Breast cancer postoperative rehabilitation posture management system based on deep learning

Abstract

The invention discloses a breast cancer postoperative rehabilitation posture management system based on deep learning. The breast cancer postoperative rehabilitation posture management system comprises a data acquisition module, an image processing module, a breast cancer postoperative rehabilitation posture monitoring model building module, a model hyper-parameter optimization module and a breast cancer postoperative rehabilitation posture management module. The invention relates to the technical field of postoperative rehabilitation management, in particular to a breast cancer postoperative rehabilitation posture management system based on deep learning, and according to the scheme, the overall performance and practicability of the system are improved through smooth convolution kernel determination, gradient reconstruction, threshold setting and edge detection; a model is established by designing a model structure, designing an activation function, introducing channel attention, designing a loss function, calculating gradient momentum and correcting deviation, so that the monitoring precision and generalization ability of the model are improved; the model hyper-parameter is searched by designing the search sensitivity, the chaos factor, the chaos search function and the local refinement search function, and the hyper-parameter optimization efficiency and effect are improved.

Description

Technical Field

[0001] This invention relates to the field of postoperative rehabilitation management technology, specifically to a deep learning-based postoperative rehabilitation posture management system for breast cancer. Background Technology

[0002] The deep learning-based breast cancer post-operative rehabilitation posture management system utilizes deep learning and data processing technologies to process and analyze image data of breast cancer patients performing rehabilitation postures after surgery, establishing a monitoring model for real-time monitoring of patients' post-operative rehabilitation postures. However, traditional breast cancer post-operative rehabilitation posture management systems suffer from insufficient accuracy in image processing, and edge detection is susceptible to noise interference, resulting in low overall system performance and practicality. Traditional breast cancer post-operative rehabilitation posture monitoring models suffer from simple model structures, poor adaptability of activation functions, insufficient feature extraction, and inflexible learning rate adjustment, leading to low monitoring accuracy and generalization ability. Traditional model hyperparameter optimization methods suffer from low search efficiency, susceptibility to local optima, and inability to effectively balance global and local search. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a deep learning-based post-masturbation rehabilitation posture management system for breast cancer. Addressing the problems of insufficient accuracy in image processing and susceptibility to noise interference in edge detection, which lead to low overall system performance and practicality in traditional post-masturbation rehabilitation posture management systems, this solution improves image processing accuracy and enhances overall system performance and practicality by determining smooth convolution kernels, defining gradient component matrices, performing gradient reconstruction, setting thresholds, and implementing edge detection. Furthermore, it addresses the limitations of traditional post-masturbation rehabilitation posture monitoring models, such as simple model structure, poor activation function adaptability, insufficient feature extraction, and learning limitations. To address the issue of low monitoring accuracy and generalization ability of models due to inflexible behavior adjustment, this solution improves the monitoring accuracy, generalization ability, flexibility, and adaptability of the model by designing the model structure, activation function, introducing channel attention, loss function, calculating gradient momentum, and bias correction. Furthermore, to address the problems of low search efficiency, susceptibility to local optima, and inability to effectively balance global and local searches in traditional hyperparameter optimization methods, this solution improves the efficiency and effectiveness of hyperparameter optimization by conducting search preparation, designing search sensitivity, designing chaos factors, designing chaotic search functions, and designing local refinement search functions.

[0004] The technical solution adopted by the present invention is as follows: The postoperative rehabilitation posture management system for breast cancer based on deep learning provided by the present invention includes a data acquisition module, an image processing module, a module for establishing a postoperative rehabilitation posture monitoring model for breast cancer, a model hyperparameter optimization module, and a postoperative rehabilitation posture management module for breast cancer.

[0005] The data acquisition module collects image data and posture standardization data of breast cancer patients performing rehabilitation postures after surgery. The posture standardization data includes standard posture, poor posture, non-standard posture and dangerous posture.

[0006] Furthermore, the image processing module processes the image by determining a smooth convolution kernel, defining a gradient component matrix, reconstructing the gradient, setting a threshold, performing edge detection, and image annotation, specifically including the following steps:

[0007] Determine the smooth convolution kernel and perform a smooth convolution on the image, as shown below:

[0008] ;

[0009] in, This represents the convolution kernel used for smooth convolution, where x and y represent the x-coordinate and y-coordinate indices of pixels in the image, respectively. Represents pi (π). The standard deviation represents the smoothing standard deviation, and e represents the natural constant.

