Hip Joint Anatomical Site Prediction System, Preoperative and Postoperative Evaluation System, and Storage Medium

Abstract

The present invention provides a hip joint anatomical site prediction system, a pre-operative and post-operative evaluation system, and a storage medium, belonging to the technical field of medical data processing. The system of the present invention includes: an input module for inputting the hip joint image data to be recognized of a patient; a data processing module for inputting the hip joint image data to be recognized into a deep learning network model to obtain the anatomical sites and anatomical regions in the hip joint image data to be recognized; an output module for outputting the recognized anatomical sites and anatomical regions, calculating parameters according to the anatomical sites and anatomical regions, and outputting a predicted pre-operative plan or post-operative evaluation based on anatomical information. The system of the present invention has high accuracy and robustness, not only improves the surgical efficiency of clinical hip dysplasia, but also promotes the digitization of the pre-operative planning and post-operative evaluation systems for total hip arthroplasty, reduces the burden on patients, is conducive to standardizing parameters, and has broad application prospects.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical data processing, and particularly relates to a hip joint anatomical site prediction system, a preoperative and postoperative evaluation system, and a storage medium. Background Art

[0002] Total hip arthroplasty (THA) is a surgical technique that replaces the diseased femoral head and acetabulum of the human body by implanting an artificial total hip joint prosthesis, and is mainly used to treat end-stage diseases of the hip joint, such as hip arthritis, hip joint ankylosis, femoral head necrosis, etc. This surgery can significantly relieve joint pain, improve joint function, correct deformities, and has the advantage of a low recurrence rate.

[0003] For the preoperative planning method of total hip arthroplasty, it is mainly based on the anteroposterior X-ray film of the pelvis in two dimensions. Among them, the reconstruction planning of the acetabular side is extremely important in actual clinical work. Currently, the two-dimensional preoperative planning in clinical practice mainly relies on manual measurement of key parameters on the pelvic plane by doctors and template measurement of prostheses. Therefore, there are problems such as cumbersome work, low efficiency, and the inability to uniformly store measurement information in a digital form, which seriously reduces the efficiency of preoperative planning and hinders the digital development of preoperative planning. Currently, the artificial intelligence-based segmentation and key information extraction algorithms have relatively good accuracy on the pelvic radiographs with relatively normal anatomical structures, so that the placement of the acetabular prosthesis and acetabular reconstruction can be more accurately achieved. The existing technology can quickly and accurately identify anatomical sites based on CT and match the required prosthesis models. However, there are the following problems: (1) The time, cost, and radiation dose received by patients for taking CT are significantly higher than those of X-ray films. (2) For Developmental Dysplasia of the Hip (DDH), due to the abnormal development of the acetabulum in anatomical structure (shallow acetabular development, acetabular upward displacement, and joint dislocation), and even severe bone defects, there is no special calculation module. Therefore, more human intervention is required for the reconstruction of complex acetabular anatomical forms, and the boundary values in the vertical direction (acetabular upward / downward displacement) and the inner and outer directions (acetabular inward and outward displacement) of the safe range of acetabular reconstruction and the acetabular coverage rate are not synchronously and real-time displayed during the intervention process.

[0004] In addition, the evaluation of the postoperative prosthesis placement position is also a key link to ensure the success of the surgery and long-term effects. Therefore, it is of great significance to develop a system that meets the preoperative planning of total hip arthroplasty and the postoperative prosthesis placement position through artificial intelligence technology. However, the processing accuracy of existing artificial intelligence models for hip joint medical images still needs to be improved. How to improve the artificial intelligence model to make it more suitable for the processing of hip joint medical images is still an urgent problem to be solved in this field. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a hip joint anatomical site prediction system, a pre- and post-operative evaluation system and a storage medium.

[0006] The present invention provides a hip joint anatomical site prediction system, including:

[0007] An input module, configured to input the hip joint image data to be recognized of a patient;

[0008] A data processing module, configured to input the hip joint image data to be recognized into a deep learning network model to recognize anatomical sites and anatomical regions;

[0009] An output module, configured to output the recognized anatomical sites and anatomical regions;

[0010] Wherein, the deep learning network model is composed of a region segmentation model and a key point localization model Res-SwinFusion; the key point localization model Res-SwinFusion is a model constructed by combining a U-shaped architecture with a Swintransformer feature extraction module and a ResNet backbone network.

[0011] Furthermore, the hip joint image data to be recognized is obtained by data augmentation and data conversion of medical digital imaging and communication data, and the hip joint image data to be recognized is a pelvic anteroposterior X-ray film.

[0012] Furthermore, the region segmentation model is an improved nnU-NetV2 model obtained by removing the random flipping part in the preprocessing of nnU-NetV2;

[0013] The key point localization model Res-SwinFusion includes a feature fusion structure, and the feature fusion structure includes an FIA module.

[0014] Furthermore, the Swin transformer feature extraction module in the key point localization model Res-SwinFusion includes a DFG module.

[0015] Furthermore, the key point localization model Res-SwinFusion includes the following structure:

[0016] Set up an input layer, with a linear embedding layer, a first convolutional layer, and a second convolutional layer connected to the backend of the input layer respectively. The backend of the linear embedding layer is sequentially connected to a first Swin-transformer feature extraction module, a second Swin-transformer feature extraction module, a third Swin-transformer feature extraction module, and a fourth Swin-transformer feature extraction module. The backend of the first convolutional layer is sequentially connected to a first Resnet feature extraction module, a second Resnet feature extraction module, a third Resnet feature extraction module, and a fourth Resnet feature extraction module;

[0017] The feature maps output by the first Swin-transformer feature extraction module and the first Resnet feature extraction module are fused through a first feature interaction aggregation module. The feature maps output by the second Swin-transformer feature extraction module and the second Resnet feature extraction module are fused through a second feature interaction aggregation module. The feature maps output by the third Swin-transformer feature extraction module and the third Resnet feature extraction module are fused through a third feature interaction aggregation module. The feature maps output by the fourth Swin-transformer feature extraction module and the fourth Resnet feature extraction module are fused through a fourth feature interaction aggregation module;

[0018] The backend of the fourth feature interaction aggregation module is connected to a first feature connection module through upsampling. The third feature interaction aggregation module is connected to the first feature connection module through a skip connection. The backend of the first feature connection module is connected to a second feature connection module through upsampling. The second feature interaction aggregation module is connected to the second feature connection module through a skip connection. The backend of the second feature connection module is connected to a third feature connection module through upsampling. The first feature interaction aggregation module is connected to the third feature connection module through a skip connection. The backend of the third feature connection module is connected to a fourth feature connection module through upsampling. The first convolutional layer is connected to the fourth feature connection module through a skip connection. The backend of the fourth feature connection module is connected to a fifth feature connection module through upsampling. The second convolutional layer is connected to the fifth feature connection module through a skip connection;

[0019] The backend of the fifth feature connection module is connected to a third convolutional layer.

