Patient body position intelligent monitoring system based on deep learning

Abstract

The invention relates to the technical field of medical monitoring, in particular to a patient body position intelligent monitoring system based on deep learning, which comprises a data acquisition module, a preprocessing module, a body position recognition model construction and training module, a body position monitoring module, an early warning module and a database module. The data acquisition module acquires body position image data of a radiotherapy patient; the preprocessing module cleans and standardizes the data; the body position recognition model is trained according to the labeled data and the requirements of a doctor; the body position monitoring module inputs the preprocessed data into a model to obtain an identification result, and compares the identification result with a standard body position to monitor consistency and a treatment position; the early warning module gives out early warning when abnormity occurs; the database module stores various data. The system can improve the radiotherapy precision, guarantee the safety of patients, improve the working efficiency, optimize the radiotherapy process, provide data support for medical research and promote medical informatization, and has important clinical value and wide application prospects.

Description

Technical Field

[0001] The present invention relates to the field of medical monitoring technology, and in particular to a patient posture intelligent monitoring system based on deep learning. Background Art

[0002] During radiotherapy, the accuracy of the patient's position is crucial to the effectiveness of radiotherapy. Traditional patient position monitoring methods mainly rely on visual observation and manual measurement by medical staff, which is not only inefficient but also easily affected by human factors, resulting in inaccurate monitoring results. In addition, with the continuous development of radiotherapy technology, the requirements for the accuracy and real-time performance of patient position monitoring are becoming increasingly higher.

[0003] In recent years, deep learning technology has achieved remarkable results in image recognition and other fields, providing new ideas and methods for intelligent patient position monitoring. However, the existing deep learning-based position monitoring system still has some problems in practical applications, such as the accuracy and robustness of the model needs to be improved, and the adaptability to different patients and radiotherapy scenarios is insufficient. Therefore, a deep learning-based patient position intelligent monitoring system is proposed to address the above problems. Summary of the invention

[0004] The purpose of the present invention is to provide a patient posture intelligent monitoring system based on deep learning to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A patient posture intelligent monitoring system based on deep learning, comprising:

[0007] Data acquisition module: used to collect the patient's body position image data during radiotherapy;

[0008] Data preprocessing module: connected to the data acquisition module to perform data cleaning and standardization on the body image data;

[0009] Body position recognition model construction and training module: Based on the deep learning algorithm, an intelligent real-time patient position recognition model is constructed. The body position image data of radiotherapy patients marked with accurate body position information and the body position requirements specified in the radiotherapy positioning application form issued by the doctor are used as constraints to train the recognition model so that it can recognize the patient's body position;

[0010] Body position monitoring module: connects to the data preprocessing module and the body position recognition model building and training module respectively, inputs the preprocessed body position image data into the trained body position recognition model to obtain the body position recognition result, and monitors the consistency between the patient's body position during simulation positioning and the daily treatment position by comparing the current body position with the standard body position in the radiotherapy positioning application form, and whether the patient is in the correct treatment position during treatment;

[0011] Early warning module: connected to the body position monitoring module, when the body position monitoring module detects that the patient's body position is inconsistent or not in the correct treatment position and the deviation exceeds the preset threshold, an early warning message is issued;

[0012] Database module: used to store body position image data, pre-processed body position image data, body position recognition result data and warning record data.

[0013] As a preferred solution, the image data set collected by the data acquisition module is ,in, Represents each frame of body position image data, and the dimension of the image data is ,in, Indicates the number of pixels in the vertical direction of the image data. Indicates the number of pixels in the horizontal direction of the image data. Indicates the number of channels of image data. The sampling frequency of the acquisition device is set to The value range is set according to the real-time demand for monitoring body position changes during radiotherapy.

[0014] As a preferred solution, the data cleaning operation of the data preprocessing module includes removing noise, blur and abnormal pixels caused by external interference in the image, and the standardization processing includes uniformly standardizing the resolution, brightness and contrast of the image. The specific algorithm is as follows:

[0015] Resolution standardization: Assume that the image data The native resolution is ;in, Indicates the number of pixels in the vertical direction of the original image data. Represents the number of pixels in the horizontal direction of the original image data; the resolution after standardized processing is By adjusting the function ,in, Indicates the scaling factor in the height direction, Indicates the scaling factor in the width direction;

[0016] Brightness standardization: Assume that the brightness value set of the original image is , the contrast ratio is ; The set of brightness values ​​after standardization is , the contrast ratio is ;Brightness is adjusted by the function Make corresponding adjustments, including: represents the original brightness mean, represents the original brightness standard deviation, represents the mean target brightness, represents the standard deviation of target brightness;

[0017] Contrast normalization: using contrast adjustment function Adjust the contrast so that the image data meets the model input requirements, where: Adjust the scale factor for contrast.