[0010] Define the gradient component matrix to define the gradient components of the image in the horizontal and vertical directions, as follows:

[0011] ;

[0012] ;

[0013] in, and These represent the component matrices of the gradient of the image in the horizontal and vertical directions, respectively. To represent the square root;

[0014] Gradient reconstruction reconstructs the gradient magnitude and direction, as shown below:

[0015] ;

[0016] ;

[0017] in, This represents the magnitude of the reconstructed gradient. Indicates the reconstructed gradient direction. Represents the arctangent function;

[0018] To set a threshold, first calculate the mean and standard deviation of the image gradient magnitude. Then, use the mean and standard deviation to set the threshold, as shown below:

[0019] ;

[0020] ;

[0021] in, and These represent the high threshold and the low threshold, respectively. This represents the mean of the image gradient magnitude. The standard deviation of the image gradient magnitude;

[0022] Edge detection marks pixels with gradient magnitudes greater than a high threshold as strong edges. It examines pixels with gradient magnitudes between the low and high thresholds. If these weak edge pixels are connected to the marked strong edge pixels, they are also marked as part of the edge. Starting from the marked strong edge pixels, it connects the weak edge pixels in their neighborhood to the strong edge to build a complete edge structure.

[0023] Image annotation involves labeling image data of breast cancer patients performing rehabilitation postures after surgery using data on the standardization of these postures, and then setting the standardization data as the label data.

[0024] Furthermore, the module for establishing a post-mass breast cancer rehabilitation posture monitoring model builds the model by designing the model structure, activation function, convolutional layer, introducing channel attention, loss function, dynamic decay rate, calculating gradient momentum, bias correction, and model parameter learning. Specifically, it includes the following:

[0025] The model structure for monitoring post-mastectomy rehabilitation posture consists of the following components connected in sequence: First, the input layer is connected to the first convolutional module, which has 32 3x3 convolutional kernels. Next is the first pooling module, which follows the first convolutional module and uses max pooling with a 2x2 pooling window. Then, it is connected to the second and third convolutional modules, which have 64 3x3 convolutional kernels and 128 3x3 convolutional kernels respectively. This is followed by the second pooling layer, which uses max pooling with a 2x2 pooling window. Feature flattening is then performed, followed by two fully connected layers: the first fully connected layer has 512 neurons, and the second fully connected layer has 256 neurons. Finally, it is connected to the output layer, which outputs the monitoring results of post-mastectomy rehabilitation posture.

[0026] Design an activation function, expressed as follows:

[0027] ;

[0028] in, This represents the activation function. This represents the input to the activation function. This represents the switch weight, and log(·) represents taking the logarithm. Indicates the nonlinear factor. This represents the slope control parameter;

[0029] The convolutional layer is designed as follows:

[0030] ;

[0031] Where i and j represent the position indices of the feature maps, and n represents the layer index of the convolutional layer. This indicates that the output feature map of the nth convolutional layer is at position The values ​​at the specified positions, k and p represent the row and column indices of the convolution kernel, respectively, and K and P represent the number of rows and columns of the convolution kernel, respectively. The output feature map of layer n-1 is at position The value at that location, This indicates the position of the convolution kernel in the nth convolutional layer. The value at that location, This represents the bias of the nth convolutional layer;

[0032] Channel attention is introduced, and the input features are subjected to global average pooling and global max pooling operations in sequence, and then processed by the activation function to obtain the channel weights.

[0033] The loss function is designed as follows:

[0034] ;

[0035] in, Let d represent the loss function, and d represent the index of the data sample. This indicates the degree of standardization of the actual posture of the d-th sample. This indicates the standardity of the pose predicted by the model's output for the d-th sample. Indicates the adjustment weight. Indicates channel attention weights. This represents the expected channel attention weight. Represents the square of the L2 norm;

[0036] The dynamic attenuation rate is designed as follows:

[0037] ;

[0038] in, This indicates the number of times the model's parameters have been learned. This represents the decay rate during the iterth iteration of model parameter learning. This represents the initial decay rate, and iterM represents the maximum number of parameter learning iterations. This indicates taking the maximum value. Indicates the lower limit of decay;

[0039] The gradient momentum is calculated as follows:

[0040] ;

[0041] ;

[0042] in, and Let represent the first and second gradient momentum of the loss function with respect to the model parameters during the iterth iteration of model parameter learning, respectively. and Let represent the first and second gradient momentum of the loss function with respect to the model parameters during the (iter-1)th iteration of model parameter learning. Let represent the gradient of the loss function with respect to the model parameters during the iterth iteration of model parameter learning. Indicates the second attenuation rate. Indicates the momentum change factor;

[0043] Deviation correction is expressed as follows:

[0044] ;

[0045] ;

[0046] in, and Let represent the first and second gradient momentum of the loss function with respect to the model parameters after bias correction during the iterth model parameter learning process, respectively.

[0047] Model parameter learning is represented as follows:

[0048] ;

[0049] in, This represents the model parameters at the iterth iteration of model parameter learning. This represents the model parameters during the (iter-1)th iteration of model parameter learning. Indicates the learning rate. Indicates the learning rate scaling weights. This represents a micro-parameter.

[0050] Furthermore, the model hyperparameter optimization module searches for model hyperparameters through search preparation, designing search sensitivity, designing chaos factor, designing chaotic search function, and designing local refinement search function, specifically including the following:

[0051] Search preparation involves selecting search hyperparameters, including nonlinear factors, slope control parameters, adjustment weights, initial decay rate, and lower decay limit; creating a parameter search space by setting the reciprocal of the model's loss function value as the model parameter fit rate.

[0052] The search sensitivity is designed as follows:

[0053] ;

[0054] Where t represents the current number of hyperparameter searches, This represents the search sensitivity during the t-th hyperparameter search. Indicates the initial search sensitivity. This represents the sensitivity control factor. This represents a random number between -1 and 1.

[0055] The chaos factor is designed as follows:

[0056] ;

[0057] in, This represents the chaos factor during the t-th hyperparameter search. Let Ctl represent the chaos factor during the (t-1)th hyperparameter search, where Ctl is a random number between 3.5 and 4.5, representing the initial chaos factor. It is a random number between 0 and 1;

[0058] Design a chaotic search function, represented as follows:

[0059] ;

[0060] in, This represents the position of the hyperparameters found by the chaotic search function during the (t+1)th hyperparameter search. This indicates the position with the highest hyperparameter fit rate. This represents a directional control factor with a random direction and a size of unit length, where u represents the chaos intensity.

[0061] Design a local refinement search function, as follows:

[0062] ;

[0063] in, This represents the hyperparameter position found by the local refinement search function during the (t+1)th hyperparameter search. This represents the hyperparameter position found by the local refinement search function during the t-th hyperparameter search. This indicates taking the absolute value. This represents the position of the hyperparameters found by the chaotic search function during the t-th hyperparameter search.

[0064] Perform the search, set the hyperparameter fitness rate threshold, set the maximum number of hyperparameter searches, generate an initial hyperparameter search cluster, select the position with the highest fitness rate in the initial hyperparameter search cluster as the current position with the highest fitness rate, first perform a hyperparameter search for the chaotic search function, update the current position with the highest fitness rate, then perform a hyperparameter search for the local refinement search function, update the current position with the highest fitness rate again. If the fitness rate of a search point exceeds the hyperparameter fitness rate threshold during the search process, output the hyperparameters of the parameter point with the highest fitness rate; if the maximum number of hyperparameter searches is reached, start the search again; otherwise, increment the search count and proceed to the next round of search.

[0065] Furthermore, the breast cancer postoperative rehabilitation posture management module collects image data of breast cancer patients performing rehabilitation postures in real time and inputs the collected data into the breast cancer postoperative rehabilitation posture monitoring model module. The model outputs the standardity of the patient's posture. When the system detects that the patient's posture is a dangerous posture, it will issue a warning message to remind the patient and medical staff to pay attention to adjusting the posture.

[0066] The beneficial effects achieved by the present invention using the above solution are as follows:

[0067] (1) In view of the problems of insufficient accuracy in image processing and low overall system performance and practicality caused by noise interference in edge detection in traditional breast cancer postoperative rehabilitation posture management system, this solution improves the accuracy of image processing and enhances the overall performance and practicality of the system by determining smooth convolution kernel, defining gradient component matrix, gradient reconstruction, setting threshold and edge detection.