[0020] Further, the key point localization model Res-SwinFusion includes the following structure:

[0021] An input layer, configured to input with a size of image data, representing the number of channels of the image, respectively representing the height and width of the image;

[0022] A linear embedding layer, configured to implement segmentation and linear mapping of the input image;

[0023] A first convolutional layer, configured to output image data of size as a feature map of size through a convolutional operation;

[0024] A second convolutional layer, configured to perform a convolutional operation on image data of input size ;

[0025] A Swin-transformer feature extraction module, configured that each Swin-transformer feature extraction module includes a window-based multi-head self-attention module, a shifted window multi-head self-attention module, and a DFG module. The window-based multi-head self-attention module and the shifted window multi-head self-attention module are connected, and the DFG module is connected before the window-based multi-head self-attention module and after the shifted window multi-head self-attention module; wherein, the first Swin-transformer feature extraction module is configured to output the segmented and linearly mapped image data output by the linear embedding layer as a feature map through 2 feature extraction modules; the second Swin-transformer feature extraction module is configured to first perform downsampling and hierarchical feature construction, and then output a feature map of size as a feature map of size through 2 feature extraction modules; the third Swin-transformer feature extraction module is configured to first perform downsampling and hierarchical feature construction, and then output a feature map of size as a feature map of size through 18 feature extraction modules; the fourth Swin-transformer feature extraction module is configured to first perform downsampling and hierarchical feature construction, and then output a feature map of size as a feature map of size through 2 feature extraction modules;

[0026] A first Resnet feature extraction module, configured to output the feature map of size output by the first convolutional layer as a feature map of size after passing through 1 residual block of Resnet50; the second Resnet feature extraction module is configured to output the feature map of size The feature map outputs a feature map with a size of after passing through 1 residual block of Resnet50; The third Resnet feature extraction module is configured to The feature map outputs a feature map with a size of after passing through 1 residual block of Resnet50; The fourth Resnet feature extraction module is configured to The feature map outputs a feature map with a size of after passing through 1 residual block of Resnet50;

[0027] The first feature interaction aggregation module is configured to fuse the feature maps output by the first Swin-transformer feature extraction module and the first Resnet feature extraction module and then output a feature fusion map with a size of ; The second feature interaction aggregation module is configured to fuse the feature maps output by the second Swin-transformer feature extraction module and the second Resnet feature extraction module and then output a feature fusion map with a size of ; The third feature interaction aggregation module is configured to fuse the feature maps output by the third Swin-transformer feature extraction module and the third Resnet feature extraction module and then output a feature fusion map with a size of ; The fourth feature interaction aggregation module is configured to fuse the feature maps output by the fourth Swin-transformer feature extraction module and the fourth Resnet feature extraction module and then output a feature fusion map with a size of ;

[0028] The first feature connection module is configured to output the feature maps output by the fourth feature interaction aggregation module and the third feature interaction aggregation module as a feature map with a size of ; The second feature connection module is configured to output the feature maps output by the first feature connection module and the second feature interaction aggregation module as a feature map with a size of ; The third feature connection module is configured to output the feature maps output by the second feature connection module and the first feature interaction aggregation module as a feature map with a size of ; The fourth feature connection module is configured to output the feature maps output by the third feature connection module and the first convolutional layer as a feature map with a size of ; The fifth feature connection module is configured to output the feature maps output by the fourth feature connection module and the second convolutional layer as a feature map with a size of ;

[0029] The third convolutional layer is configured to perform a convolutional operation on the output of the fifth feature connection module to obtain a key point heat map with a size .

[0030] Furthermore, the DFG module passes the shallow representation to the deep layer by connecting the features before and after calculating the attention in each Swin Transformer block, and constructs additional spatial information on the residual edge to further assist in global information encoding. The algorithm of the DFG module is carried out through the following operations:

[0031] For different stages , ; the input of the Swin-transformer feature extraction module is reshaped into , where represents the height of the feature map in the nth stage, ; represents the width of the feature map in the nth stage, ; represents the height of the original input image, represents the width of the original input image; then an atrous convolutional layer is applied to compress the number of channels to , obtaining a feature , and the expression of the feature is:

[0032]

[0033] where is a reshaping function for matching the size;

[0034] Next, is subjected to global average pooling operation and then reshaped to obtain a feature , and the expression of the feature is:

[0035]

[0036] The reshaped feature of is subjected to a tensor product with the feature , and then the result obtained by normalization , and the expression of the result is:

[0037]

[0038] Finally, and explicitly modeled by the sigmoid functionPerform element-wise multiplication and apply the dilated convolutional layer to obtain the spatial attention feature , the spatial attention feature has the following expression:

[0039]

[0040] Finally, by adding the spatial attention feature to the output of the shifted window multi-head self-attention module , the stage output feature is obtained. The expression of the stage output feature is as follows:

[0041]

[0042] The FIA module fuses local structural features into the global. The stage output features from Resnet feature extraction and the stage output features from Swin-transformer feature extraction will be input into the FIA module at the same level to obtain the multi-level fusion feature , which is used as the input to the decoder. The output resolution of the multi-level fusion feature is . The algorithm of the multi-level fusion feature is carried out according to the following operations:

[0043] For the feature , the adaptive average pooling layer is used to extract the corresponding spatial feature statistics in the horizontal and vertical directions respectively. The expressions for extracting the corresponding spatial feature statistics in the horizontal and vertical directions are:

[0044]

[0045]

[0046] where , ;

[0047] Multiply the two and to highlight the spatial position information of the landmark points and establish channel dependencies through two convolutional operations. The expression for establishing channel dependencies is:

[0048]

[0049] Then, enhance the spatial encoding ability of the feature with the spatial attention map generated by the CBAM block. The expression is as follows:

[0050]

[0051]

[0052] wherein and represent two different channel information, represents a splicing operation;

[0053] Next, the fine-grained features are integrated, and a convolutional layer with a small convolutional kernel is used to adjust the two different encoded features to the same dimension. The expression is as follows:

[0054]

[0055] Finally, the processed features are spliced, and the output features are formed through a residual operation , the output feature has the following expression:

[0056] .