[0018] As a preferred solution, in the posture recognition model construction and training module, an adaptive convolutional neural network CNN is used in combination with a deconstructed human posture key point regression method during the training process. The specific algorithm formula is as follows:

[0019] Convolutional layer feature extraction: Assume that the input preprocessed image data is , the feature map after the convolution layer is expressed as , through adaptive convolution kernel According to the convolution operation Perform feature extraction, where * represents the convolution operation. is the bias term; for the convolution kernel The size is set to ,in, represents the height of the convolution kernel, Represents the width of the convolution kernel and the step size is , filled with , then the size of the output feature map is and It can be calculated according to the following formula: ; ;

[0020] Deconstructed human posture key point regression: Assume that the real coordinate set of human posture key points is ;in, is the number of key points, and the key point coordinate set predicted by the model is , loss function The mean square error (MSE) is used to measure the difference between the predicted coordinates and the true coordinates. The calculation formula is: ;

[0021] Activation function: In the convolutional neural network, in order to introduce nonlinear factors, the ReLU activation function is used to process the output of the convolution layer. Suppose the input is , the output is , and its calculation formula is: ;

[0022] Pixel-level spatial transformer part: let the input feature map After linear transformation, the query is obtained key value , whose dimensions are , the calculation formulas are: ; ; ;in, is the corresponding learnable weight matrix; then the attention score is calculated , the formula is: ,in, Represents the query vector With key vector Perform matrix multiplication by transposing ;

[0023] Then the attention weight is obtained through softmax normalization ; Finally, the output feature is obtained by weighted summation ;

[0024] The posture recognition model uses the back propagation algorithm during the training phase to update the model parameters based on the calculated loss value. ,in, Represents all learnable parameters of the model so that the model can be continuously optimized.

[0025] As a preferred solution, when the body position monitoring module determines the patient's body position deviation, the currently identified patient's body position is expressed as The standard position in the radiotherapy positioning application form is expressed as , by calculating the difference between the two To determine whether the body position is consistent, the difference measurement uses the Euclidean distance. For the key point position in the two-dimensional plane coordinates, if the number of key points is , the Euclidean distance calculation formula is: ;in, for Middle The coordinates of the key points, for Middle The coordinates of the key points, when When the body position is determined to be inconsistent, is the preset threshold, The value is pre-set according to the accuracy requirements of radiotherapy and the characteristics of the site.

[0026] As a preferred solution, when the body position monitoring module determines the patient's body position deviation, the key parts to be compared include the position and angle relationship of the patient's chest, shoulders, and head relative to the radiotherapy bed. The chest position coordinate is , the shoulder position coordinates are The head position coordinates are , the angle relationship can be expressed by the vector angle between the parts, and the vector connecting the head and chest and any fixed axial vector of the radiotherapy bed The angle is , the angle calculation formula is:

[0027] ; Position deviation of chest, shoulders and head and angular deviation Thresholds can be flexibly set according to the radiotherapy site and accuracy requirements , , , , when or or or When the position deviation of the part exceeds the threshold, it is determined that the overall position is inconsistent.

[0028] As a preferred solution, when the body position recognition model is constructed and trained, the labeled body position image data is divided into training data sets. , validation dataset and test dataset , perform model training, verification and performance testing according to the preset ratio until the model reaches the preset accuracy on the test data set and performance index requirements, the accuracy and performance index requirements include that the accuracy of automatic patient position recognition is not less than the preset value, and the recognition time is not more than the preset duration;

[0029] Assume the training round is , in each round, based on the training data set Calculate the loss function through forward propagation , and then use back propagation to update the model parameters , and then in the validation data set Evaluate the model performance on the test data set, adjust the training strategy based on the verification results, and finally Calculate the accuracy ,in, The number of samples that correctly identify body positions on the test dataset, is the total number of samples in the test data set, when And identify time When , the model is judged to meet the performance index requirements, where is the preset accuracy value, The preset duration.

[0030] It can be seen from the technical solutions provided by the present invention that the patient posture intelligent monitoring system based on deep learning provided by the present invention has the following beneficial effects:

[0031] Improving radiotherapy accuracy:

[0032] The present invention adopts a body position recognition model constructed by a deep learning algorithm, which can accurately identify the coordinates of key points of the patient's body position. During radiotherapy, slight deviations in the patient's body position will affect the effect. The system accurately analyzes the body position imaging data, promptly detects and corrects deviations, ensures that the radiotherapy rays accurately irradiate the tumor site, reduces damage to surrounding normal tissues, improves the accuracy of radiotherapy, and ensures the treatment effect.

