Pelvic bone medical image key point detection system and method based on deep learning
Through the deep learning-based pelvic medical imaging key point detection system, using the YOLOv7-pose network model training and verification, the problem of slow detection of image key point in the existing technology of traditional Chinese medicine has been solved, and a faster and more accurate detection effect is achieved.
Patent Information
- Application Number
- CN202411110944.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-13
- Publication Date
- 2024-11-26
AI Technical Summary
The existing medical image key point detection algorithm relies on computing resources, and the reasoning takes a long time, making it difficult to meet the real-time needs of medical image detection.
The pelvic medical imaging key point detection system based on deep learning, including data acquisition, data annotation, model training and image inference calculation modules, is used to optimize the model to improve detection speed and accuracy through YOLOv7-pose network model training and verification.
The detection speed of medical image key point with advantages over traditional algorithms is realized, which improves the performance of the model on unseen data, reduces inference time, and enhances the accuracy and robustness of the detection.
Smart Images

Figure CN119027391A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of medical image detection, and in particular relates to a pelvic medical image key point detection system and method based on deep learning. Background Art
[0002] At present, the key point detection algorithms of medical images are generally divided into two types: Top-down and Bottom-up. Different from target detection, key point detection algorithms based on heat maps or detector processing are more dependent on computing resources and take a little longer to infer. In order to solve the problem of key point detection speed in medical images, we propose a pelvic medical image key point detection system and method based on deep learning. Summary of the invention
[0003] The object of the present invention is to address the problems raised in the above background technology.
[0004] The present invention adopts the following technical solutions:
[0005] A pelvic medical image key point detection system based on deep learning, comprising:
[0006] Data acquisition module, the data acquisition module consists of an image acquisition and storage unit, a data cleaning and preprocessing unit and a data desensitization unit. The image acquisition and storage unit is responsible for acquiring pelvic medical image data from different medical image sources, including different types of medical images such as X-ray, MRI, and CT, and storing them in the system's database. The data cleaning and preprocessing unit cleans and preprocesses the above-mentioned acquired medical images affected by noise, artifacts or other interference factors. The data desensitization unit is used to remove the patient's privacy information to prevent data leakage;
[0007] The data annotation module includes a region annotation unit and a key point annotation unit. The region annotation unit is used to annotate the detected region to provide the system with more detailed information. The key point annotation unit is used to annotate the key points in the annotated region.
[0008] A model training module, which is used to train a deep learning model using labeled data so that it can accurately detect key points in medical images. The model training module includes a data preparation unit, a model training unit, a model verification and tuning unit, and a model preservation unit;
[0009] An image reasoning calculation module, which includes an image input and preprocessing unit, a model loading unit, a forward propagation unit, and a result output unit;
[0010] User management module, which is used to manage user information, permissions and access control in the system. The user management module includes a user registration and authentication unit, a user information management unit, a role allocation and management unit, a role management unit, a user notification and communication unit, and a password retrieval and account recovery unit;
[0011] The role authority management module includes a role definition unit, a authority allocation unit, a role inheritance unit and a user group management unit.
[0012] Preferably, the data preparation unit is used to import the labeled data to be trained and divide it into a training set, a validation set and a test set. At the same time, the training data set is expanded by rotation, scaling and flipping to improve the robustness of the model. The model training unit uses the prepared training data set to train the YOLOv7-pose network model. During the training process, the model continuously adjusts parameters through back propagation and optimization algorithms to minimize the loss function. The model verification and tuning unit uses the validation set to evaluate the trained model and adjusts the hyperparameters of the model to further improve the performance and prevent overfitting, ensuring that the model performs well on unseen data. After the training is completed, the model saving unit saves the trained model and names it as the YOLOv7-pose key point detection model for subsequent reasoning and application on new data.