[0057] The present invention also provides a hip joint postoperative evaluation system, including:

[0058] An anatomical site prediction module, integrated with the above-mentioned hip joint anatomical site prediction system, is configured to identify anatomical sites and anatomical regions according to the postoperative hip joint image data of the patient;

[0059] A parameter calculation module, configured to calculate postoperative prosthesis placement parameters according to the anatomical sites and anatomical regions;

[0060] A report generation module, configured to generate a postoperative prosthesis position evaluation report;

[0061] The output module also outputs the postoperative prosthesis placement parameters and the postoperative prosthesis position evaluation report.

[0062] The present invention also provides a hip joint preoperative evaluation system, including:

[0063] An anatomical site prediction module, integrated with the above-mentioned hip joint anatomical site prediction system, is configured to identify anatomical sites and anatomical regions according to the preoperative hip joint image data of the patient;

[0064] A reinforcement learning module, configured to use a reinforcement learning algorithm agent to adjust the anatomical sites and anatomical regions to meet clinical constraints, and output an optimal solution and multiple candidate solutions;

[0065] A parameter calculation module, configured to calculate preoperative prosthesis placement parameters according to the optimal solution and the predicted optimal and candidate anatomical sites and anatomical regions among the multiple candidate solutions;

[0066] A planning generation module, configured to generate a preoperative planning UI interface, which can be interacted with and provide real-time feedback;

[0067] An output module, configured to output the predicted optimal and candidate anatomical sites and anatomical regions, and output the preoperative prosthesis placement parameters and the preoperative planning UI interface.

[0068] Furthermore, the anatomical sites include the center point of the prosthesis, the outer edge point of the acetabulum, and the diameter of the prosthesis;

[0069] The clinical constraints include clinical parameter constraints and anatomical rationality constraints.

[0070] Furthermore, the clinical parameter constraints include: (1) The coverage rate is greater than 80%; (2) The abduction angle is 40°.

[0071] The anatomical rationality constraints include: (1) The diameter of the prosthesis is an integer within 38 - 56 mm; (2) The outer edge point of the acetabulum is the highest point among the anatomical sites related to the prosthesis; (3) The outer edge point of the acetabulum is within the safe range of acetabular reconstruction.

[0072] Furthermore, the reinforcement learning module is obtained through training. The training process of the reinforcement learning module is as follows: Input the anatomical sites and anatomical regions obtained by the anatomical site prediction module; Train the reinforcement learning policy network using a reward function that only includes clinical parameter constraints; Add anatomical rationality constraints to the reward function and adjust the weights in the reward function, and continue to train the policy network; Adjust the temperature parameter and different sampling strategies to generate multiple solutions, and select the top 3 solutions according to the comprehensive score ranking as candidate solutions.

[0073] Furthermore, the reward function is 60% for clinical parameters, 35% for anatomical rationality, and 5% for deviation penalty. Clinical parameter rewards: (1) The coverage rate is greater than 80%; (2) The abduction angle is 40°. Anatomical rationality rewards: (1) The diameter of the prosthesis conforms to the actual clinical size; (2) The outer extension point of the acetabulum should be the highest point among the anatomical sites related to the prosthesis; (3) The outer extension point of the acetabulum should be within the safe range of acetabular reconstruction. Deviation penalty: To prevent the reinforcement learning algorithm from over-adjusting the point position.

[0074] Furthermore, the loss function in the deviation penalty uses the mean squared error loss function.

[0075] Furthermore, the loss function in the deviation penalty is: (1) Anatomical site prediction loss: MSE loss and key point heat map cross-entropy loss; (2) Region prediction loss: Dice loss; (3) Constraint condition loss: Coverage rate constraint loss (>=80%); Abduction angle constraint loss (=40°).

[0076] The present invention also provides a storage medium, in which a computer program is stored. Wherein, the computer program is configured to execute the above hip joint anatomical site prediction system, the above hip joint postoperative evaluation system or the above hip joint preoperative evaluation system when running.

[0077] In the description of the present invention, the ordinal numbers such as "first", "second", "third", "fourth", etc. are only used for descriptive purposes to indicate the relative positions or orders of different technical features, and are not used to limit the importance, functions or roles of these features. The use of these ordinal numbers is for the convenience of understanding and explaining the technical solution of the present invention, and does not represent any form of priority or functional difference. Those skilled in the art should understand that the use of these ordinal numbers is to simplify the description and does not constitute any form of limitation to the protection scope of the present invention.

[0078] Compared with the prior art, the present invention has achieved the following beneficial effects:

[0079] (1) Improve the clinical DDH surgery efficiency: Before the operation, the present invention can provide the different effects of different DDH surgery plans, can display different surgery plans before the operation, doctors do not need to manually measure the relevant parameters, saving time; Patients can also intuitively watch the effects of the surgery plans, facilitating patients to understand the surgery and improving the doctor-patient communication efficiency.

[0080] (2) Promote the digitization of the THA preoperative planning and postoperative evaluation system: Whether before or after the operation, the X-rays of DDH patients and their measured relevant parameters can be digitized, which is convenient for storage, sharing and subsequent analysis, and supports the research of preoperative planning and postoperative evaluation.

[0081] (3) Reduce the burden on patients: Preoperative planning can be carried out through X-rays, which can greatly reduce the shooting cost, radiation dose and time compared with CT.

[0082] (4) Standardize parameters: Using artificial intelligence, there will be no problems with the definition of anatomical positions and cognitive errors, which is conducive to standardization.

[0083] (5) Effectively improve the situation where there is no dedicated calculation module for the case of acetabular dysplasia.

[0084] Obviously, based on the above content of the present invention, according to the common technical knowledge and customary means in the art, without departing from the above basic technical idea of the present invention, various other forms of modifications, substitutions or changes can also be made.

[0085] The following is a further detailed description of the above content of the present invention through specific embodiments in the form of examples. However, this should not be understood as limiting the scope of the above subject matter of the present invention to the following examples. All technologies implemented based on the above content of the present invention belong to the scope of the present invention. Brief Description of the Drawings

[0086] Figure 1 This is the overall technical roadmap of the hip joint anatomical site prediction system and the pre- and post-operative evaluation system of the present invention.

[0087] Figure 2 This is the network structure of the region segmentation model nnU-NetV2.