[0033] During radiotherapy for brain tumors, the system can accurately monitor the position and angle of the patient's head to ensure that the radiotherapy rays are accurately focused on the tumor area, avoid deviations in radiotherapy doses due to changes in head position, and improve the success rate of treatment.

[0034] Ensure patient safety:

[0035] The body position monitoring module monitors the patient's body position changes in real time. If the body position is inconsistent or the deviation exceeds the preset threshold, the early warning module will immediately issue an early warning message. Medical staff will adjust the patient's body position in time according to the early warning information to prevent radiotherapy complications caused by improper body position and ensure patient safety.

[0036] If the patient moves during radiotherapy and the chest position deviates too much, it may affect the radiotherapy dose to the heart or lungs. The system will issue a timely warning, and medical staff can suspend radiotherapy and readjust the body position to avoid unnecessary damage to important organs.

[0037] Improve work efficiency:

[0038] Traditional manual checking of patient posture is inefficient and prone to errors; the intelligent monitoring system of the present invention automatically collects, processes and analyzes posture image data, quickly and accurately determines whether the posture meets the requirements, and reduces the workload of medical staff; at the same time, the real-time monitoring function of the system enables medical staff to focus more on other medical work, improving overall work efficiency;

[0039] Taking the radiotherapy department as an example, before using this system, it took a long time for medical staff to manually check the patient's body position, which might take several minutes for each patient. After using this system, body position monitoring and identification can be completed in a short time, and the body position check time for each patient can be shortened to several seconds, which significantly improves the work efficiency of the department.

[0040] Optimizing radiotherapy process:

[0041] The application of this system makes the radiotherapy process more standardized and intelligent. From the moment the patient enters the radiotherapy room, the data acquisition module automatically collects body position image data, which is processed by the data preprocessing module and the body position recognition model. The body position monitoring module monitors in real time, and the early warning module warns in time, forming a complete closed-loop monitoring system. This helps optimize the radiotherapy process, reduce radiotherapy interruptions and delays caused by body position problems, and improve the efficiency of medical resource utilization.

[0042] In the traditional radiotherapy process, if there is a problem with the patient's position, medical staff need to manually adjust and reposition the scan, wasting time and resources; this system can detect and adjust position problems in a timely manner, making the radiotherapy process smoother and more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a schematic diagram of the overall structure of a patient posture intelligent monitoring system based on deep learning in the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0045] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0046] like Figure 1 As shown, an embodiment of the present invention provides a patient posture intelligent monitoring system based on deep learning, including a data acquisition module, a data preprocessing module, a posture recognition model construction and training module, a posture monitoring module, an early warning module and a database module.

[0047] In this embodiment, the data acquisition module is used to collect the patient's body position image data during radiotherapy. Specifically, the data acquisition module collects data in the following steps:

[0048] Device initialization:

[0049] Turn on the data acquisition device (such as a high-definition camera, infrared camera, etc.), and set the device parameters, including but not limited to the resolution (set to ,in, is the number of pixels in the vertical direction, is the number of pixels in the horizontal direction), frame rate (i.e. sampling frequency , unit: frames / second), contrast, brightness, etc., to ensure that the equipment is in normal working condition;

[0050] Establishing communication connection with subsequent data processing module and storage module;

[0051] Spatial positioning and calibration:

[0052] Determine the positional relationship between the radiotherapy equipment and the patient in space, and use a coordinate system (e.g., establish a three-dimensional coordinate system with the center of the radiotherapy bed as the origin) to spatially locate and calibrate the camera to ensure that the collected image data can accurately reflect the patient's position changes during radiotherapy;

[0053] Calculate the camera's shooting angle and coverage to ensure that the patient's activity area on the radiotherapy bed can be effectively captured. Suppose the camera's shooting angle is , covering the range of (It can be expressed in terms of area or angle range, etc.) These parameters are determined through geometric calculation and actual measurement to meet the data collection requirements;

[0054] Image data acquisition:

[0055] According to the set sampling frequency , continuously collect the patient's body position image data during radiotherapy; suppose the collected image data set is ,in, Represents each frame of body position image data, and the dimension of the image data is ( Indicates the number of pixels in the vertical direction of the image data. Indicates the number of pixels in the horizontal direction of the image data. Represents the number of channels of image data, for example, for color images , grayscale image );

[0056] In order to ensure the integrity and accuracy of the data, some data verification algorithms can be used during the acquisition process, such as calculating the checksum for each frame of image data. (You can use a simple hash algorithm or other verification algorithm), the formula is: , verify whether the data is erroneous or lost through verification during subsequent transmission and storage;