[0068] (2) In view of the problems of low monitoring accuracy and generalization ability of traditional breast cancer postoperative rehabilitation posture monitoring models, such as simple model structure, poor adaptability of activation function, insufficient feature extraction, and inflexible learning rate adjustment, this scheme improves the monitoring accuracy, generalization ability, flexibility and adaptability of the model by designing model structure, designing activation function, introducing channel attention, designing loss function, calculating gradient momentum and bias correction.

[0069] (3) In view of the problems of low search efficiency, easy to get trapped in local optima and inability to effectively balance global search and local search in traditional model hyperparameter optimization methods, this scheme searches for model hyperparameters by search preparation, designing search sensitivity, designing chaos factor, designing chaos search function and designing local refinement search function, thereby improving the efficiency and effect of hyperparameter optimization. Attached Figure Description

[0070] Figure 1 A schematic diagram of the deep learning-based post-breast cancer rehabilitation posture management system provided by the present invention;

[0071] Figure 2 This is a schematic diagram of the image processing module;

[0072] Figure 3 A schematic diagram of a post-operative rehabilitation posture monitoring model module for breast cancer surgery;

[0073] Figure 4 This is a schematic diagram of the model hyperparameter optimization module;

[0074] Figure 5 A schematic diagram of the fitness rate detection process for hyperparameter optimization.

[0075] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0076] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0077] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0078] Example 1, see Figure 1The deep learning-based post-breast cancer rehabilitation posture management system provided by this invention includes a data acquisition module, an image processing module, a module for establishing a post-breast cancer rehabilitation posture monitoring model, a model hyperparameter optimization module, and a post-breast cancer rehabilitation posture management module.

[0079] The data acquisition module collects image data and standardization data of breast cancer patients performing rehabilitation postures after surgery. The standardization data of the movements includes standard postures, suboptimal postures, non-standard postures, and dangerous postures.

[0080] The image processing module processes images by determining smooth convolution kernels, defining gradient component matrices, reconstructing gradients, setting thresholds, performing edge detection, and annotating images.

[0081] The module for establishing a post-operative rehabilitation posture monitoring model for breast cancer patients establishes the model by designing the model structure, activation function, convolutional layer, introducing channel attention, loss function, dynamic decay rate, calculating gradient momentum, bias correction, and model parameter learning.

[0082] The model hyperparameter optimization module searches for model hyperparameters through search preparation, designing search sensitivity, designing chaos factor, designing chaotic search function, and designing local refinement search function;

[0083] The breast cancer postoperative rehabilitation posture management module collects image data of breast cancer patients performing rehabilitation postures in real time and inputs the collected data into the breast cancer postoperative rehabilitation posture monitoring model module. The model outputs the standard of the patient's posture. When the system detects that the patient's posture is dangerous, it will issue a warning message to remind the patient and medical staff to adjust their posture.

[0084] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. The image processing module processes the image by determining the smooth convolution kernel, defining the gradient component matrix, reconstructing the gradient, setting a threshold, edge detection, and image annotation. Specifically, it includes the following steps:

[0085] Determine the smooth convolution kernel and perform a smooth convolution on the image, as shown below:

[0086] ;

[0087] in, This represents the convolution kernel used for smooth convolution, where x and y represent the x-coordinate and y-coordinate indices of pixels in the image, respectively. Represents pi (π). The standard deviation represents the smoothing standard deviation, and e represents the natural constant.

[0088] Define the gradient component matrix to define the gradient components of the image in the horizontal and vertical directions, as follows:

[0089] ;

[0090] ;

[0091] in, and These represent the component matrices of the gradient of the image in the horizontal and vertical directions, respectively. To represent the square root;

[0092] Gradient reconstruction reconstructs the gradient magnitude and direction, as shown below:

[0093] ;

[0094] ;

[0095] in, This represents the magnitude of the reconstructed gradient. Indicates the reconstructed gradient direction. Represents the arctangent function;

[0096] To set a threshold, first calculate the mean and standard deviation of the image gradient magnitude. Then, use the mean and standard deviation to set the threshold, as shown below:

[0097] ;

[0098] ;

[0099] in, and These represent the high threshold and the low threshold, respectively. This represents the mean of the image gradient magnitude. The standard deviation of the image gradient magnitude;

[0100] Edge detection marks pixels with gradient magnitudes greater than a high threshold as strong edges. It examines pixels with gradient magnitudes between the low and high thresholds. If these weak edge pixels are connected to the marked strong edge pixels, they are also marked as part of the edge. Starting from the marked strong edge pixels, it connects the weak edge pixels in their neighborhood to the strong edge to build a complete edge structure.