[0013] Preferably, the image input and preprocessing unit receives the medical image to be detected as input, the model loading unit loads the trained key point detection model, the forward propagation unit inputs the input medical image into the loaded model, performs the forward propagation operation, and thus generates the model's prediction for the position of each key point, the result output unit outputs the result of the key point detection, draws the key point position on the image and provides the coordinates of the key point as output, and outputs the calculation results of the corresponding parameters.
[0014] Preferably, the user registration and authentication unit allows users to register with the system and perform identity authentication to ensure the legitimacy of their identities. The user information management unit maintains the basic information of each registered user, including name, contact information, and affiliated organization. The role allocation and management unit assigns different roles to different users to control their access and operation scope in the system. The role management unit assigns users to different roles. The user notification and communication unit sends notifications to users. The password retrieval and account recovery unit is used for users to retrieve forgotten passwords or recover accounts to reduce access problems caused by identity authentication problems.
[0015] Preferably, the role definition unit defines different user roles in the system, each role represents a set of specific responsibilities and operating permissions, the permission allocation unit grants specific operations and access permissions to specific roles, including permissions to system functions, data and operations, the role inheritance unit implements the role inheritance mechanism so that higher-level roles automatically inherit the permissions of lower-level roles, and the user group management unit allows administrators to group users and assign specific roles and permissions to the entire user group.
[0016] Preferably, the working steps of the YOLOv7-pose key point detection model are as follows:
[0017] 1. Input image preprocessing: First, the model receives the input image and performs preprocessing, which includes resizing the image to the required size and performing normalization operations to ensure that the image can be effectively processed in the model;
[0018] 2. Feature extraction: On the preprocessed images, the model uses a convolutional neural network (CNN) for feature extraction. The network structure of the model contains specially designed layers to better capture the key point information of the image;
[0019] 3. Key point prediction: On the extracted feature map, the model predicts key points. Each key point consists of coordinate information and a confidence score associated with it. The goal of the model is to accurately locate key points in medical images.
[0020] 4. Loss function definition: In order to optimize the training process, the model defines a loss function, which takes into account the error of the key point position and the error of the confidence score. By continuously optimizing the loss function, the model learns how to predict the key points of the human body more accurately;
[0021] 5. Training and optimization: Using the training data with labeled key points, the model begins training. At this stage, the model fine-tunes weights and parameters through multiple iterations using the gradient descent optimization algorithm to better adapt to the key point detection task;
[0022] 6. Evaluation and tuning: When model training is completed, the model is evaluated using a validation set or a test set. The evaluation process includes but is not limited to analyzing the positioning accuracy, precision, and recall performance indicators of key points.
[0023] A method for detecting key points of pelvic medical images based on deep learning, comprising the following steps:
[0024] 1. First, the administrator assigns a role to the user, and the user obtains system usage permissions through the role;
[0025] 2. Users collect training data through the data collection module, perform data cleaning and preprocessing, and desensitize the data;
[0026] 3. The user annotates key points. Before annotating key points, the detected area needs to be annotated. Area annotation provides the system with more detailed information, which helps the model understand the image content more comprehensively.
[0027] 4. Perform model training and use labeled data to train the deep learning model so that it can accurately detect key points in medical images;
[0028] 5. Image reasoning calculation module, which uses the trained model to predict the newly input medical images and obtain the calculation results.