[0057] Data transmission and storage:

[0058] The collected image data is transmitted to the data preprocessing module in real time for subsequent processing; during the transmission process, the image data can be compressed using a data compression algorithm (such as JPEG, H.264, etc.) to reduce the amount of data transmission and storage space occupied; the compressed image data is assumed to be , the compression rate is ,but ,in, represents the compression function;

[0059] At the same time, the original uncompressed image data is stored in the database module for subsequent query, analysis and backtracking; the image data can be organized and managed in chronological order or other reasonable ways during storage, such as adding a timestamp to each frame of image data. , storing image data as files or data records named with timestamps.

[0060] In this embodiment, the data preprocessing module is connected to the data acquisition module to perform data cleaning and standardization processing on the body image data; specifically, the following operation steps are included:

[0061] Data reception:

[0062] Receive the collected body position image data set from the data acquisition module ;

[0063] Data cleaning:

[0064] Remove noise: Use filtering algorithms (such as median filtering, Gaussian filtering, etc.) to remove noise from each frame of image data. Processing is performed to remove noise points in the image; taking median filtering as an example, the filter window size is , for image data The median coordinate is The pixel point, its filtered value For centered The median of the pixel values ​​in the neighborhood, that is ;

[0065] Deblurring: Use image sharpening algorithms (such as Laplace operator, high-pass filtering, etc.) to enhance the clarity of the image and reduce blur. Take the Laplace operator as an example, image data Perform convolution operation and set the Laplace operator kernel to be , then the processed image data ,in, Represents the convolution operation;

[0066] Remove abnormal pixels: Detect and remove abnormal pixels caused by external interference by setting thresholds and other methods; for example, calculate image data The difference between the gray value of each pixel and the gray value of the surrounding pixels in the image. If the difference is greater than the set threshold , then mark the pixel as an abnormal pixel and remove or correct it;

[0067] Standardization:

[0068] Resolution standardization: Assume that the image data The native resolution is ;in, Indicates the number of pixels in the vertical direction of the original image data. Represents the number of pixels in the horizontal direction of the original image data; the resolution after standardized processing is By adjusting the function ,in, Indicates the scaling factor in the height direction, Indicates the scaling factor in the width direction;

[0069] Brightness standardization: Assume that the brightness value set of the original image is , the contrast ratio is ; The set of brightness values ​​after standardization is , the contrast ratio is ;Brightness is adjusted by the function Make corresponding adjustments, including: represents the original brightness mean, represents the original brightness standard deviation, represents the mean target brightness, represents the standard deviation of target brightness;

[0070] Contrast normalization: using contrast adjustment function Adjust the contrast so that the image data meets the model input requirements, where: Adjust the scale factor for contrast;

[0071] Data output:

[0072] The body position image data after data cleaning and standardization are output to the body position recognition model construction and training module and the body position monitoring module for subsequent processing and analysis.

[0073] In this embodiment, the body position recognition model construction and training module constructs a patient body position intelligent real-time recognition model based on a deep learning algorithm, and uses radiotherapy patient body position image data marked with accurate body position information and the body position requirements specified in the radiotherapy positioning application form issued by the doctor as constraints to train the recognition model so that it can recognize the patient's body position; the specific operation steps are as follows:

[0074] Model architecture design:

[0075] It was determined to use an adaptive convolutional neural network (CNN) combined with a deconstructed human posture key point regression method to build an intelligent real-time patient posture recognition model;

[0076] Design the network structure of the model, including the number and parameter settings of convolutional layers, pooling layers, and fully connected layers; for example, set up multiple convolutional layers, each of which uses convolution kernels of different sizes (such as 3×3, 5×5, etc.) for feature extraction, and the number of convolution kernels gradually increases with the depth of the network; appropriately add pooling layers (such as maximum pooling layers) between convolutional layers for downsampling, reduce data dimensions and extract main features; finally, add a fully connected layer for classification or regression prediction;

[0077] Data preparation:

[0078] Collect radiotherapy patient position imaging data with accurate position information. These data should include images of different patients, different positions, and different radiotherapy stages.

