A method and system for monitoring and evaluating the canceration of benign breast tumors

Abstract

The present invention relates to the field of information and communication technology for handling or processing health data, and specifically to a method and system for monitoring and evaluating the canceration of benign breast tumors. The method includes five steps, and the system includes a breast tumor case data preparation unit, a breast tumor case data expansion unit, a breast tumor case data feature processing unit, and a breast tumor case data identification unit. Existing methods and systems for monitoring and evaluating the canceration of benign breast tumors have the problem of low evaluation accuracy. The method and system for monitoring and evaluating the canceration of benign breast tumors provided by the present invention have high evaluation accuracy.

Description

Technical Field

[0001] The present invention relates to the field of information and communication technology for handling or processing health data, and in particular to a method and system for monitoring and evaluating the canceration of benign breast tumors. Background Art

[0002] Breast tumors are tumors that grow in the breast and can be benign or malignant. Fibroadenomas are the most common benign breast tumors, followed by intraductal papilloma, and benign phyllodes tumors are third. There are nearly 20 other benign breast tumors, but they account for less than 3% of the total benign tumors. Benign breast tumors may turn into malignant tumors, i.e. cancer. Breast cancer is one of the common malignant tumors. If it can be detected and treated in time, the cure rate can be improved.

[0003] However, the morphology and tissue structure of breast tumors are relatively complex. Traditional imaging examination methods (such as X-rays, ultrasound, etc.) often have difficulty in accurately identifying the nature of tumors in the early stages, especially in distinguishing between benign and malignant tumors. Therefore, a benign breast tumor canceration monitoring and evaluation system is introduced to monitor and evaluate the canceration process of benign breast tumors. However, the existing benign breast tumor canceration monitoring and evaluation system faces the following three problems:

[0004] 1. Insufficient real case data, resulting in a lack of training data for the system. Although existing benign breast tumor canceration monitoring and assessment systems use simple data augmentation methods, such as rotation and flipping, to generate case data, these methods often fail to generate diverse and realistic generated case data, resulting in insufficient diversity in case training data sets, which in turn affects the system's training effect and classification accuracy.

[0005] Second, it is difficult to effectively capture the nonlinear characteristics in the data when processing high-dimensional complex data. Cases are usually a kind of high-dimensional complex data. The existing breast benign tumor canceration monitoring and evaluation system mostly uses traditional neural network models to process data. When processing high-dimensional complex data, it is easy to have gradient vanishing or gradient explosion phenomena, which makes it easy to make errors in the subsequent data classification;

[0006] 3. When dealing with cases with imbalanced categories or cases that are difficult to distinguish, the classification accuracy of existing benign breast tumor canceration monitoring and assessment systems is still difficult to meet the standards for clinical application. Many existing classification methods fail to dynamically adjust the impact of error terms during training, resulting in the inability to effectively optimize samples that are difficult to classify, which in turn leads to low accuracy in benign breast tumor canceration monitoring and assessment.

[0007] Based on the above reasons, the existing benign breast tumor canceration monitoring and assessment methods and systems have the problem of low assessment accuracy. Summary of the invention

[0008] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a method and system for monitoring and evaluating the canceration of benign breast tumors with high evaluation accuracy.

[0009] In order to solve the above technical problems, the present invention provides a method for monitoring and evaluating the canceration of benign breast tumors, comprising:

[0010] S1. Collect multiple real case data, manually annotate the benign tumor stage corresponding to each real case data, and then store them;

[0011] Each real case data includes the patient’s age, gender, physical monitoring data, family medical history, tumor location, tumor size, imaging features, cytological grade, treatment method, follow-up time, and follow-up results;

[0012] Benign tumor stages include early benign stage, intermediate benign stage and advanced benign stage;

[0013] S2, using multiple real case data to train the data expansion model to obtain a trained data expansion model, using the trained data expansion model to generate multiple generated case data, and using the multiple real case data and the multiple generated case data as case training data sets;

[0014] The training data expansion model uses a generative adversarial network algorithm based on symmetry adjustment;

[0015] S3, inputting the case training data set into the feature extraction model, training the feature extraction model, and obtaining a trained feature extraction model and a case training data set after feature extraction;

[0016] The training feature extraction model uses a neural network algorithm guided by quantum fluctuations;

[0017] S4, inputting the case training data set after feature extraction into the classifier, training the classifier, and obtaining a trained classifier;

[0018] The classifier is trained using a fractional-order neural network algorithm based on error adjustment;

[0019] S5. Collect new real case data, input the new real case data into the trained feature extraction model to obtain new real case data after feature extraction, input the new real case data after feature extraction into the trained classifier to obtain the benign tumor stage corresponding to the new real case data.