[0029] Compared with the prior art, the present invention can obtain pelvic medical image data from different medical image sources by setting a data acquisition module, and store it in a database of the system. The data cleaning and preprocessing unit will clean and preprocess the collected medical images affected by noise, artifacts or other interference factors, and mark the detected area by setting a data annotation module to provide more detailed information for the system, and use a key point annotation unit to mark the key points in the marked area. The deep learning model is trained by the model training module, which can accurately detect the key points in the medical image, and the image reasoning calculation module outputs the result of the key point detection, draws the key point position on the image and provides the coordinates of the key point as output; in addition, the YOLOv7-pose network model is trained by the model training unit using the prepared training data set to further obtain a mature YOLOv7-pose key point detection model, which has certain advantages in processing speed over traditional medical image key point detection algorithms such as Top-down and Bottom-up. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 This is a schematic diagram of the key point annotation of medical images in a pelvic medical image key point detection system and method based on deep learning proposed by the present invention. DETAILED DESCRIPTION
[0031] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0032] A pelvic medical image key point detection system based on deep learning, comprising:
[0033] Data acquisition module, the data acquisition module consists of an image acquisition and storage unit, a data cleaning and preprocessing unit and a data desensitization unit. The image acquisition and storage unit is responsible for acquiring pelvic medical image data from different medical image sources, including different types of medical images such as X-ray, MRI, CT, etc., and storing them in the system database. The data cleaning and preprocessing unit cleans and preprocesses the above-mentioned acquired medical images affected by noise, artifacts or other interference factors. The data desensitization unit is used to remove the patient's privacy information to prevent data leakage;
[0034] The data annotation module includes a region annotation unit and a key point annotation unit. The region annotation unit is used to annotate the detected region to provide the system with more detailed information. The key point annotation unit is used to annotate the key points in the annotated region.
[0035] Model training module: The model training module is used to train the deep learning model with labeled data so that it can accurately detect key points in medical images. It includes a data preparation unit, a model training unit, a model verification and tuning unit, and a model preservation unit. The data preparation unit is used to import the labeled data to be trained and divide it into a training set, a verification set, and a test set. At the same time, the training data set is expanded by rotation, scaling, and flipping to improve the robustness of the model. The model training unit uses the prepared training data set to train the YOLOv7-pose network model. During the training process, the model continuously adjusts parameters through back propagation and optimization algorithms to minimize the loss function. The model verification and tuning unit uses the verification set to evaluate the trained model and adjusts the model's hyperparameters to further improve performance, prevent overfitting, and ensure that the model performs well on unseen data. After the training is completed, the model preservation unit saves the trained model and names it the YOLOv7-pose key point detection model for subsequent reasoning and application on new data.
[0036] An image reasoning calculation module, the image reasoning calculation module includes an image input and preprocessing unit, a model loading unit, a forward propagation unit and a result output unit. The image input and preprocessing unit receives a medical image to be detected as an input, the model loading unit loads a trained key point detection model, the forward propagation unit inputs the input medical image into the loaded model, performs a forward propagation operation, and thus generates a prediction of the model for each key point position. The result output unit outputs the result of key point detection, draws the key point position on the image and provides the coordinates of the key point as output, and outputs the calculation result of the corresponding parameter;
[0037] User management module, the user management module is used to manage user information, permissions and access control in the system. The user management module includes a user registration and authentication unit, a user information management unit, a role allocation and management unit, a role management unit, a user notification and communication unit, and a password retrieval and account recovery unit. The user registration and authentication unit allows users to register the system and perform identity authentication to ensure the legitimacy of their identities. The user information management unit maintains the basic information of each registered user, including name, contact information, and affiliated organization. The role allocation and management unit assigns different roles to different users to control their access and operation scope in the system. The role management unit assigns users to different roles, each role has a set of specific permissions, which helps to simplify permission management, especially when there are multiple users in the system. The user notification and communication unit sends notifications to users, such as system updates, new function releases, or important information related thereto. The password retrieval and account recovery unit is used by users to retrieve forgotten passwords or recover accounts to reduce access problems caused by identity authentication problems;
[0038] The role permission management module includes the role definition unit, the permission assignment unit, the role inheritance unit and the user group management unit. The role definition unit defines different user roles in the system, such as doctors, technicians, administrators, etc. Each role represents a specific set of responsibilities and operating permissions. The permission assignment unit grants specific operations and access rights to specific roles. This includes permissions to system functions, data and operations, such as model training and key point annotation. The role inheritance unit implements the role inheritance mechanism so that higher-level roles automatically inherit the permissions of lower-level roles. This simplifies permission management and ensures that higher-level users have broader permissions. The user group management unit allows administrators to group users and assign specific roles and permissions to the entire user group, which makes it easier to uniformly manage permissions for a group of users.