[0079] At the same time, the body position requirements specified in the radiotherapy positioning application form issued by the doctor are obtained and used as constraints for model training;

[0080] Divide the image data into training datasets , validation dataset and test dataset , usually divided according to a certain proportion (such as 70%, 20%, 10%);

[0081] Model initialization:

[0082] Initialize the parameters of the model, such as randomly initializing the convolution kernel weights and bias terms;

[0083] Training process:

[0084] Forward propagation:

[0085] For each sample in the training data set, the preprocessed image data x is input into the model, and forward propagation calculation is performed through each layer of the model to obtain the output prediction value of the model;

[0086] In the convolution layer, the adaptive convolution kernel According to the convolution operation Perform feature extraction, where represents the convolution operation, is the bias term; for the convolution kernel The size is set to , the step length is , filled with , then the size of the output feature map is and According to the formula and calculate;

[0087] After a series of convolutions, pooling, and activation functions (such as the ReLU activation function ) is processed, and finally the predicted body position key point coordinate set is obtained through the fully connected layer ;

[0088] Calculate the loss function:

[0089] Assume that the real coordinate set of the key points of human posture is , loss function The mean square error (MSE) is used to measure the difference between the predicted coordinates and the true coordinates. The calculation formula is: ,in, is the number of key points, Represents all learnable parameters of the model;

[0090] Backward Propagation:

[0091] According to the calculated loss function value, the back propagation algorithm is used to calculate the gradient of each layer parameter;

[0092] Using the chain rule, starting from the output layer, the partial derivative of each parameter with respect to the loss function is calculated in turn to obtain the gradient update value of the parameters of each layer;

[0093] Parameter update:

[0094] Based on the calculated gradient update value, use an optimization algorithm (such as stochastic gradient descent, Adam optimizer, etc.) to update the model parameters; for example, when using the Adam optimizer, adjust the parameters according to its update rules so that the loss function gradually decreases;

[0095] Model evaluation and optimization:

[0096] During training, periodically re-validate the validation dataset. Evaluate the performance of the model and calculate indicators such as loss value and accuracy on the validation set;

[0097] Adjust the model's hyperparameters (such as learning rate, convolution kernel size, number of network layers, etc.) based on the validation results to optimize model performance. For example, if the loss value on the validation set no longer decreases or the accuracy no longer increases, you may need to reduce the learning rate or adjust other hyperparameters.

[0098] Model testing:

[0099] After the model converges on the training data set, use the test data set Test the final trained model, evaluate the model's generalization ability on unseen data, and calculate the accuracy on the test set ,in, The number of samples that correctly identify body positions on the test dataset, is the total number of samples in the test data set;

[0100] Model deployment:

[0101] The trained model is deployed to the actual patient position intelligent monitoring system for real-time patient position identification.

[0102] In this embodiment, the body position monitoring module is connected to the data preprocessing module and the body position recognition model construction and training module respectively, and the preprocessed body position image data is input into the trained body position recognition model to obtain the body position recognition result, and the consistency of the patient's body position during simulation positioning and the daily treatment position is monitored by comparing the current body position with the standard body position in the radiotherapy positioning application form, and whether the patient is in the correct treatment position during treatment; the specific steps are as follows:

[0103] Data reception:

[0104] receiving the preprocessed body image data from the data preprocessing module;

[0105] Obtaining body position recognition results:

[0106] Input the received body position image data into the trained body position recognition model;

[0107] Through the forward propagation calculation of the model, the recognition result of the current body position is obtained, that is, the predicted body position key point coordinate set and other information;

[0108] Standard body position acquisition:

[0109] Obtain the standard body position information specified by the doctor from the radiotherapy positioning application form, including the standard body position key point coordinate set, etc.;

[0110] Posture consistency monitoring:

[0111] Calculate the difference between the current body position and the standard body position, which can be done by using methods such as Euclidean distance;

[0112] By comparing the difference measure with the preset threshold, it is determined whether the body position is consistent. If the difference measure is less than or equal to the threshold, it is determined that the current body position and the standard body position are well consistent during simulation positioning or daily treatment; if the difference measure is greater than the threshold, it is determined that the body position is inconsistent;

[0113] Treatment Position Monitoring:

[0114] During the treatment, the patient's body position is continuously monitored to see if it is in the correct treatment position, also by calculating the difference between the current body position and the standard treatment position and comparing it with the corresponding threshold;

[0115] If the patient's body position deviates from the correct treatment position and the deviation exceeds the preset range, record the relevant information and proceed to the next step;

[0116] Result output:

[0117] Outputting the body position recognition results, body position consistency judgment results and treatment position monitoring results to the early warning module and the database module;

[0118] Among them, when determining the patient's body position deviation, the currently identified patient's body position is expressed as The standard position in the radiotherapy positioning application form is expressed as , by calculating the difference between the two To determine whether the body position is consistent, the difference measurement uses the Euclidean distance. For the key point position in the two-dimensional plane coordinates, if the number of key points is ,The Euclidean distance calculation formula is: ;in, for Middle The coordinates of the key points, for Middle The coordinates of the key points, when When the body position is determined to be inconsistent, is the preset threshold, The value of is preset according to the accuracy requirements of radiotherapy and the characteristics of the site;