[0088] Figure 3 This is the network structure of the key point localization model Res-SwinFusion.

[0089] Figure 4 These are the independently proposed DFG and FIA modules.

[0090] Figure 5 This is the distribution of anatomical sites.

[0091] Figure 6 This is the situation of the postoperative prosthesis placement parameters.

[0092] Figure 7 This is the size and placement position of the acetabular cup for acetabular reconstruction. Detailed Implementation Modes

[0093] It should be specifically noted that the algorithms for data acquisition, transmission, storage, and processing steps not specifically described in the embodiments, as well as the hardware structures and circuit connections not specifically described, can all be implemented through the content publicly disclosed in the prior art.

[0094] Example 1: Hip Joint Anatomical Site Prediction System

[0095] Figure 1 This is the overall technical roadmap of the hip joint anatomical site prediction system of the present invention, including:

[0096] An input module, configured to input the hip joint image data to be recognized of a patient, where the hip joint image data to be recognized is obtained by data augmentation and data conversion of medical digital imaging and communication data, and the hip joint image data to be recognized is a pelvic anteroposterior X-ray film.

[0097] In the present invention, first, the medical digital imaging and communication (Digital Imaging and Communications in Medicine, abbreviated as DICOM) data of the patient's total hip joint is input, the DICOM data is subjected to data augmentation, and then the DICOM data is converted into a png format picture, that is, the hip joint image data to be recognized is obtained.

[0098] In the present invention, the data augmentation refers to image augmentation, and the process of image augmentation includes: (1) applying a VOI LUT to enhance the DICOM contrast; (2) normalizing and performing percentile clipping on the image (the low percentile is 2% and the high percentile is 100%); (3) stretching the pixel values to the range of [0, 255] and converting the image to 8-bit unsigned integer type; (4) using CLAHE to enhance the local contrast, where the clipLimit is set to 100 and the tileGridSize is set to (25, 6).

[0099] A data processing module, configured to input the hip joint image data to be recognized into a deep learning network model to recognize the anatomical site and anatomical region.

[0100] In the present invention, the hip joint image data to be recognized is recognized by a deep learning network model. The deep learning network model is composed of a region segmentation model and a key point localization model Res-SwinFusion after joint training. The hip joint image data to be recognized obtains the anatomical site and anatomical region through the region segmentation model and the key point localization model Res-SwinFusion, and the final anatomical site and anatomical region are obtained after synthesizing the two results.

[0101] In the present invention, joint training means connecting multiple networks in parallel and verifying each other, which is achieved through a joint loss function and a cross-validation mechanism. It includes: (1) Joint loss function: The site loss uses the mean square error (MSE) to measure the accuracy of key point prediction; Segmentation loss: The Dice loss is used to measure the accuracy of the segmented region; Constraint loss: Check whether the key point is on the edge or inside the segmented region. For each site , if it is not in the segmented region inside, then a penalty term is added: , where represents the shortest distance from the key point to the edge of the region. The form of the joint loss function: ( is a weight hyperparameter). (2) Cross-validation: The site result verifies the segmented region. If the site is not in the segmented region, calculate the error distance; The segmentation result verifies the site. If the segmented region does not cover the key point, calculate the coverage error.

[0102] In the present invention, the distribution of the anatomical sites is as Figure 5As shown in the figure, the definition of the anatomical sites includes: Points 1 and 2 are the lowest points of the lower edges of the bilateral teardrops; Points 3 and 4 are the highest points of the bilateral iliac crests; Point 5 is the medial point of the upper edge of the obturator foramen; Point 6 is the lowest point of the lateral edge of the ilium; Point 7 is the outermost edge point of the prosthesis; Point 8 is the outermost edge point on the other side of the prosthesis; Point 9 is the outer edge of the acetabulum; Point 10 is the center of the prosthesis; Point 11 is the center of the femoral head on the contralateral side; Points 12 and 13 are the lowest points of the lower edges of the bilateral ischia; Point 14 is the point where a parallel line to the teardrop (the line connecting Points 1 and 2) intersects the lateral edge of the ilium passing through the outer edge of the acetabulum; Point 15 is the point where a parallel line to the Kohler line is tangent to the edge of the prosthesis.

[0103] The network structure of the region segmentation model is as Figure 2 shown. In the present invention, the region segmentation model is an improved nnU-NetV2 model obtained by removing the random flipping part in the preprocessing of nnU-NetV2. Before removing the random flipping, the average Dicewei of the segmentation was 0.862, and after removing the random flipping, the average Dice of the segmentation was improved to 0.926. The Dice coefficient (Dice Similarity Coefficient, DSC) is an index used to evaluate the performance of the segmentation model, which measures the similarity between the predicted segmentation result and the true annotation. The value range of the Dice coefficient is [0, 1], and the closer the value is to 1, the higher the similarity between the predicted result and the true annotation, and the better the segmentation effect.

[0104] In the present invention, the key point localization model Res-SwinFusion is a model constructed by combining a U-shaped architecture with a Swin transformer feature extraction module and a ResNet backbone network. By capturing multi-scale context information and modeling long-range dependencies, the feature discrimination ability for anatomical landmarks is significantly improved. The key point localization model Res-SwinFusion includes a feature fusion structure, and the feature fusion structure includes an FIA module. The Swin transformer feature extraction module in the key point localization model Res-SwinFusion includes a DFG module.

[0105] As Figure 3 、 4 shown. The key point localization model Res-SwinFusion includes the following structures:

[0106] Set up an input layer. The back end of the input layer is respectively connected to a linear embedding layer, a first convolutional layer, and a second convolutional layer. The back end of the linear embedding layer is sequentially connected to a first Swin-transformer feature extraction module, a second Swin-transformer feature extraction module, a third Swin-transformer feature extraction module, and a fourth Swin-transformer feature extraction module. The back end of the first convolutional layer is sequentially connected to a first Resnet feature extraction module, a second Resnet feature extraction module, a third Resnet feature extraction module, and a fourth Resnet feature extraction module.

[0107] The feature maps output by the first Swin-transformer feature extraction module and the first Resnet feature extraction module are fused through a first feature interaction aggregation module. The feature maps output by the second Swin-transformer feature extraction module and the second Resnet feature extraction module are fused through a second feature interaction aggregation module. The feature maps output by the third Swin-transformer feature extraction module and the third Resnet feature extraction module are fused through a third feature interaction aggregation module. The feature maps output by the fourth Swin-transformer feature extraction module and the fourth Resnet feature extraction module are fused through a fourth feature interaction aggregation module.