[0101] Image annotation involves labeling image data of breast cancer patients performing rehabilitation postures after surgery using data on the standardization of these postures, and then setting the standardization data as the label data.

[0102] By performing the above operations, this solution addresses the problems of insufficient accuracy in image processing and low overall system performance and practicality caused by noise interference in edge detection in traditional post-masturbation posture management systems for breast cancer. This solution improves image processing accuracy and enhances the overall system performance and practicality by determining smooth convolution kernels, defining gradient component matrices, gradient reconstruction, setting thresholds, and edge detection.

[0103] Example 3, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. The module for establishing a post-breast cancer rehabilitation posture monitoring model establishes the model by designing the model structure, designing the activation function, designing the convolutional layer, introducing channel attention, designing the loss function, designing the dynamic decay rate, calculating the gradient momentum, bias correction, and learning the model parameters. Specifically, it includes the following:

[0104] The model structure for monitoring post-mastectomy rehabilitation posture consists of the following components connected in sequence: First, the input layer is connected to the first convolutional module, which has 32 3x3 convolutional kernels. Next is the first pooling module, which follows the first convolutional module and uses max pooling with a 2x2 pooling window. Then, it is connected to the second and third convolutional modules, which have 64 3x3 convolutional kernels and 128 3x3 convolutional kernels respectively. This is followed by the second pooling layer, which uses max pooling with a 2x2 pooling window. Feature flattening is then performed, followed by two fully connected layers: the first fully connected layer has 512 neurons, and the second fully connected layer has 256 neurons. Finally, it is connected to the output layer, which outputs the monitoring results of post-mastectomy rehabilitation posture.

[0105] Design an activation function, expressed as follows:

[0106] ;

[0107] in, This represents the activation function. This represents the input to the activation function. This represents the switch weight, and log(·) represents taking the logarithm. Indicates the nonlinear factor. This represents the slope control parameter;

[0108] The convolutional layer is designed as follows:

[0109] ;

[0110] Where i and j represent the position indices of the feature maps, and n represents the layer index of the convolutional layer. This indicates that the output feature map of the nth convolutional layer is at position The values ​​at the specified positions, k and p represent the row and column indices of the convolution kernel, respectively, and K and P represent the number of rows and columns of the convolution kernel, respectively. The output feature map of layer n-1 is at position The value at that location, This indicates the position of the convolution kernel in the nth convolutional layer. The value at that location, This represents the bias of the nth convolutional layer;

[0111] Channel attention is introduced, and the input features are subjected to global average pooling and global max pooling operations in sequence, and then processed by the activation function to obtain the channel weights.

[0112] The loss function is designed as follows:

[0113] ;

[0114] in, Let d represent the loss function, and d represent the index of the data sample. This indicates the degree of standardization of the actual posture of the d-th sample. This indicates the standardity of the pose predicted by the model's output for the d-th sample. Indicates the adjustment weight. Indicates channel attention weights. This represents the expected channel attention weight. Represents the square of the L2 norm;

[0115] The dynamic attenuation rate is designed as follows:

[0116] ;

[0117] in, This indicates the number of times the model's parameters have been learned. This represents the decay rate during the iterth iteration of model parameter learning. This represents the initial decay rate, and iterM represents the maximum number of parameter learning iterations. This indicates taking the maximum value. Indicates the lower limit of decay;

[0118] The gradient momentum is calculated as follows:

[0119] ;

[0120] ;

[0121] in, and Let represent the first and second gradient momentum of the loss function with respect to the model parameters during the iterth iteration of model parameter learning, respectively. and Let represent the first and second gradient momentum of the loss function with respect to the model parameters during the (iter-1)th iteration of model parameter learning. Let represent the gradient of the loss function with respect to the model parameters during the iterth iteration of model parameter learning. Indicates the second attenuation rate. Indicates the momentum change factor;

[0122] Deviation correction is expressed as follows:

[0123] ;

[0124] ;

[0125] in, and Let represent the first and second gradient momentum of the loss function with respect to the model parameters after bias correction during the iterth model parameter learning process, respectively.