[0020] As a further improvement of the present invention: the training data expansion model in S2 includes:

[0021] S201, initializing the parameters of the generator and discriminator of the generative adversarial network;

[0022] S202, in the first training stage, adaptively adjusting the weights of the generator and the discriminator, and adjusting the direction and amplitude of the weight perturbation according to the generation effect of the initial batch of generated case data;

[0023] S203, in the second training stage, each cycle includes processing multiple batches of case data, in each batch, the generator generates generated case data that conforms to the target distribution, and the discriminator distinguishes between real case data and generated case data;

[0024] S204, in the third training stage, according to the classification accuracy of the discriminator for each type of case data, the weights of each type of case data in the loss function are adjusted, and for case data categories with low classification accuracy, their weights in the loss function are increased to encourage the generator to pay more attention to these categories;

[0025] S205. In the fourth training stage, the symmetry of the case data distribution during the entire training process of the generative adversarial network is monitored. If it is found that the generated case data deviates from the distribution of the real case data, the symmetry of the case data distribution is restored by adjusting the parameters of the non-uniform weight perturbation to ensure the diversity and authenticity of the generated case data.

[0026] S206, repeating iterations S202 to S205 until a preset stop iteration condition is met, indicating that the data augmentation model training is completed.

[0027] As a further improvement of the present invention: the training feature extraction model in S3 includes:

[0028] S301, initializing the parameters of the neural network;

[0029] S302, the input training data is processed by each layer of the neural network, the activation of each neuron is determined by the quantum dynamic equation, and the activation value is calculated based on the input of the current neuron;

[0030] S303, through a dynamically adjusted feedback loop control mechanism, during the training process of the neural network, the feature activity and the response of the neural network in the forward propagation process are monitored in real time, and the activation function parameters of each layer of the neural network are dynamically adjusted;

[0031] S304, adjusting the weights through a back-propagation algorithm according to the error between the actual output of the breast tumor case training data and the output of the neural network, wherein the update of each weight depends not only on the gradient but also on the fluctuation of the weight in the quantum field;

[0032] S305, dynamically adjust the fluctuation guidance coefficient according to the complexity of the output features of each layer, the first Fluctuation guidance coefficient of the layer The adjustment strategy is as follows:

[0033] ,

[0034] In the formula, is the adjustment coefficient of the fluctuation guidance coefficient, is the feature sensitivity adjustment factor, is the entropy measurement function of the feature, For the neural network The input feature set of the layer;

[0035] S306. During the iteration process, the training progress of the neural network is continuously monitored. Once the feature extraction model reaches the preset performance threshold, the training is stopped. The convergence of the neural network is tested by the following quantum stability indicators:

[0036] ,

[0037] In the formula, is the quantum stability index of the neural network, is the number of layers of the neural network; represents the natural exponential function; is a correction factor, which is adjusted based on the performance of the feature extraction model on the training set or the stability in past iterations; represents the L2 norm; For the neural network The weight of the layer, For the neural network The weight of the layer; It is a stability adjustment factor, which is used to evaluate the smoothness of weight changes to determine whether to terminate training.

[0038] Preferably, in S306, during the iteration process, the neural network Layer weights The update method is expressed as:

[0039] ,

[0040] In the formula, is the parameter update operation, is the learning rate of the neural network, is the loss function of the neural network with respect to Layer weight gradients; is the hyperbolic tangent function; is the cosine function; To adjust the intensity of the foundation; is the sensitivity coefficient of the neural network, is the change in the neural network weights, The neural network The fluctuation guidance coefficient of the layer.