[0108] The back end of the fourth feature interaction aggregation module is connected to a first feature connection module by upsampling. The third feature interaction aggregation module is connected to the first feature connection module through a skip connection. The back end of the first feature connection module is connected to a second feature connection module by upsampling. The second feature interaction aggregation module is connected to the second feature connection module through a skip connection. The back end of the second feature connection module is connected to a third feature connection module by upsampling. The first feature interaction aggregation module is connected to the third feature connection module through a skip connection. The back end of the third feature connection module is connected to a fourth feature connection module by upsampling. The first convolutional layer is connected to the fourth feature connection module through a skip connection. The back end of the fourth feature connection module is connected to a fifth feature connection module by upsampling. The second convolutional layer is connected to the fifth feature connection module through a skip connection.

[0109] The back end of the fifth feature connection module is connected to a third convolutional layer.

[0110] Further, the key point localization model Res-SwinFusion includes the following structure:

[0111] Input layer, configured to input with a size of The image data, representing the number of channels of the image, respectively represent the height and width of the image;

[0112] A linear embedding layer, configured to implement segmentation and linear mapping of the input image;

[0113] The first convolutional layer, configured to output the image data with a size of as a feature map with a size of through a convolutional operation;

[0114] The second convolutional layer, configured to perform a convolutional operation on the image data with an input size of ;

[0115] The Swin-transformer feature extraction module, configured to each Swin-transformer feature extraction module includes a window-based multi-head self-attention module, a shifted window multi-head self-attention module, and a DFG module. The window-based multi-head self-attention module and the shifted window multi-head self-attention module are connected, and the DFG module is connected before the window-based multi-head self-attention module and after the shifted window multi-head self-attention module; among them, the first Swin-transformer feature extraction module is configured to output the segmented and linearly mapped image data output by the linear embedding layer as through 2 feature extraction modules; The second Swin-transformer feature extraction module is configured to first perform downsampling and hierarchical feature construction, and then output the feature map with a size of as a feature map with a size of through 2 feature extraction modules; The third Swin-transformer feature extraction module is configured to first perform downsampling and hierarchical feature construction, and then output the feature map with a size of as a feature map with a size of through 18 feature extraction modules; The fourth Swin-transformer feature extraction module is configured to first perform downsampling and hierarchical feature construction, and then output the feature map with a size of as a feature map with a size of through 2 feature extraction modules;

[0116] The first Resnet feature extraction module is configured to output the feature map with a size of output by the first convolutional layer as a feature map with a size of after passing through 1 residual block of Resnet50; The second Resnet feature extraction module is configured to output the feature map with a size of The feature map outputs a feature map with a size of after passing through 1 residual block of Resnet50 The third Resnet feature extraction module is configured to output a feature map with a size of The feature map outputs a feature map with a size of after passing through 1 residual block of Resnet50 The fourth Resnet feature extraction module is configured to output a feature map with a size of The feature map outputs a feature map with a size of after passing through 1 residual block of Resnet50 The feature map;

[0117] The first feature interaction aggregation module is configured to fuse the feature maps output by the first Swin-transformer feature extraction module and the first Resnet feature extraction module and then output a feature fusion map with a size of The second feature interaction aggregation module is configured to fuse the feature maps output by the second Swin-transformer feature extraction module and the second Resnet feature extraction module and then output a feature fusion map with a size of The third feature interaction aggregation module is configured to fuse the feature maps output by the third Swin-transformer feature extraction module and the third Resnet feature extraction module and then output a feature fusion map with a size of The fourth feature interaction aggregation module is configured to fuse the feature maps output by the fourth Swin-transformer feature extraction module and the fourth Resnet feature extraction module and then output a feature fusion map with a size of The feature fusion map;

[0118] The first feature connection module is configured to output a feature map with a size of from the feature maps output by the fourth feature interaction aggregation module and the third feature interaction aggregation module The second feature connection module is configured to output a feature map with a size of from the feature maps output by the first feature connection module and the second feature interaction aggregation module The third feature connection module is configured to output a feature map with a size of from the feature maps output by the second feature connection module and the first feature interaction aggregation module The fourth feature connection module is configured to output a feature map with a size of from the feature maps output by the third feature connection module and the first convolutional layer The fifth feature connection module is configured to output a feature map with a size of from the feature maps output by the fourth feature connection module and the second convolutional layer The feature map;

[0119] The third convolutional layer is configured to perform a convolution operation on the output of the fifth feature connection module to obtain a key point heat map with a size .

[0120] Among them, the DFG module passes the shallow representation to the deep layer by connecting the features before and after calculating the attention in each Swin Transformer block, and constructs additional spatial information on the residual edge to further assist in global information encoding. This design can increase the proportion of target landmark points in the encoding weights, enabling the model to establish global attention through more discriminative features and reducing the occurrence of detection ambiguity.

[0121] The detailed description of the DFG module of the present invention is as follows:

[0122] represents the nth stage (for example, n is 1, 2, 3, 4 corresponding to the Swin-transformer feature extraction of the 1st, 2nd, 3rd, and 4th parts respectively), ; for different stages, the input of the Swin-transformer feature extraction module is first reshaped to , where represents the height of the feature map in the nth stage, ; represents the width of the feature map in the nth stage, ; represents the height of the original input image, represents the width of the original input image; then an atrous convolutional layer ( with a convolutional kernel for convolution) is applied to expand the receptive field and compress the number of channels to , obtaining the feature :

[0123]

[0124] where is the reshaping function for matching the size;

[0125] Next, global average pooling ( ) is performed on and then reshaped (Reshape) to obtain the feature :

[0126]

[0127] The feature after reshaping (Reshape) of is subjected to a tensor product with the feature ( ), and then the result obtained by normalization :

[0128]

[0129] Finally, for and explicitly modeled by the sigmoid function, perform element-wise multiplication (Hadamard product, ), apply a dilated convolutional layer ( 's convolutional kernel for convolution) to obtain the spatial attention feature :

[0130]

[0131] Finally, by adding the spatial attention feature to the output of the shifted window multi-head self-attention module (SW-Trans module), the stage output feature is obtained:

[0132]

[0133] The initial inspiration for the design of the FIA module comes from BiFusion. The FIA module provides more accurate location information by fusing local structural features into the global. Specifically, the stage output feature from Resnet feature extraction and the stage output feature from Swin-transformer feature extraction will be input into the FIA module at the same level to obtain the multi-level fusion feature , with the output resolution of , which is used as the input to the decoder.