[0126] Model parameter learning is represented as follows:

[0127] ;

[0128] in, This represents the model parameters at the iterth iteration of model parameter learning. This represents the model parameters during the (iter-1)th iteration of model parameter learning. Indicates the learning rate. Indicates the learning rate scaling weights. This represents a micro-parameter.

[0129] By performing the above operations, this solution addresses the problems of low monitoring accuracy and generalization ability of traditional post-masturbation posture monitoring models, which suffer from simple model structure, poor adaptability of activation functions, insufficient feature extraction, and inflexible learning rate adjustment. This solution improves the monitoring accuracy, generalization ability, flexibility, and adaptability of the model by designing the model structure, designing the activation function, introducing channel attention, designing the loss function, calculating gradient momentum, and bias correction.

[0130] Example 4, see Figure 1 , Figure 4 and Figure 5 This embodiment is based on the above embodiment. The model hyperparameter optimization module searches for model hyperparameters through search preparation, designing search sensitivity, designing chaos factor, designing chaotic search function, and designing local refinement search function. Specifically, it includes the following:

[0131] Search preparation involves selecting search hyperparameters, including nonlinear factors, slope control parameters, adjustment weights, initial decay rate, and lower decay limit; creating a parameter search space by setting the reciprocal of the model's loss function value as the model parameter fit rate.

[0132] The search sensitivity is designed as follows:

[0133] ;

[0134] Where t represents the current number of hyperparameter searches, This represents the search sensitivity during the t-th hyperparameter search. Indicates the initial search sensitivity. This represents the sensitivity control factor. This represents a random number between -1 and 1.

[0135] The chaos factor is designed as follows:

[0136] ;

[0137] in, This represents the chaos factor during the t-th hyperparameter search. Let Ctl represent the chaos factor during the (t-1)th hyperparameter search, where Ctl is a random number between 3.5 and 4.5, representing the initial chaos factor. It is a random number between 0 and 1;

[0138] Design a chaotic search function, represented as follows:

[0139] ;

[0140] in, This represents the position of the hyperparameters found by the chaotic search function during the (t+1)th hyperparameter search. This indicates the position with the highest hyperparameter fit rate. This represents a directional control factor with a random direction and a size of unit length, where u represents the chaos intensity.

[0141] Design a local refinement search function, as follows:

[0142] ;

[0143] in, This represents the hyperparameter position found by the local refinement search function during the (t+1)th hyperparameter search. This represents the hyperparameter position found by the local refinement search function during the t-th hyperparameter search. This indicates taking the absolute value. This represents the position of the hyperparameters found by the chaotic search function during the t-th hyperparameter search.

[0144] Perform the search, set the hyperparameter fitness rate threshold, set the maximum number of hyperparameter searches, generate an initial hyperparameter search cluster, select the position with the highest fitness rate in the initial hyperparameter search cluster as the current position with the highest fitness rate, first perform a hyperparameter search for the chaotic search function, update the current position with the highest fitness rate, then perform a hyperparameter search for the local refinement search function, update the current position with the highest fitness rate again. If the fitness rate of a search point exceeds the hyperparameter fitness rate threshold during the search process, output the hyperparameters of the parameter point with the highest fitness rate; if the maximum number of hyperparameter searches is reached, start the search again; otherwise, increment the search count and proceed to the next round of search.

[0145] By performing the above operations, this scheme addresses the problems of low search efficiency, susceptibility to local optima, and inability to effectively balance global and local searches in traditional model hyperparameter optimization methods. It improves the efficiency and effectiveness of hyperparameter optimization by performing search preparation, designing search sensitivity, designing chaos factors, designing chaotic search functions, and designing local refinement search functions to search for model hyperparameters.