[0119] When the body position monitoring module determines the patient's body position deviation, the key parts to be compared include the position and angle relationship of the patient's chest, shoulders, and head relative to the radiotherapy bed. The chest position coordinates are set as , the shoulder position coordinates are The head position coordinates are , the angle relationship can be expressed by the vector angle between the parts, and the vector connecting the head and chest and any fixed axial vector of the radiotherapy bed The angle is , the angle calculation formula is:

[0120] ; Position deviation of chest, shoulders and head and angular deviation Thresholds can be flexibly set according to the radiotherapy site and accuracy requirements , , , , when or or or When the position deviation of the part exceeds the threshold, it is determined that the overall position is inconsistent;

[0121] At the same time, the weights of each key part can be set according to different radiotherapy plans and individual differences of patients, and the position deviation can be comprehensively judged by weighted summation. The weights of each part are , comprehensive deviation measure The calculation formula can be: ;

[0122] when ( The body position is determined to be inconsistent when the value is the comprehensive threshold, thereby further improving the flexibility and accuracy of body position judgment.

[0123] In this embodiment, the warning module is connected to the body position monitoring module. When the body position monitoring module detects that the patient's body position is inconsistent or not in the correct treatment position and the deviation exceeds the preset threshold, a warning message is issued; the specific steps are as follows:

[0124] Monitoring information reception:

[0125] receiving body position monitoring results from the body position monitoring module, including body position recognition results, body position consistency judgment results, and treatment position monitoring results;

[0126] Determine whether to issue an early warning:

[0127] Check the position consistency judgment results. If the patient's position is inconsistent and the deviation exceeds the preset threshold , then enter the early warning process;

[0128] At the same time, check the treatment position monitoring results. If the patient is not in the correct treatment position and the deviation exceeds the preset threshold , and also enter the early warning process;

[0129] Warning information generation:

[0130] Generate corresponding warning information based on the abnormal situation detected. The warning information should include the patient's basic information (such as name, medical record number, etc.), the specific situation of the abnormal position (such as which key points exceed the threshold, the deviation value, etc.), and timestamp and other information;

[0131] Early warning method selection:

[0132] Select the appropriate warning method according to the system settings, such as sound warning, pop-up window prompt, SMS notification, etc.

[0133] Warning information sent:

[0134] The warning information is sent to relevant medical staff through the selected warning method to ensure that medical staff can receive and handle abnormal body position in time.

[0135] In this embodiment, the database module is used to store body position image data, pre-processed body position image data, body position recognition result data and warning record data. The specific steps are as follows:

[0136] Database initialization:

[0137] Create a database and establish a corresponding data table structure to store body position image data, pre-processed body position image data, body position recognition result data, and early warning record data;

[0138] Configure database connection parameters to ensure normal communication with other modules of the system;

[0139] Data Storage:

[0140] After the data acquisition module acquires the body position image data, it stores the original body position image data in the specified data table in the database, and adds a unique identifier (such as a timestamp or an auto-increment ID, etc.) and related patient information to each frame of image data;

[0141] After the data preprocessing module processes the body position image data, it stores the preprocessed body position image data in the corresponding data table in the database, and also associates the corresponding original image data identifier and patient information;

[0142] After the body position recognition model construction and training module obtains the body position recognition result, the body position recognition result data (such as the predicted body position key point coordinates, etc.) is stored in the database and associated with the corresponding pre-processed image data;

[0143] When the early warning module issues an early warning message, the early warning record data (including the early warning time, the reason for the early warning, the patient information, etc.) is stored in the database;

[0144] Data query and management:

[0145] Provides a data query interface, allowing other modules of the system to query data in the database as needed. For example, medical staff can query the body position imaging data, body position recognition results, and warning records of specific patients;

[0146] Regularly maintain and manage the database, such as data backup, data cleanup, etc., to ensure database stability and data security.