[0039] Among them, the working steps of the YOLOv7-pose key point detection model are as follows:
[0040] 1. Input image preprocessing: First, the model receives the input image and performs preprocessing, which includes resizing the image to the required size and performing normalization operations to ensure that the image can be effectively processed in the model;
[0041] 2. Feature extraction: On the preprocessed images, the model uses a convolutional neural network (CNN) for feature extraction. The network structure of the model contains specially designed layers to better capture the key point information of the image;
[0042] 3. Key point prediction: On the extracted feature map, the model predicts key points. Each key point consists of coordinate information and a confidence score associated with it. The goal of the model is to accurately locate key points in medical images.
[0043] 4. Loss function definition: In order to optimize the training process, the model defines a loss function, which takes into account the error of the key point position and the error of the confidence score. By continuously optimizing the loss function, the model learns how to predict the key points of the human body more accurately;
[0044] 5. Training and optimization: Using the training data with labeled key points, the model begins training. At this stage, the model fine-tunes weights and parameters through multiple iterations using optimization algorithms such as gradient descent to better adapt to the key point detection task;
[0045] 6. Evaluation and tuning: When model training is completed, the model is evaluated using a validation set or a test set. The evaluation process includes analyzing performance indicators such as key point positioning accuracy, precision, and recall. Through this process, the model can be adjusted to improve its effectiveness in practical applications.
[0046] The present invention also proposes a pelvic medical image key point detection method based on deep learning, comprising the following steps:
[0047] 1. First, the administrator assigns a role to the user, and the user obtains system usage permissions through the role;
[0048] 2. Users collect training data through the data collection module, perform data cleaning and preprocessing, and desensitize the data;
[0049] 3. The user annotates key points. Before annotating key points, the detected area needs to be annotated. Area annotation provides the system with more detailed information, which helps the model understand the image content more comprehensively.
[0050] 4. Perform model training and use labeled data to train the deep learning model so that it can accurately detect key points in medical images;
[0051] 5. Image reasoning calculation module, which uses the trained model to predict the newly input medical images and obtain the calculation results.
[0052] The following is combined with Figure 1 The following table explains the key point annotations for the pelvic medical image in:
[0053]
[0054]
[0055] (1) Acetabular prosthesis abduction angle: the angle between the line connecting the outermost edge of the acetabular prosthesis and the innermost edge of the acetabular prosthesis and the line connecting the lower edge of the ischial tuberosity on both sides. (The acetabular prosthesis abduction angle in the left replacement model is the angle between the line connecting the L_P3 point and the L_P4 point and the line connecting the L_P8 point and the L_P9 point; the calculation of the acetabular prosthesis abduction angle in the right replacement model and the bilateral replacement model is similar to that in the left replacement model)
[0056] (2) Acetabular prosthesis anteversion angle: The acetabular prosthesis anteversion angle is calculated according to the Lewis formula AVA = arcs in (r / R), where r is one times the perpendicular line from the lower edge of the short axis of the acetabular prosthesis cup opening to the line connecting the outermost edge of the acetabular prosthesis and the innermost edge of the acetabular prosthesis, and R is the distance between the outermost edge of the acetabular prosthesis and the innermost edge of the acetabular prosthesis. (The acetabular prosthesis anteversion angle in the left replacement model is: R = the distance between the L_P3 point and the L_P4 point; r = twice the distance between the L_P5 point and the perpendicular line connecting the L_P3 point and the L_P4 point; the calculation of the acetabular prosthesis anteversion angle in the right replacement model and the bilateral replacement model is similar to that of the left replacement model)
[0057] (3) Vertical position of the rotation center: the vertical distance from the center of the femoral head (or the center of the femoral head prosthesis) to the line connecting the lower edges of the teardrops on both sides. (The vertical position of the rotation center on both sides in the left replacement model is the vertical distance from the line connecting the L_P1 point and the L_P2 point to the L_P6 point and the L_P7 point respectively; the calculation of the vertical position of the rotation center in the right replacement model and the bilateral replacement model is similar to that of the left replacement model)