[0146] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0147] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0148] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A postoperative rehabilitation posture management system for breast cancer based on deep learning, characterized by: It includes data acquisition module, image processing module, module for establishing postoperative rehabilitation posture monitoring model for breast cancer, model hyperparameter optimization module, and postoperative rehabilitation posture management module for breast cancer; The data acquisition module collects image data of breast cancer patients performing rehabilitation postures after surgery and posture standard degree data, wherein the posture standard degree data includes standard postures, poor postures, non-standard postures and dangerous postures; The image processing module processes the image by determining a smooth convolution kernel, defining a gradient component matrix, gradient reconstruction, setting a threshold, edge detection and image annotation; The module for establishing a breast cancer postoperative rehabilitation posture monitoring model is constructed by designing a model structure, designing an activation function, designing a convolutional layer, introducing channel attention, designing a loss function, designing a dynamic decay rate, calculating gradient momentum, bias correction, and model parameter learning; The model hyperparameter optimization module searches for model hyperparameters by search preparation, designing search sensitivity, designing chaos factor, designing chaos search function and designing local refinement search function; The postoperative rehabilitation posture management module for breast cancer collects image data of adenocarcinoma patients performing rehabilitation postures in real time, and inputs the collected data into the postoperative rehabilitation posture monitoring model module for breast cancer. The model outputs the standard degree of the patient's posture. When the patient's posture is detected to be a dangerous posture, the system will issue a warning message to remind the patient and medical staff to pay attention to adjusting the posture. The image processing module processes the image by determining a smoothing convolution kernel, defining a gradient component matrix, gradient reconstruction, setting a threshold, edge detection and image annotation, and specifically includes the following steps: Determine the smooth convolution kernel and perform smooth convolution on the image, which is expressed as follows: ； in, Represents the convolution kernel for smooth convolution, x and y represent the horizontal and vertical coordinate indexes of the pixels of the image, respectively. represents pi, represents the smoothing standard deviation, and e represents a natural constant; Define the gradient component matrix, which defines the gradient components of the image in the horizontal and vertical directions, as follows: ； ； in, and Represent the component matrices of the image gradient in the horizontal and vertical directions, Indicates square root; Gradient reconstruction reconstructs the gradient magnitude and gradient direction, which is expressed as follows: ； ； in, represents the reconstructed gradient magnitude, represents the reconstructed gradient direction, represents the inverse tangent function; To set the threshold, first calculate the mean value of the image gradient amplitude and the standard deviation of the image gradient amplitude, and use the mean and standard deviation to set the threshold, as shown below: ； ； in, and Represent the high threshold and low threshold respectively, represents the mean value of the image gradient amplitude, Represents the standard deviation of the image gradient amplitude; Edge detection: mark pixels with gradient magnitude greater than the high threshold as strong edges, check pixels with gradient magnitude between the low threshold and the high threshold, and if these weak edge pixels are connected to the marked strong edge pixels, mark them as part of the edge. Starting from the marked strong edge pixels, check the weak edge pixels in their neighborhood, connect them to the strong edge, and build a complete edge structure. Image annotation, annotating the image data of the breast cancer patients performing rehabilitation postures through the standard degree data of the rehabilitation postures after surgery, and setting the standard degree data as label data.