[0147] In this embodiment, in the posture recognition model construction and training module, an adaptive convolutional neural network CNN combined with a deconstructed human posture key point regression method is used in the training process, and its specific algorithm formula is as follows:

[0148] Convolutional layer feature extraction: Assume that the input preprocessed image data is , the feature map after the convolution layer is expressed as , through adaptive convolution kernel According to the convolution operation Perform feature extraction, where * represents the convolution operation. is the bias term; for the convolution kernel The size is set to ,in, represents the height of the convolution kernel, Represents the width of the convolution kernel and the step size is , filled with , then the size of the output feature map is and It can be calculated according to the following formula: ; ;

[0149] Deconstructed human posture key point regression: Assume that the real coordinate set of human posture key points is ;in, is the number of key points, and the key point coordinate set predicted by the model is , loss function The mean square error (MSE) is used to measure the difference between the predicted coordinates and the true coordinates. The calculation formula is: ;

[0150] Activation function: In the convolutional neural network, in order to introduce nonlinear factors, the ReLU activation function is used to process the output of the convolution layer. Suppose the input is , the output is , and its calculation formula is: ;

[0151] Pixel-level spatial transformer part: let the input feature map After linear transformation, the query is obtained key value , whose dimensions are , the calculation formulas are: ; ; ;in, is the corresponding learnable weight matrix; then the attention score is calculated , the formula is: ,in, Represents the query vector With key vector Perform matrix multiplication by transposing ;

[0152] Then the attention weight is obtained through softmax normalization ; Finally, the output feature is obtained by weighted summation ;

[0153] The posture recognition model uses the back propagation algorithm during the training phase to update the model parameters based on the calculated loss value. ,in, Represents all learnable parameters of the model so that the model can be continuously optimized.

[0154] In this embodiment, when the body position recognition model is constructed and the training module trains the recognition model, the labeled body position image data is divided into a training data set. , validation dataset and test dataset , perform model training, verification and performance testing according to the preset ratio until the model reaches the preset accuracy on the test data set and performance index requirements, the accuracy and performance index requirements include that the accuracy of automatic patient position recognition is not less than the preset value, and the recognition time is not more than the preset duration;

[0155] Assume the training round is , in each round, based on the training data set Calculate the loss function through forward propagation , and then use back propagation to update the model parameters , and then in the validation data set Evaluate the model performance on the test data set, adjust the training strategy based on the verification results, and finally Calculate the accuracy ,in, The number of samples that correctly identify body positions on the test dataset, is the total number of samples in the test data set, when And identify time When , the model is judged to meet the performance index requirements, where is the preset accuracy value, The preset duration;

[0156] During the training process, regularization techniques can also be used, such as Regularization, add a regularization term to the loss function, and set the regularization coefficient to Then add The loss function after regularization The calculation formula is: (in, Represents a set of model parameters The parameters) can be further optimized through the above method.

[0157] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A patient posture intelligent monitoring system based on deep learning, characterized in that: include: Data acquisition module: used to collect the patient's body position image data during radiotherapy; Data preprocessing module: connected to the data acquisition module, to perform data cleaning and standardization processing on the body position image data; Body position recognition model construction and training module: Based on the deep learning algorithm, an intelligent real-time patient position recognition model is constructed. The patient position image data of radiotherapy patients marked with accurate position information and the position requirements specified in the radiotherapy positioning application form issued by the doctor are used as constraints to train the recognition model so that it can recognize the patient position; Body position monitoring module: connected to the data preprocessing module and the body position recognition model construction and training module respectively, inputs the preprocessed body position image data into the trained body position recognition model to obtain the body position recognition result, and monitors the consistency between the patient's body position during simulation positioning and the daily treatment body position by comparing the current body position with the standard body position in the radiotherapy positioning application form, and whether the patient is in the correct treatment position during treatment; Early warning module: connected to the body position monitoring module, when the body position monitoring module detects that the patient's body position is inconsistent or not in the correct treatment position and the deviation exceeds a preset threshold, an early warning message is issued; Database module: used to store the body position image data, pre-processed body position image data, body position recognition result data and warning record data.

2. The patient posture intelligent monitoring system based on deep learning according to claim 1, characterized in that: The image data set collected by the data collection module is ,in, Represents each frame of body position image data, and the dimension of the image data is ,in, Indicates the number of pixels in the vertical direction of the image data. Indicates the number of pixels in the horizontal direction of the image data. Indicates the number of channels of image data. The sampling frequency of the acquisition device is set to The value range is set according to the real-time demand for monitoring body position changes during radiotherapy.

3. The patient posture intelligent monitoring system based on deep learning according to claim 1, characterized in that: The data cleaning operation of the data preprocessing module includes removing noise, blur and abnormal pixels caused by external interference in the image. The standardization processing includes uniformly standardizing the resolution, brightness and contrast of the image. The specific algorithm is as follows: Resolution standardization: Assume that the image data The native resolution is ;in, Indicates the number of pixels in the vertical direction of the original image data. Represents the number of pixels in the horizontal direction of the original image data; the resolution after standardized processing is By adjusting the function ,in, Indicates the scaling factor in the height direction, Indicates the scaling factor in the width direction; Brightness standardization: Assume that the brightness value set of the original image is , the contrast ratio is ; The set of brightness values ​​after standardization is , the contrast ratio is ;Brightness is adjusted by the function Make corresponding adjustments, including: represents the original brightness mean, represents the original brightness standard deviation, represents the mean target brightness, represents the standard deviation of target brightness; Contrast normalization: using contrast adjustment function Adjust the contrast so that the image data meets the model input requirements, where: Adjust the scale factor for contrast.