[0041] As a further improvement of the present invention: the training classifier in S4 includes:

[0042] S401, initializing the parameters of the fractional-order neural network, including the weights and biases of the fractional-order neural network; using fractional-order perturbations for initialization, and applying perturbations to the weights during the initialization process of the fractional-order neural network, the magnitude of which is determined by the fractional-order adjustment factor;

[0043] S402, in each training iteration, the case data after feature extraction is forward propagated through the fractional-order neural network, and finally the classification result is obtained through the output layer;

[0044] S403, calculating the gradient of the fractional-order neural network using a back-propagation algorithm;

[0045] S404, during the training process of the fractional-order neural network, dynamically adjusting the learning rate according to the changing trend of the loss function;

[0046] S405, repeating S402-S404 until a preset stop iteration condition is met, which means that the classifier training is completed.

[0047] Preferably, the forward propagation in S402 is expressed as:

[0048] ,

[0049] In the formula, is the fractional order neural network The input of the layer, The fractional-order neural network The weight matrix of the layer, The fractional-order neural network The output of the layer, The fractional-order neural network The bias term of the layer, is the fractional order neural network The weight coefficients associated with each feature of the layer, is the fractional order neural network The input feature matrix of the layer, is the fractional order neural network The output of the layer, is the Sigmoid activation function.

[0050] The present invention also provides a system for monitoring and evaluating the canceration of benign breast tumors, which uses the above-mentioned method for monitoring and evaluating the canceration of benign breast tumors, and comprises:

[0051] Breast tumor case data preparation unit, used to collect and store multiple real case data;

[0052] A breast tumor case data expansion unit, used to generate multiple pieces of generated case data through a trained data expansion model; the data expansion model is trained using a generative adversarial network algorithm based on symmetry regulation;

[0053] A breast tumor case data feature processing unit, used to extract features from the real case data or the generated case data through a trained feature extraction model; the feature extraction model is trained using a neural network algorithm guided by quantum fluctuations;

[0054] The breast tumor case data identification unit is used to identify the benign tumor stage corresponding to the real case data or the generated case data after feature extraction through a trained classifier, and the classifier is trained using a fractional order neural network algorithm based on error adjustment.

[0055] The beneficial effects of the present invention are as follows: the present invention provides a method and system for monitoring and evaluating the canceration of benign breast tumors, which have high evaluation accuracy.

[0056] This method uses a generative adversarial network training data augmentation model based on symmetry adjustment, and uses the trained data augmentation model to generate case data. The symmetry adjustment mechanism is used to control the symmetry between the distribution of generated data and the real data, thereby ensuring the similarity between the generated data and the real data, enhancing the diversity and authenticity of the training data, effectively improving the generalization ability of the model, and solving the problem of insufficient case data.

[0057] A feature extraction model is trained based on a neural network guided by quantum fluctuations. The optimization mechanism guided by quantum fluctuations simulates the random fluctuations in quantum mechanics, effectively avoiding problems such as gradient vanishing and gradient explosion that may occur in traditional neural networks, improving the learning ability and generalization performance of neural networks, and making the trained feature extraction model applicable to complex data.

[0058] The classifier is trained using a fractional-order neural network algorithm based on error adjustment. In traditional neural networks, the error term is generally static. This method adjusts the weight of the error term so that the classifier can adapt to learning tasks at different stages, especially for case data that is more difficult to classify, thereby improving the stability and accuracy of the classifier.

[0059] In summary, this method and the system have higher evaluation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1This is a principle block diagram of a method for monitoring and evaluating the canceration of benign breast tumors provided by the present invention.

[0061] Figure 2 A schematic diagram of a benign breast tumor canceration monitoring and evaluation system provided by the present invention. DETAILED DESCRIPTION

[0062] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0063] like Figure 1 As shown, the present invention provides a method for monitoring and evaluating the canceration of benign breast tumors, comprising:

[0064] S1. Collect multiple real case data, manually annotate the benign tumor stage corresponding to each real case data, and store them in CSV format;

[0065] Each real case data includes the patient's age, gender, physical monitoring data, family medical history, tumor location, tumor size, imaging characteristics, cytological grade, treatment method, follow-up time and follow-up results; age, gender, family medical history, tumor location, tumor size, imaging characteristics, cytological grade, treatment method, follow-up time and follow-up results are all from the confirmed breast tumor case data collected by hospitals and diagnostic centers; physical monitoring data comes from monitoring equipment, and physical monitoring data includes blood pressure data monitored by a sphygmomanometer, blood sugar data monitored by a blood glucose meter, body temperature data monitored by an electronic thermometer, electrocardiogram data monitored by an electrocardiograph, and blood oxygen saturation data monitored by an oximeter. The above data can also be monitored by a comprehensive monitor or wearable device.