[0134] The detailed description of the FIA module of the present invention is as follows:

[0135] First, for the feature , use the adaptive average pooling layer to extract the corresponding spatial feature statistics in the horizontal and vertical directions respectively, as shown in the following formula:

[0136]

[0137]

[0138] where , ;

[0139] Next, multiply the two aggregation tensors to highlight the spatial position information of the landmark points, and establish channel dependencies through two convolutional operations, that is:

[0140]

[0141] Then, for the feature , enhance the spatial encoding ability of with the spatial attention map generated by the CBAM block. The calculation process is as follows:

[0142]

[0143]

[0144] Where and represent two different channel information, represents the concatenation operation; then perform element-wise multiplication to integrate the fine-grained features, and apply a convolutional layer with a small convolutional kernel to adjust the two different encoded features to the same dimension. The formula is as follows:

[0145]

[0146] Finally, concatenate the processed features and form the output features through a residual operation , and the formula is as follows:

[0147]

[0148] In the present invention, each module can be used alone or in combination with and integrated with the network structures of the Swin transformer and ResNet.

[0149] An output module, configured to output the identified anatomical site and anatomical region.

[0150] Example 2, Hip Joint Postoperative Evaluation System

[0151] Figure 1 is the overall technical roadmap of the hip joint postoperative evaluation system of the present invention, including:

[0152] An anatomical site prediction module, integrated with the hip joint anatomical site prediction system of Example 1, configured to identify the anatomical site and anatomical region according to the preoperative hip joint image data of the patient;

[0153] A parameter calculation module, configured to calculate the postoperative prosthesis placement parameters according to the anatomical site and anatomical region;

[0154] A report generation module, configured to generate a postoperative prosthesis position evaluation report;

[0155] The output module also outputs the postoperative prosthesis placement parameters and the postoperative prosthesis position evaluation report.

[0156] In the present invention, the postoperative prosthesis placement parameters are as Figure 6 shown. The definitions of the postoperative prosthesis placement parameters include: (1) Pelvic height (PH): the distance between the line connecting points 3 and 4 and the line connecting points 12 and 13; (2) Distance between the iliac crest line and the teardrop line (IS_to_TD): the distance between the line connecting points 3 and 4 and the line connecting points 1 and 2; (3) Acetabular roof thickness (ATD): the distance from point 9 to point 14; (4) Acetabular roof height (ADH): the distance from the line connecting points 9 and 14 to the line connecting points 1 and 2; (5) Coverage rate α (coverage_rate): the acute angle formed by the line connecting points 9 and 10 and the line connecting points 7 and 10; (6) Inclination angle (Inclination_angle): the acute angle formed by the line connecting points 7 and 8 and the line connecting points 1 and 2; (7) Distance from the prosthesis center to the teardrop line (Hip_center_height): the distance from point 10 to the line connecting points 1 and 2; (8) Distance from the contralateral center to the teardrop line (other_center_height): the distance from point 11 to the line connecting points 1 and 2; (9) Prosthesis diameter (Cup_size): the distance from point 7 to point 8; (10) Minimum distance from the cup to the Kohler line (MDCK): the distance from point 15 to the line connecting points 5 and 6; (11) Minimum distance from the cup edge to the teardrop line (Cup_to_TD): the distance from point 8 to the line connecting points 1 and 2.

[0157] Example 3. Preoperative evaluation system for hip joint

[0158] Figure 1 is the overall technical roadmap of the preoperative evaluation system for hip joint of the present invention, including:

[0159] Anatomical site prediction module, integrated with the hip joint anatomical site prediction system of Example 1, is configured to identify anatomical sites and anatomical regions according to the preoperative hip joint image data of the patient;

[0160] Reinforcement learning module, is configured to use the reinforcement learning algorithm agent to adjust the anatomical sites and anatomical regions to meet the clinical constraints, and output the optimal solution and multiple candidate solutions;

[0161] Parameter calculation module, is configured to calculate the preoperative prosthesis placement parameters according to the optimal solution and the predicted optimal and candidate anatomical sites and anatomical regions among the multiple candidate solutions;

[0162] Planning generation module, is configured to generate a preoperative planning UI interface, which can be interacted with and provide real-time feedback;

[0163] Output module, is configured to output the predicted optimal and candidate anatomical sites and anatomical regions, output the preoperative prosthesis placement parameters and the preoperative planning UI interface.

[0164] In the present invention, the preoperative prosthesis placement parameters are obtained by combining a multi-task regression model and the safe area for acetabular reconstruction.

[0165] In the present invention, the multi-task regression model utilizes preoperative and postoperative data (preoperative X-ray films, preoperatively segmented regions, postoperative relevant parameters, and postoperative anatomical sites). With Resnet as the backbone network and combined with transformer, through reinforcement learning, the following effects are achieved: input the preoperative anteroposterior pelvic X-ray film; output the optimal placement position, orientation, and size of the prosthesis when the coverage rate is above 80% and the abduction angle is 40°, and multiple candidate solutions will be provided, allowing the doctor to manually adjust the prosthesis position and display the feedback after changing the prosthesis position in real time.

[0166] The training process of the multi-task regression model is as follows: input data; pre-train the backbone network; initially predict the placement position, orientation, and size of the preoperative prosthesis (determined by key points); the initial key point prediction results enter the reinforcement learning environment; the reinforcement learning agent adjusts the key point positions to meet clinical constraints; output the optimal solution and multiple candidate solutions.

[0167] The reinforcement learning algorithm is the Soft Actor-Critic (SAC) algorithm architecture, which includes a policy network (outputting action mean and variance), a dual Q network to improve training stability, and entropy regularization to promote exploration. The discount factor γ = 0.99. The reward function is 60% for clinical parameters, 35% for anatomical rationality, and 5% for deviation penalty. Clinical parameter rewards: (1) The coverage rate is greater than 80%; (2) The abduction angle is 40°. Anatomical rationality rewards: (1) The prosthesis diameter conforms to the actual clinical size; (2) The acetabular extension point should be the highest point among the prosthesis-related anatomical sites; (3) The acetabular extension point should be within the safe area for acetabular reconstruction. Deviation penalty: Prevent the reinforcement learning algorithm from over-adjusting the point positions.