[0058] (4) Horizontal position of the rotation center: the distance from the projection of the center of the femoral head (or the center of the femoral head prosthesis) on the line connecting the lower edges of the teardrop on both sides to the lower edge of the teardrop on the same side. (The horizontal position of the rotation center on both sides in the left replacement model is: the distance from the vertical projection point of point L_P1 on the line connecting points L_P6 and L_P7 to point L_P7; the distance from the vertical projection point of point L_P2 on the line connecting points L_P6 and L_P7 to point L_P6; the calculation of the horizontal position of the rotation center in the right replacement model and bilateral replacement model is similar to that of the left replacement model)
[0059] (5) Femoral eccentricity: the vertical distance from the center of the femoral head to the proximal femoral axis (the proximal femoral axis is the line connecting the midpoint of the medullary cavity at the level of the lower edge of the lesser trochanter and the midpoint of the medullary cavity at the level 5 cm below the lower edge of the lesser trochanter). (The femoral eccentricity on both sides of the left replacement model is: the vertical distance from the line connecting the L_P1 point to the L_P14 point and the L_P15 point; the vertical distance from the L_P2 to the L_P16 and L_P17 points; the calculation of the femoral eccentricity of the right replacement model and the bilateral replacement model is similar to that of the left replacement model)
[0060] (6) Bilateral leg length difference: draw parallel lines through the most prominent points of the middle of the lesser trochanters on both sides and parallel to the line connecting the lower edge of the teardrop on both sides. The distance between the two parallel lines is defined as the difference in leg length between the two sides. (The difference in leg length between the two sides in the left replacement model is: draw parallel lines through L_P10 and L_P11 and parallel to the line connecting L_P8 and L_P9 respectively. The distance between these two parallel lines is the LLD; the calculation of the difference in leg length between the two sides in the right replacement model and the bilateral replacement model is similar to that in the left replacement model)
[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A pelvic medical image key point detection system based on deep learning, characterized in that: include: Data acquisition module, the data acquisition module consists of an image acquisition and storage unit, a data cleaning and preprocessing unit and a data desensitization unit. The image acquisition and storage unit is responsible for acquiring pelvic medical image data from different medical image sources, including different types of medical images such as X-ray, MRI, and CT, and storing them in the system's database. The data cleaning and preprocessing unit cleans and preprocesses the above-mentioned acquired medical images affected by noise, artifacts or other interference factors. The data desensitization unit is used to remove the patient's privacy information to prevent data leakage; The data annotation module includes a region annotation unit and a key point annotation unit. The region annotation unit is used to annotate the detected region to provide the system with more detailed information. The key point annotation unit is used to annotate the key points in the annotated region. A model training module, which is used to train a deep learning model using labeled data so that it can accurately detect key points in medical images. The model training module includes a data preparation unit, a model training unit, a model verification and tuning unit, and a model preservation unit; An image reasoning calculation module, which includes an image input and preprocessing unit, a model loading unit, a forward propagation unit, and a result output unit; User management module, which is used to manage user information, permissions and access control in the system. The user management module includes a user registration and authentication unit, a user information management unit, a role allocation and management unit, a role management unit, a user notification and communication unit, and a password retrieval and account recovery unit; The role authority management module includes a role definition unit, a authority allocation unit, a role inheritance unit and a user group management unit.
2. A pelvic medical image key point detection system based on deep learning according to claim 1, characterized in that: The data preparation unit is used to import the labeled data to be trained and divide it into training set, validation set and test set. At the same time, the training data set is expanded by rotation, scaling and flipping to improve the robustness of the model. The model training unit uses the prepared training data set to train the YOLOv7-pose network model. During the training process, the model continuously adjusts parameters through back propagation and optimization algorithms to minimize the loss function. The model verification and tuning unit uses the verification set to evaluate the trained model and adjust the model's hyperparameters to further improve performance and prevent overfitting, ensuring that the model performs well on unseen data. After the training is completed, the model saving unit saves the trained model and names it as the YOLOv7-pose key point detection model for subsequent reasoning and application on new data.