[0066] Age is a continuous variable, expressed as an integer; gender is a discrete variable, 0 represents male and 1 represents female; family medical history is a discrete variable, 0 represents no relevant medical history, and 1 represents a history of breast cancer; tumor location is a discrete variable, specifically divided into left breast and right breast; tumor size is a continuous variable, measured in centimeters; imaging features are discrete vectors that describe the imaging characteristics of the tumor; cytological grade is a discrete variable, graded from 1 to 3, and the higher the grade, the more obvious the atypia; treatment methods are discrete variables, including surgery, radiotherapy, chemotherapy, etc.; follow-up time is a continuous variable, indicating the number of months of follow-up; follow-up results are discrete variables, 0 represents no recurrence and 1 represents recurrence.

[0067] The stages of benign tumors include early benign stage, intermediate benign stage and late benign stage.

[0068] S2, using multiple real case data to train the data expansion model to obtain a trained data expansion model, using the trained data expansion model to generate multiple generated case data, and using the multiple real case data and the multiple generated case data as case training data sets;

[0069] The data augmentation model is built based on the generative adversarial network algorithm, and the training data augmentation model adopts the generative adversarial network algorithm based on symmetry adjustment;

[0070] The generative adversarial network algorithm based on symmetry adjustment refers to monitoring the symmetry of the distribution of generated case data and real case data in the fourth training stage of the generative adversarial network. If it is found that the generated case data deviates from the distribution of the real case data, the symmetry of the distribution of the generated case data and the real case data is restored by adjusting the parameters of the non-uniform weight perturbation to ensure the diversity and authenticity of the generated case data. Symmetry adjustment includes the above-mentioned monitoring of symmetry, judging whether the generated case data deviates from the distribution of the real case data, and restoring the symmetry of the distribution of the generated case data and the real case data through adjustment.

[0071] Based on multiple real case data, a generative adversarial network algorithm based on symmetry adjustment is used to generate multiple generated case data, including:

[0072] S201, initializing the parameters of the generator and discriminator of the generative adversarial network, including:

[0073] ,

[0074] In the formula, are the parameters of the generator, is the parameter update operation, The mean is 0, is the normal distribution with variance, is the initialization variance of the generated adversarial network parameters, are the parameters of the discriminator;

[0075] S202. In the first training stage, the weights of the generator and the discriminator are adaptively adjusted. According to the generation effect of the initial batch of generated case data, the direction and amplitude of the weight perturbation are adjusted, and the generated case data category with lower generation quality is preferentially improved. The generated case data category with lower generation quality refers to the generated case data that is judged by the discriminator to be greatly different from the real case data or the generated case data with a larger divergence value after calculating the Jensen-Shannon divergence between the generated case data and the real case data in each category; preferential improvement means that the generator is used to first perform data expansion on the generated case data category with lower generation quality, for example, The Jensen-Shannon divergence value of the case data generated by the first category is 0.8, and the Jensen-Shannon divergence value of the case data generated by the second category is 0.2. 0.8>0.2, the Jensen-Shannon divergence value of the case data generated by the first category is larger, indicating that the quality of the case data generated by the first category is low. First, the case data generated by the first category is expanded until the Jensen-Shannon divergence value of the case data generated by the first category is greater than the divergence value of the case data generated by the second category; the adaptive weight adjustment of the initial batch is optimized by back propagation, which is expressed as:

[0076] ,

[0077] In the formula, is the parameter of the generator; is the learning rate of the first stage of the generative adversarial network, Set to 0.01; For the parameters of the generator The gradient of is the loss function of the generative adversarial network, For the generator, is the discriminator; is the case data input to the discriminator, which is the real case data or the generated case data generated by the generator;