4. The patient posture intelligent monitoring system based on deep learning according to claim 1, characterized in that: In the posture recognition model construction and training module, an adaptive convolutional neural network CNN combined with a deconstructed human posture key point regression method is used in the training process, and the specific algorithm formula is as follows: Convolutional layer feature extraction: Assume that the input preprocessed image data is , the feature map after the convolution layer is expressed as , through adaptive convolution kernel According to the convolution operation Perform feature extraction, where * represents the convolution operation. is the bias term; for the convolution kernel The size is set to ,in, represents the height of the convolution kernel, Represents the width of the convolution kernel and the step size is , filled with , then the size of the output feature map is and It can be calculated according to the following formula: ； ； Deconstructed human posture key point regression: Assume that the real coordinate set of human posture key points is ;in, is the number of key points, and the key point coordinate set predicted by the model is , loss function The mean square error (MSE) is used to measure the difference between the predicted coordinates and the true coordinates. The calculation formula is: ; Activation function: In the convolutional neural network, in order to introduce nonlinear factors, the ReLU activation function is used to process the output of the convolution layer. Suppose the input is , the output is , and its calculation formula is: ; Pixel-level spatial transformer part: let the input feature map After linear transformation, the query is obtained key value , whose dimensions are , the calculation formulas are: ; ; ;in, is the corresponding learnable weight matrix; then the attention score is calculated , the formula is: ,in, Represents the query vector With key vector Perform matrix multiplication by transposing ; Then the attention weight is obtained through softmax normalization ; Finally, the output feature is obtained by weighted summation ; The posture recognition model uses the back propagation algorithm during the training phase to update the model parameters based on the calculated loss value. ,in, Represents all learnable parameters of the model so that the model can be continuously optimized.

5. The patient posture intelligent monitoring system based on deep learning according to claim 1, characterized in that: When the body position monitoring module determines the patient's body position deviation, the currently identified patient's body position is expressed as The standard position in the radiotherapy positioning application form is expressed as , by calculating the difference between the two To determine whether the body position is consistent, the difference measurement uses the Euclidean distance. For the key point position in the two-dimensional plane coordinates, if the number of key points is , the Euclidean distance calculation formula is: ;in, for Middle The coordinates of the key points, for Middle The coordinates of the key points, when When the body position is determined to be inconsistent, is the preset threshold, The value is pre-set according to the accuracy requirements of radiotherapy and the characteristics of the site.

6. The patient posture intelligent monitoring system based on deep learning according to claim 1, characterized in that: When the body position monitoring module determines the patient's body position deviation, the key parts to be compared include the position and angle relationship of the patient's chest, shoulders, and head relative to the radiotherapy bed. Suppose the chest position coordinate is , the shoulder position coordinates are The head position coordinates are , the angle relationship can be expressed by the vector angle between the parts, and the vector connecting the head and chest and any fixed axial vector of the radiotherapy bed The angle is , the angle calculation formula is: ； Positional deviation of chest, shoulders and head and angular deviation Thresholds can be flexibly set according to the radiotherapy site and accuracy requirements , , , , when or or or When the position deviation of the part exceeds the threshold, it is determined that the overall position is inconsistent.

7. The patient posture intelligent monitoring system based on deep learning according to claim 1, characterized in that: When the body position recognition model construction and training module trains the recognition model, the labeled body position image data is divided into training data sets. , validation dataset And the test dataset , perform model training, verification and performance testing according to the preset ratio until the model reaches the preset accuracy on the test data set and performance index requirements, wherein the accuracy and performance index requirements include that the accuracy of automatic patient position recognition is not less than a preset value, and the recognition time is not more than a preset duration; Assume the training round is , in each round, based on the training data set Calculate the loss function through forward propagation , and then use back propagation to update the model parameters , and then in the validation data set Evaluate the model performance on the test data set, adjust the training strategy based on the verification results, and finally Calculate the accuracy ,in, The number of samples that correctly identify body positions on the test dataset, is the total number of samples in the test data set, when And identify time When , the model is judged to meet the performance index requirements, where is the preset accuracy value, The preset duration.

Images