or have different components. For example, the computer device 1000 may further include input / output devices, network access devices, a bus, etc.

[0116] The processor 1010 may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0117] The memory 1020 may be an internal storage unit of the computer device 1000, such as the hard disk or memory of the computer device 1000. The memory 1020 may also be an external storage device of the computer device 1000, such as a plug-in hard disk equipped on the computer device 1000, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Further, the memory 1020 may also include both the internal storage unit and the external storage device of the computer device 1000. The memory 1020 is used to store the computer program 1021 and other programs and data required by the computer device 1000. The memory 1020 may also be used to temporarily store data that has been output or is to be output.

[0118] An embodiment of this application also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the methods described in the foregoing various embodiments are implemented.

[0119] An embodiment of this application also discloses a computer-readable storage medium storing a computer program, and when the computer program is executed by a computer, the methods described in the foregoing various embodiments are implemented.

[0120] An embodiment of this application also discloses a computer program product including a computer program, and when the computer program runs on a computer, the computer is caused to execute the methods described in the foregoing various embodiments.

[0121] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limiting them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements 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 the present application, and should all be included within the protection scope of the present application.

Claims

1. A method for identifying joint landmark points, characterized in that, Comprising: Obtain medical image data with a first resolution, where the first resolution is less than the original resolution of the medical image data; Determine the initial positions of respective joint landmark points in the medical image data; According to the initial positions, register an initial landmark point set formed by the respective joint landmark points with a pre-constructed statistical shape model to obtain a target registration matrix, where the statistical shape model is constructed based on a standard data set, and the standard data set is a data set pre-labeled with joint landmark points; In the case where there are missing landmark points in the medical image data, use the target registration matrix to complete the missing landmark points; Project the completed respective joint landmark points onto the surface of a joint model corresponding to the medical image data, and output joint landmark point recognition result data with a second resolution, where the second resolution is greater than the first resolution.

2. The method according to claim 1, characterized in that The determining the initial positions of respective joint landmark points in the medical image data includes: Use a deep neural network with multiple channels to segment the medical image data to obtain one or more connected regions corresponding to the respective joint landmark points, where the number of channels of the deep neural network is equal to the number of joint landmark points to be recognized, and any one channel is used to segment data in a region where one joint landmark point is located; Based on the one or more connected regions corresponding to the respective joint landmark points, calculate the initial positions of the respective joint landmark points.

3. The method according to claim 2, wherein The calculating the initial positions of the respective joint landmark points based on the one or more connected regions corresponding to the respective joint landmark points includes: For each of the one or more connected regions segmented by each channel of the deep neural network, respectively determine the largest connected region among the one or more connected regions corresponding to each channel, and the largest connected region is the region where the corresponding joint landmark point is located; Calculate the center point of the largest connected region corresponding to each channel as the initial position of the corresponding joint landmark point.

4. The method according to any one of claims 1 to 3, characterized in that The registering the initial landmark point set formed by the respective joint landmark points with the pre-constructed statistical shape model according to the initial positions to obtain a target registration matrix includes: According to the initial positions, perform null value detection on the respective joint landmark points in the medical image data to obtain a non-empty landmark index number set, and the non-empty landmark index number set is obtained by sorting the respective joint landmark points already recognized from the medical image data; Determine a plurality of optional point number sets based on the non-empty landmark index number set; By registering each of the optional point number sets with the statistical shape model, obtain an initial number set and a set of numbers to be updated, where the set of numbers to be updated is the difference between the non-empty landmark index number set and the initial number set; After incorporating one or more joint landmark points in the set of numbers to be updated into the initial number set, repeat registering the number set incorporating other joint landmark points with the statistical shape model to obtain a target index number set, and the registration matrix corresponding to the target index number set is the target registration matrix.

5. The method according to claim 4, wherein After incorporating one or more joint boundary punctuation points in the to-be-updated sequence number set into the initial sequence number set, repeatedly registering the sequence number set incorporating other joint boundary punctuation points with the statistical shape model to obtain a target index sequence number set, including: According to the index of each joint boundary punctuation point in the to-be-updated sequence number set, incorporate the joint boundary punctuation point with the smallest index value into the initial sequence number set, and register the sequence number set after incorporating the joint boundary punctuation point with the smallest index value with the statistical shape model to obtain the corresponding Euclidean distance and distance difference between point clouds; If the distance difference is greater than or equal to the distance threshold, discard the joint boundary punctuation point with the smallest current index value from the to-be-updated sequence number set; If the distance difference is less than the distance threshold, retain the joint boundary punctuation point with the smallest current index value in the initial sequence number set; Repeatedly execute the step of incorporating the joint boundary punctuation point with the smallest index value into the initial sequence number set according to the index of each joint boundary punctuation point in the to-be-updated sequence number set and performing registration to obtain a target index sequence number set.

6. The method according to claim 4, characterized in that, Any one of the optional point sequence number sets includes at least 4 identified joint boundary punctuation points, and the number of joint boundary punctuation points included in each optional point sequence number set is equal.

7. The method according to claim 4, wherein The complementing the missing boundary punctuation points by using the target registration matrix includes: Determine an empty landmark index sequence number set based on the target index sequence number set; Use the target registration matrix to transform each standard landmark point corresponding to the empty landmark index sequence number set in the statistical shape model into the medical image data to complement the missing boundary punctuation points.

8. The method according to any one of claims 1 to 3 or 5 to 7, characterized in that The projecting each of the complemented joint boundary punctuation points onto the surface of the joint model corresponding to the medical image data includes: Determine the surface of the joint model corresponding to the medical image data by performing morphological erosion on the medical image data; For any one of the complemented joint boundary punctuation points, calculate the distances between each point on the surface of the joint model and other joint boundary punctuation points except the current joint boundary punctuation point respectively; Take the point corresponding to the minimum distance as the projection point of the current joint boundary punctuation point on the surface of the joint model.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, the computer device implements the method according to any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program runs, the method according to any one of claims 1 to 8 is executed.

Images