[0078] S203. In the second training phase, each cycle includes processing of multiple batches of case data. In each batch, the generator generates generated case data that conforms to the target distribution, and the discriminator distinguishes between real case data and generated case data, which is expressed as:

[0079] ,

[0080] In the formula, is the generated case data generated by the generator, For the generator, is the noise distribution of the generator input, are the parameters of the generator, is the probability distribution of noise, are the parameters of the discriminator; is the learning rate of the second-stage generative adversarial network, Set to 0.03; is the parameter of the discriminator The gradient of is the loss function of the generative adversarial network, For the generator, is the discriminator; The case data input to the discriminator is real case data or generated case data generated by the generator;

[0081] S204. In the third training stage, according to the classification accuracy of the discriminator for each type of case data, the weights of each type of case data in the loss function are adjusted. For case data categories with low classification accuracy, their weights in the loss function are increased to encourage the generator to pay more attention to these categories, which is expressed as:

[0082] ,

[0083] In the formula, is the dynamic weight of the discriminator for the i-th case data, representing the classification performance of the discriminator for the i-th case data, where the case data includes real case data and generated case data; i is the index of the case data; is the L2 norm, is the discriminator function; is the i-th case data input to the discriminator, which is the real case data or the generated case data generated by the generator; is the true label of the i-th case data, is the variance of the characteristic value of case data in the third training stage, is the parameter of the generator; is the learning rate of the third stage of the generative adversarial network, Set to 0.05; nc is the number of generated case data generated by the current batch generator, For the parameters of the generator The gradient of is the loss function of the generative adversarial network, For the generator, is the discriminator; is the case data input to the discriminator, which is the real case data or the generated case data generated by the generator;

[0084] In the above formula Represents the response output of the discriminator to the input, which is a composite function containing activation function and multi-layer perception. The calculation formula is:

[0085] ,

[0086] In the formula, is the discriminator function; is the i-th case data input to the discriminator, which is the real case data or the generated case data generated by the generator; is the Sigmoid function, which is used to output the discriminant probability of the discriminator; is the number of neurons in the discriminator output layer; is the index of the number of neurons in the discriminator output layer; is the output layer of the discriminator The weight of each neuron is a training parameter; is the hyperbolic tangent function; is the output layer of the discriminator The input scaling factor of each neuron is a training parameter; is the output layer of the discriminator The offset of each neuron is a training parameter; is the offset of the discriminator output layer and is a random parameter.

[0087] S205. In the fourth training stage, the symmetry of the case data distribution during the entire training process of the generative adversarial network is monitored. If it is found that the generated case data deviates from the distribution of the real case data, the symmetry of the case data distribution is restored by adjusting the parameters of the non-uniform weight perturbation to ensure the diversity and authenticity of the generated case data. The calculation formula of the distribution adjustment item of the generator in the fourth training stage is:

[0088] ,

[0089] In the formula, is the change in the parameters of the generator, is the adjustment coefficient of the generator, is the symbolic function, For the parameters of the generator The gradient of is the distance function between two distributions, is the real case data distribution, The generated case data distribution generated by the data augmentation model, is the parameter of the generator;

[0090] The Jensen-Shannon divergence is used to measure the similarity between the real case data distribution and the generated case data distribution generated by the generator, that is, The calculation formula is:

[0091] ,

[0092] In the formula, is the distance function between two distributions, is the real case data distribution, The generated case data distribution generated by the data augmentation model, is the Kullback-Leibler divergence function, is the average of the real case data distribution and the generated case data distribution generated by the data augmentation model, .

[0093] S206, repeating iterations S202 to S205 until a preset stop iteration condition is met, indicating that the data augmentation model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0094] S3, inputting the case training data set into the feature extraction model, training the feature extraction model, and obtaining a trained feature extraction model and a case training data set after feature extraction;

[0095] The core task of feature extraction is to extract representative features from new samples and map them to the latent feature space. The feature extraction model is built based on a neural network, and the neural network algorithm guided by quantum fluctuations is used to train the feature extraction model. The neural network algorithm guided by quantum fluctuations refers to the use of the fluctuation theorem in quantum mechanics to simulate and optimize the weight and parameter update process of the neural network, and improve the learning ability and generalization of the neural network through the random fluctuations of the quantum state to adapt to the complex data characteristics of benign breast tumors and breast tumor cases.