[0168] The loss function in the deviation penalty uses the mean square error loss function. The loss function in the deviation penalty is: (1) Anatomical site prediction loss: MSE loss and key point heat map cross-entropy loss; (2) Region prediction loss: Dice loss; (3) Constraint condition loss: Coverage rate constraint loss (>=80%); Abduction angle constraint loss (=40°).

[0169] In the present invention, the preoperative prosthesis placement parameters refer to the size and placement position of the acetabular cup for acetabular reconstruction, as Figure 7 shown.

[0170] In the present invention, the preoperative prosthesis placement parameters are obtained from the parameter calculation module, and through post-processing, information such as the size of the acetabular cup for acetabular reconstruction within the safe area, the coverage rate, and the difference in lower limb lengths can be dynamically displayed in real time.

[0171] The beneficial effects of the present invention are demonstrated by the following experimental examples.

[0172] Experimental Example 1, Influence of Different Modules on the Key Point Localization Model Res-SwinFusion

[0173] 1. Experimental Method

[0174] To explore the contributions of different modules to the model, the present invention conducted a set of ablation experiments on Test1 of a publicly available cephalometric dataset (ISBI 2015 Grand Challenge). The baseline network was UNet with a ResNet101 encoder. The Swin Transformer backbone adopted the "base" configuration. To only discuss the influence of introducing the Swin Transformer, the present invention fused the two encoding features through simple skip connections and denoted it as "Swin-U" in the results. BiFusion was the initial inspiration source for the FIA module design. The present invention used two commonly used metrics, the mean radial error (MRE) and the successful detection rate (SDR), to evaluate the localization performance; MRE was defined as the average Euclidean distance between the ground truth and the predicted position, and SDR was the percentage of predicted landmarks with MRE less than a given threshold; for a test set with N images and M landmarks, the calculation formulas are as follows:

[0175] MRE =

[0176] SD =

[0177] SDR = ×100

[0178] 2. Experimental Results

[0179] Table 1 Results of Ablation Experiments

[0180]

[0181] The results in Table 1 show that in the present invention, introducing the DFG module, BiFusion module, or FIA module alone can reduce the mean radial error of the model and improve the successful detection rate of the Swin-U model. Introducing two modules simultaneously has a better effect on reducing the mean radial error of the model and improving the successful detection rate of the model. Among them, compared with the BiFusion module, the improved FIA module has a better effect.

[0182] In summary, the present invention provides a hip joint anatomical site prediction system, a pre-operative and post-operative evaluation system, and a storage medium. The system of the present invention includes: an input module for inputting the hip joint image data to be recognized of a patient; a data processing module for inputting the hip joint image data to be recognized into a deep learning network model to obtain the anatomical sites and anatomical regions in the hip joint image data to be recognized; and an output module for outputting the recognized anatomical sites and anatomical regions, calculating parameters based on the anatomical sites and anatomical regions, and outputting a pre-operative plan or a post-operative evaluation of anatomical prediction. The system of the present invention has high accuracy and robustness, not only improves the surgical efficiency of clinical hip dysplasia, but also promotes the digitization of the pre-operative planning and post-operative evaluation systems for total hip arthroplasty, reduces the burden on patients, is conducive to standardizing parameters, and has broad application prospects.

Claims

1. A hip joint anatomical site prediction system, characterized in that, Including: An input module configured to input hip joint image data to be recognized of a patient; A data processing module configured to input the hip joint image data to be recognized into a deep learning network model to recognize anatomical sites and anatomical regions; An output module configured to output the recognized anatomical sites and anatomical regions; Wherein, the deep learning network model is composed of a region segmentation model and a key point localization model Res-SwinFusion; the key point localization model Res-SwinFusion is a model constructed by combining a U-shaped architecture with a Swin transformer feature extraction module and a ResNet backbone network; The key point localization model Res-SwinFusion includes the following structure: An input layer is provided, and the backend of the input layer is respectively connected to a linear embedding layer, a first convolutional layer and a second convolutional layer. The backend of the linear embedding layer is sequentially connected to a first Swin-transformer feature extraction module, a second Swin-transformer feature extraction module, a third Swin-transformer feature extraction module, and a fourth Swin-transformer feature extraction module; the backend of the first convolutional layer is sequentially connected to a first Resnet feature extraction module, a second Resnet feature extraction module, a third Resnet feature extraction module, and a fourth Resnet feature extraction module; The feature maps output by the first Swin-transformer feature extraction module and the first Resnet feature extraction module are fused through a first feature interaction aggregation module; the feature maps output by the second Swin-transformer feature extraction module and the second Resnet feature extraction module are fused through a second feature interaction aggregation module; the feature maps output by the third Swin-transformer feature extraction module and the third Resnet feature extraction module are fused through a third feature interaction aggregation module; the feature maps output by the fourth Swin-transformer feature extraction module and the fourth Resnet feature extraction module are fused through a fourth feature interaction aggregation module; The backend of the fourth feature interaction aggregation module is connected to the first feature connection module through upsampling, and the third feature interaction aggregation module is connected to the first feature connection module through a skip connection; the backend of the first feature connection module is connected to the second feature connection module through upsampling, and the second feature interaction aggregation module is connected to the second feature connection module through a skip connection; the backend of the second feature connection module is connected to the third feature connection module through upsampling, and the first feature interaction aggregation module is connected to the third feature connection module through a skip connection; the backend of the third feature connection module is connected to the fourth feature connection module through upsampling, and the first convolutional layer is connected to the fourth feature connection module through a skip connection; the backend of the fourth feature connection module is connected to the fifth feature connection module through upsampling, and the second convolutional layer is connected to the fifth feature connection module through a skip connection; The backend of the fifth feature connection module is connected to a third convolutional layer; Swin-transformer feature extraction module, configured such that each Swin-transformer feature extraction module includes a window-based multi-head self-attention module, a shifted window multi-head self-attention module, and a DFG module. The window-based multi-head self-attention module and the shifted window multi-head self-attention module are connected, and the DFG module is connected before the window-based multi-head self-attention module and after the shifted window multi-head self-attention module. The detailed description of the DFG module is as follows: Denote the n-th stage, ; For different stages, the input of the Swin-transformer feature extraction module is first reshaped into , where represents the height of the feature map in the n-th stage, ; represents the width of the feature map in the n-th stage, ; represents the height of the original input image, represents the width of the original input image; then apply a dilated convolutional layer to expand the receptive field and compress the number of channels to , obtaining the feature : Among them, is a reshaping function used to match dimensions; Next, for perform global average pooling operation and then reshape to obtain the feature : Pair The reshaped feature and the feature Perform a tensor product and then normalize the resulting : Finally, for and explicitly modeled by the sigmoid function, perform element-wise multiplication, and apply a dilated convolutional layer to obtain the spatial attention feature : Finally, by adding the spatial attention feature to the output of the shifted window multi-head self-attention module the stage output feature is obtained : 。 2. The hip joint anatomical site prediction system according to claim 1, wherein The hip joint image data to be recognized is obtained by data augmentation and data conversion of medical digital imaging and communication data, and the hip joint image data to be recognized is a pelvic anteroposterior X-ray film.