2. The deep learning-based postoperative rehabilitation posture management system for breast cancer according to claim 1 is characterized in that: The model module for establishing the postoperative rehabilitation posture monitoring model for breast cancer is established by designing the model structure, designing the activation function, designing the convolution layer, introducing the channel attention, designing the loss function, designing the dynamic attenuation rate, calculating the gradient momentum, bias correction and model parameter learning, and specifically includes the following contents: Design the model structure. The breast cancer postoperative rehabilitation posture monitoring model is composed of the following parts connected in sequence. First, the input layer is connected to the first convolution module. The first convolution module has 32 3×3 convolution kernels, followed by the first pooling module, which follows the first convolution module and uses maximum pooling with a pooling window of 2×2; then connected to the second and third convolution modules. The second convolution module has 64 3×3 convolution kernels and the third convolution module has 128 3×3 convolution kernels. Then connected to the second pooling layer, using maximum pooling with a pooling window of 2×2, followed by feature flattening, and then to two fully connected layers. The first fully connected layer has 512 neurons and the second fully connected layer has 256 neurons. Finally, it is connected to the output layer to output the monitoring results of breast cancer postoperative rehabilitation posture. Design the activation function, expressed as follows: ； in, represents the activation function, represents the input of the activation function, represents the switch weight, log(·) represents the logarithm, represents the nonlinear factor, represents the slope control parameter; Design the convolutional layer, expressed as follows: ； Among them, i and j represent the position index of the feature map, and n represents the layer index of the convolution layer. Indicates that the output feature map of the nth convolutional layer is at position The value at , k and p represent the row index and column index of the convolution kernel respectively, K and P represent the number of rows and columns of the convolution kernel respectively, The n-1 layer output feature map is at position The value at Indicates that the convolution kernel of the nth convolutional layer is at position The value at Represents the bias of the nth convolutional layer; Channel attention is introduced. Channel attention is introduced, and global average pooling and global maximum pooling operations are performed on the input features in sequence, and then the channel weights are obtained through activation function processing; The loss function is designed as follows: ； in, represents the loss function, d represents the index of the data sample, Indicates the standard degree of the actual posture of the d-th sample, Indicates the standard degree of the posture predicted by the d-th sample output by the model, represents the adjustment weight, represents the channel attention weight, represents the expected channel attention weight, represents the square of L2 norm; The designed dynamic attenuation rate is expressed as follows: ； in, represents the number of times the model’s parameters are learned, It represents the decay rate of the iter-th model parameter learning. represents the initial decay rate, iterM represents the maximum number of parameter learning times, Indicates taking the maximum value, Indicates the lower limit of attenuation; Calculate the gradient momentum, expressed as follows: ； ； in, and They represent the first momentum and second momentum of the gradient of the loss function for the model parameters during the iter-th model parameter learning. and They represent the first momentum and second momentum of the gradient of the loss function for the model parameters during the iter-1th model parameter learning. It represents the gradient of the loss function with respect to the model parameters during the iter-th model parameter learning. represents the second decay rate, represents the momentum change factor; Bias correction, expressed as follows: ； ； in, and They represent the first momentum and second momentum of the gradient of the loss function for the model parameters after deviation correction during the iter-th model parameter learning. Model parameter learning is expressed as follows: ； in, It represents the model parameters during the iter-th model parameter learning. It represents the model parameters during the iter-1th model parameter learning. represents the learning rate, represents the learning rate scaling weight, Represents micro parameters.

3. The deep learning-based postoperative rehabilitation posture management system for breast cancer according to claim 1 is characterized in that: The model hyperparameter optimization module searches for model hyperparameters by search preparation, designing search sensitivity, designing chaos factor, designing chaos search function and designing local refinement search function, which specifically includes the following contents: Search preparation, select search hyperparameters, including nonlinear factors, slope control parameters, adjustment weights, initial decay rate, and decay lower limit; create a parameter search space, and set the inverse of the model's loss function value as the model parameter adaptation rate; Design search sensitivity, expressed as follows: ； Among them, t represents the current number of hyperparameter searches, represents the search sensitivity during the tth hyperparameter search, represents the initial search sensitivity, represents the sensitivity control factor, Represents a random number between -1 and 1; The chaos factor is designed as follows: ； in, represents the chaos factor during the t-th hyperparameter search, represents the chaos factor during the t-1th hyperparameter search, Ctl represents a random number between 3.5 and 4.5, and the initial chaos factor is a random number between 0 and 1; Design the chaos search function, which is expressed as follows: ； in, It represents the hyperparameter position searched by the chaotic search function during the t+1th hyperparameter search. Indicates the position where the current hyperparameter adaptation rate is the highest, It represents a random direction with a unit length direction control factor, and u represents the chaos intensity; Design a local refinement search function, expressed as follows: ； in, It represents the hyperparameter position searched by the local refinement search function during the t+1th hyperparameter search. It represents the hyperparameter position searched by the local refinement search function during the t-th hyperparameter search. Indicates taking the absolute value, Indicates the hyperparameter position searched by the chaotic search function during the tth hyperparameter search; Execute the search, set the hyperparameter adaptation rate threshold, set the maximum number of hyperparameter searches, generate the initial hyperparameter search cluster points, select the position with the highest adaptation rate in the initial point group of the hyperparameter search as the position with the highest current hyperparameter adaptation rate, first perform a hyperparameter search of the chaotic search function, update the position with the highest current hyperparameter adaptation rate, then perform a hyperparameter search of the local refinement search function, and update the position with the highest current hyperparameter adaptation rate again. If the adaptation rate of the search point is greater than the hyperparameter adaptation rate threshold during the search process, output the hyperparameter of the parameter point with the highest current hyperparameter adaptation rate; if the maximum number of hyperparameter searches is reached, search again; otherwise, increase the number of searches by one and perform the next round of search.

Images