[0096] Training the feature extraction model involves:

[0097] S301, initialize the parameters of the neural network; the bias parameters of the neural network are initialized to a zero vector, and the weights of the neural network are initialized to:

[0098] ,

[0099] In the formula, is the initial weight matrix of the neural network; is a small constant, Set to 0.0001; is the random perturbation constant, Set to 2; is a uniformly distributed random variable in the interval [0, 1];

[0100] S302, the input training data is processed by each layer of the neural network. The activation of each neuron is determined by the quantum dynamic equation. The activation value is calculated based on the input of the current neuron, which is expressed as:

[0101] ,

[0102] In the formula, For the neural network The activity output of the layer, is the index of the neural network layer, is the activity function guided by quantum fluctuations; For the neural network The weighted input of the layer, For the neural network The weight of the layer, For the neural network The activity output of the layer, For the neural network Bias of the layer; is the fluctuation guidance coefficient of the neural network; The neural network Feedback adjustment function of the layer;

[0103] Will Expanded into its Taylor series form, through the enhancement of nonlinear features, the feature extraction model can show higher flexibility and adaptability when processing more complex training data to reveal the influence of quantum fluctuation guidance on the function. The Taylor series form is expanded to the third term and can be expressed as:

[0104] ,

[0105] In the formula, is the fluctuation guidance coefficient of the neural network; For the neural network The weighted input of the layer, For the neural network The weight of the layer, For the neural network The activity output of the layer, For the neural network Bias of the layer;

[0106] S303, through the dynamic adjustment feedback loop control mechanism, in the training process of the neural network, the feature activity and the response of the neural network in the forward propagation process are monitored in real time, and the activation function parameters of each layer of the neural network are dynamically adjusted; thereby optimizing the response capability and processing efficiency of the feature extraction model to complex features, and adjusting the nonlinear behavior of the activation function based on the real-time monitored feature activity to adapt to the changes in the input training data. The calculation formula of the feedback adjustment function is:

[0107] ,

[0108] In the formula, The neural network The feedback adjustment function of the layer, is the regulation strength of the neural network, Set to 2; is the neural network sensitivity adjustment factor, Set to 0.3; The neural network The average activation value of the neurons in the layer; is the target activity threshold, which aims to keep the neural network activity within a certain dynamic range;

[0109] S304, according to the error between the actual output of the breast tumor case training data and the output of the neural network, the weight is adjusted through the back propagation algorithm, and the update of each weight depends not only on the gradient, but also on the fluctuation of the weight in the quantum field; thereby enhancing the additional randomness and global optimization capabilities. During back propagation, the partial derivative calculation method through the gradient of quantum fluctuations is expressed as:

[0110] ,

[0111] In the formula, is the symbol of partial derivative, is the loss function, For the neural network The weight of the layer, For the neural network Activity output of the layer The transpose of For the neural network The error gradient of the layer is calculated by the preset Softmax function; The quantum modulation intensity is used to control the influence of the quantum term and ensure the balance between the complexity of the feature extraction model and the stability of training; is the fluctuation guidance coefficient of the neural network, For the neural network The weight of the layer;

[0112] The quantum modulation intensity is calculated through a dynamic adjustment mechanism, and the quantum modulation intensity is modulated by the local field, which is calculated by the weight difference of adjacent layers, so that the neural network can adaptively adjust its parameter update strategy during the learning process; optimize learning efficiency and stability, expressed as:

[0113] ,

[0114] In the formula, For the basic adjustment strength, set it to 0.1; is the sensitivity coefficient of the neural network; is the change in weight, For the neural network The weight of the layer;

[0115] S305. Dynamically adjust the fluctuation guidance coefficient according to the complexity of the output features of each layer to ensure that the neural network can self-regulate and adapt to complex training data and optimize the transmission and fusion of features. The adjustment strategy of the fluctuation guidance coefficient is as follows:

[0116] ,

[0117] In the formula, The neural network The fluctuation guidance coefficient of the layer increases the nonlinear dynamic behavior; is the adjustment coefficient of the fluctuation guidance coefficient, Set to 0.2; is the feature sensitivity adjustment factor, Set to 0.3; is the entropy measurement function of the feature, For the neural network The input feature set of the layer.