3. A pelvic medical image key point detection system based on deep learning according to claim 1, characterized in that: The image input and preprocessing unit receives the medical image to be detected as input, the model loading unit loads the trained key point detection model, the forward propagation unit inputs the input medical image into the loaded model, performs the forward propagation operation, and thus generates the model's prediction for the position of each key point. The result output unit outputs the result of the key point detection, draws the key point position on the image and provides the coordinates of the key point as output, and outputs the calculation results of the corresponding parameters.
4. A pelvic medical image key point detection system based on deep learning according to claim 1, characterized in that: The user registration and authentication unit allows users to register the system and perform identity authentication to ensure the legitimacy of their identity. The user information management unit maintains the basic information of each registered user, including name, contact information, and affiliated organization. The role allocation and management unit assigns different roles to different users to control their access and operation scope in the system. The role management unit assigns users to different roles. The user notification and communication unit sends notifications to users. The password retrieval and account recovery unit is used for users to retrieve forgotten passwords or recover accounts to reduce access problems caused by identity authentication problems.
5. A pelvic medical image key point detection system based on deep learning according to claim 1, characterized in that: The role definition unit defines different user roles in the system. Each role represents a specific set of responsibilities and operating permissions. The permission allocation unit grants specific operations and access permissions to specific roles, including permissions to system functions, data, and operations. The role inheritance unit implements the role inheritance mechanism so that higher-level roles automatically inherit the permissions of lower-level roles. The user group management unit allows administrators to group users and assign specific roles and permissions to the entire user group.
6. A pelvic medical image key point detection system based on deep learning according to claim 2, characterized in that: The working steps of the YOLOv7-pose key point detection model are as follows:
1. Input image preprocessing: First, the model receives the input image and performs preprocessing, which includes resizing the image to the required size and performing normalization operations to ensure that the image can be effectively processed in the model; 2. Feature extraction: On the preprocessed images, the model uses a convolutional neural network (CNN) for feature extraction. The network structure of the model contains specially designed layers to better capture the key point information of the image; 3. Key point prediction: On the extracted feature map, the model predicts key points. Each key point consists of coordinate information and a confidence score associated with it. The goal of the model is to accurately locate key points in medical images.
4. Loss function definition: In order to optimize the training process, the model defines a loss function, which takes into account the error of the key point position and the error of the confidence score. By continuously optimizing the loss function, the model learns how to predict the key points of the human body more accurately; 5. Training and optimization: Using the training data with labeled key points, the model begins training. At this stage, the model fine-tunes weights and parameters through multiple iterations using the gradient descent optimization algorithm to better adapt to the key point detection task; 6. Evaluation and tuning: When model training is completed, the model is evaluated using a validation set or a test set. The evaluation process includes but is not limited to analyzing the positioning accuracy, precision, and recall performance indicators of key points.
7. A method for detecting key points of pelvic medical images based on deep learning, characterized in that: The following steps are involved:
1. First, the administrator assigns a role to the user, and the user obtains system usage permissions through the role; 2. Users collect training data through the data collection module, perform data cleaning and preprocessing, and desensitize the data; 3. The user annotates key points. Before annotating key points, the detected area needs to be annotated. Area annotation provides the system with more detailed information, which helps the model understand the image content more comprehensively.
4. Perform model training and use labeled data to train the deep learning model so that it can accurately detect key points in medical images; 5. Image reasoning calculation module, which uses the trained model to predict the newly input medical images and obtain the calculation results.
Citation Information
Cited By
3D target detection method based on YOLOv7-pose
CN119888441A
Femoral head necrosis detection method and system
CN121904035A