3. The hip joint anatomical site prediction system according to claim 1, wherein The region segmentation model is an improved nnU-NetV2 model that removes the random flipping part in the preprocessing of nnU-NetV2; The key point localization model Res-SwinFusion includes a feature fusion structure, and the feature fusion structure includes an FIA module.

4. The hip joint anatomical site prediction system according to claim 1, characterized in that, The key point localization model Res-SwinFusion includes the following structures: The input layer is configured to input image data with a size of , representing the number of channels of the image, respectively representing the height and width of the image; Linear embedding layer, configured to implement segmentation and linear mapping of the input image; The first convolutional layer is configured to output image data of size as feature maps of size through a convolution operation; The second convolutional layer is configured to perform a convolution operation on image data with an input size of ; Among them, the first Swin-transformer feature extraction module is configured to output the segmented and linearly mapped image data output by the linear embedding layer as a feature map of through 2 feature extraction modules; the second Swin-transformer feature extraction module is configured to first perform downsampling and hierarchical feature construction, and then output the feature map of size as a feature map of size through 2 feature extraction modules; the third Swin-transformer feature extraction module is configured to first perform downsampling and hierarchical feature construction, and then output the feature map of size as a feature map of size through 18 feature extraction modules; the fourth Swin-transformer feature extraction module is configured to first perform downsampling and hierarchical feature construction, and then output the feature map of size as a feature map of size through 2 feature extraction modules; The first Resnet feature extraction module is configured to output a feature map with a size of after passing the feature map output by the first convolutional layer through 1 residual block of Resnet50 as a feature map with a size of ; The second Resnet feature extraction module is configured to output a feature map with a size of after passing the feature map output by the first Resnet feature extraction module through 1 residual block of Resnet50 as a feature map with a size of ; The third Resnet feature extraction module is configured to output a feature map with a size of after passing the feature map output by the second Resnet feature extraction module through 1 residual block of Resnet50 as a feature map with a size of ; The fourth Resnet feature extraction module is configured to output a feature map with a size of after passing the feature map output by the third Resnet feature extraction module through 1 residual block of Resnet50 as a feature map with a size of ; The first feature interaction aggregation module is configured to fuse the feature maps output by the first Swin-transformer feature extraction module and the first Resnet feature extraction module and then output a feature fusion map with a size of ; The second feature interaction aggregation module is configured to fuse the feature maps output by the second Swin-transformer feature extraction module and the second Resnet feature extraction module and then output a feature fusion map with a size of ; The third feature interaction aggregation module is configured to fuse the feature maps output by the third Swin-transformer feature extraction module and the third Resnet feature extraction module and then output a feature fusion map with a size of ; The fourth feature interaction aggregation module is configured to fuse the feature maps output by the fourth Swin-transformer feature extraction module and the fourth Resnet feature extraction module and then output a feature fusion map with a size of ; The first feature connection module is configured to output the feature maps output by the fourth feature interaction aggregation module and the third feature interaction aggregation module as feature maps with a size of ; The second feature connection module is configured to output the feature maps output by the first feature connection module and the second feature interaction aggregation module as feature maps with a size of ; The third feature connection module is configured to output the feature maps output by the second feature connection module and the first feature interaction aggregation module as feature maps with a size of ; The fourth feature connection module is configured to output the feature maps output by the third feature connection module and the first convolutional layer as feature maps with a size of ; The fifth feature connection module is configured to output the feature maps output by the fourth feature connection module and the second convolutional layer as feature maps with a size of ; The third convolutional layer is configured to perform a convolutional operation on the output of the fifth feature connection module to obtain a key point heat map with a size of ​ 5. A hip joint postoperative evaluation system, characterized in that, Including: Anatomical site prediction module, integrated with the hip joint anatomical site prediction system according to any one of claims 1 to 4, configured to identify anatomical sites and anatomical regions based on the postoperative hip joint image data of the patient; Parameter calculation module, configured to calculate postoperative prosthesis placement parameters according to the anatomical sites and anatomical regions; Report generation module, configured to generate a postoperative prosthesis position evaluation report; The output module also outputs the postoperative prosthesis placement parameters and the postoperative prosthesis position evaluation report.

6. A hip joint preoperative evaluation system, characterized in that, Including: Anatomical site prediction module, integrated with the hip joint anatomical site prediction system according to any one of claims 1 to 4, configured to identify anatomical sites and anatomical regions based on the preoperative hip joint image data of the patient; Reinforcement learning module, configured to use a reinforcement learning algorithm agent to adjust the anatomical sites and anatomical regions to meet clinical constraints, and output an optimal solution and multiple candidate solutions; A parameter calculation module, configured to calculate preoperative prosthesis placement parameters according to the optimal solution, the predicted optimal and candidate anatomical sites, and the anatomical regions among multiple candidate solutions; A planning generation module, configured to generate a preoperative planning UI interface, which can be interacted with and provide real-time feedback; An output module, configured to output the predicted optimal and candidate anatomical sites and anatomical regions, and output the preoperative prosthesis placement parameters and the preoperative planning UI interface.

7. The hip joint preoperative evaluation system according to claim 6, wherein The anatomical sites include the prosthesis center point, the acetabular outer edge point, and the prosthesis diameter; The clinical constraints include clinical parameter constraints and anatomical rationality constraints.

8. The hip joint preoperative evaluation system according to claim 7, wherein The clinical parameter constraints include: (1) The coverage rate is greater than 80%; (2) The abduction angle is 40°; The anatomical rationality constraints include: (1) The prosthesis diameter is an integer within 38 - 56 mm; (2) The acetabular outer edge point is the highest point among the prosthesis-related anatomical sites; (3) The acetabular outer edge point is within the acetabular reconstruction safety range.

Images