[0118] S306. During the iteration process, the training progress of the neural network is continuously monitored. Once the feature extraction model reaches the preset performance threshold, the training is stopped. The convergence of the neural network is tested by the following quantum stability indicators:

[0119] ,

[0120] In the formula, is the quantum stability index of the neural network, is the number of layers of the neural network; represents the natural exponential function; is a correction factor, which is adjusted based on the performance of the feature extraction model on the training set or the stability in past iterations; represents the L2 norm; For the neural network The weight of the layer, For the neural network The weight of the layer; is a stability adjustment factor, which is used to evaluate the smoothness of weight changes to determine whether to terminate training. Set to 0.1;

[0121] In the iterative process in S306, the weight update method of the neural network is expressed as:

[0122] ,

[0123] In the formula, is the parameter update operation, For the neural network The weight of the layer, is the learning rate of the neural network, is the loss function of the neural network with respect to Layer weight gradients; is the hyperbolic tangent function; is the cosine function; To adjust the intensity of the foundation; is the sensitivity coefficient of the neural network, is the change in the neural network weights, The neural network The fluctuation guidance coefficient of the layer.

[0124] The learning rate of the neural network is dynamically adjusted according to the size of its quantum entropy to reflect the uncertainty and complexity of each layer in the learning process. The calculation formula is:

[0125] ,

[0126] In the formula, is the learning rate of the tth iteration of the neural network, t is the index of the number of iterations, is the basic learning rate of the neural network, Set to 0.01; is the learning rate adjustment coefficient of the neural network, which is used to control the influence of entropy on the learning rate; is the quantum stability indicator of the t-th iteration of the neural network.

[0127] S4, inputting the case training data set after feature extraction into the classifier, training the classifier, and obtaining a trained classifier;

[0128] The classifier is constructed based on a fractional-order neural network, and the training classifier adopts a fractional-order neural network algorithm based on error adjustment; the fractional-order neural network algorithm based on error adjustment refers to adjusting the calculation method of the cross entropy in the loss function of the fractional-order neural network by adjusting the parameters of the error term, so that the error term changes continuously with the progress of training, thereby optimizing the classification performance and improving the classification accuracy.

[0129] Training a classifier involves:

[0130] S401, initializing the parameters of the fractional-order neural network, including the weights and biases of the fractional-order neural network. The method uses fractional-order perturbations for initialization, and the initialization process of the fractional-order neural network imposes perturbations on the weights. The size of the perturbation is determined by the adjustment factor of the fractional order, which is expressed as:

[0131] ,

[0132] In the formula, For the The weight matrix of the fractional-order neural network of the layer, is the layer index of the fractional-order neural network, is a fractional calculus operation, is the order of the fractional order, For the Initialization weight matrix of the fractional neural network of the layer; For the The perturbation term of the fractional-order neural network with layers, is the perturbation step size factor of the fractional-order neural network, is a random disturbance;

[0133] The perturbation step factor of the fractional-order neural network controls the amplitude of the perturbation of each layer, and the calculation formula is:

[0134] ,

[0135] In the formula, is the perturbation step size factor of the fractional-order neural network, is the initial perturbation amplitude of the fractional-order neural network, is the adjustment factor of the disturbance amplitude; is the periodic frequency related to the number of iterations, and the amplitude of the disturbance changes dynamically as the training process progresses through periodic adjustment; is the current iteration number of the fractional-order neural network;

[0136] Random perturbations are random values ​​generated from a uniform distribution to introduce unpredictable perturbations. The calculation formula is:

[0137] ,

[0138] In the formula, is the scale factor, which adjusts the range of disturbance; is a uniformly distributed random number from 0 to 1;

[0139] S402. In each training iteration, the case data after feature extraction is forward propagated through the fractional-order neural network, and finally the classification result is obtained through the output layer. Suppose the first The input of the layer is , the fractional order neural network The output of the layer is , the forward propagation method is expressed as:

[0140] ,

[0141] In the formula, is the fractional order neural network The input of the layer, The fractional-order neural network The weight matrix of the layer, The fractional-order neural network The output (activation value) of the layer, The fractional-order neural network The bias term of the layer, is the fractional order neural network The weight coefficients associated with each feature of the layer, is the fractional order neural network The input feature matrix of the layer, is the fractional order neural network The output (activation value) of the layer, is the Sigmoid activation function;

[0142] The error term contained in the loss function of the fractional-order neural network is dynamically adjusted to adapt to the characteristics of different training stages. The calculation method of the cross entropy in the loss function is adjusted by the error term adjustment factor, so that the error term changes continuously as the training progresses. The calculation formula is:

[0143] ,

[0144] In the formula, is the loss function of the fractional-order neural network, is the number of samples input to the fractional-order neural network in the current batch, For the The error term adjustment factor for the iteration, For the The true labels of samples, is the fractional order neural network for the The predicted value of samples, Index of the number of samples input to the neural network;

[0145] The error adjustment factor is updated according to the number of training iterations, so that in the later stages of training, the classifier can pay more attention to samples that are more difficult to classify, thereby improving the stability and accuracy of training. The update method is expressed as:

[0146] ,

[0147] In the formula, For the The error term adjustment factor for the iteration; is the initial error adjustment factor, Set to 0.5; is the factor that adjusts the learning rate, Set to 0.01; Control factor decay speed, Set to 0.95;

[0148] The weighting coefficient dynamically adjusts the influence of the feature and uses the nonlinear influence of the feature so that the contribution of each feature can change dynamically during the training process. The calculation formula is:

[0149] ,

[0150] In the formula, is the fractional order neural network The weight coefficients associated with each feature of the layer, is the first adjustment coefficient of the fractional-order neural network, is the second adjustment coefficient of the fractional-order neural network, Fractional order neural network The square of the L2 norm of the layer’s output (activation value);

[0151] S403. Based on the loss function result, the back propagation algorithm is used to calculate the gradient of the fractional-order neural network. Different from the conventional gradient calculation, this method uses fractional-order derivatives to calculate the gradient, so that the gradient calculation can be based on the long-term dependence of the case data, and improve the adaptability of the classifier to complex case data patterns. Specifically, the partial derivative of the loss function of the fractional-order neural network with respect to the weight parameter The calculation formula is:

[0152] ,

[0153] In the formula, is a fractional calculus operation, is the order of the fractional order, is the partial derivative of the loss function of the fractional-order neural network with respect to the activation value, is the partial derivative of the activation value of the fractional-order neural network with respect to the weight.

[0154] The update of the gradient depends on the fractional gradient operation. Considering the dynamic changes of each layer of features during the training process, the gradient of each layer not only considers the error, but also integrates the influence of the disturbance. Then the partial derivative of the activation value of the fractional neural network with respect to the weight is The calculation formula is:

[0155] ,

[0156] In the formula, For the The output of the layer, is the fractional order neural network The weighting coefficients associated with each feature at the layer.

[0157] S404. During the training process of the fractional-order neural network, the learning rate is dynamically adjusted according to the changing trend of the loss function. In the back propagation of the loss function, the learning rate of the fractional-order neural network is adjusted according to the changing trend of the loss function. The weight update amount of the layer The calculation formula is:

[0158] ,

[0159] In the formula, is the learning rate of the fractional-order neural network, represents fractional calculus operations, is the order of the fraction, is the partial derivative of the loss function of the fractional-order neural network with respect to the weight parameter;

[0160] The learning rate of the fractional-order neural network is based on the gradient size and the impact of the fractional-order perturbation on the gradient, thereby improving the adaptive adjustment ability of the training process of the fractional-order neural network. The calculation formula is:

[0161] ,

[0162] In the formula, For the The learning rate of the fractional neural network for each iteration; is the initial learning rate of the fractional-order neural network, Set to 0.01; is the hyperparameter that controls the attenuation, Set to 0.9; is the fractional order neural network The weight coefficients associated with each feature of the layer; is an adaptive adjustment factor that reflects the nonlinear characteristics of the training process. Set to 0.1;

[0163] S405, repeating iterations S402 to S404 until a preset stop iteration condition is met, indicating that the classifier training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0164] S5. Collect new real case data, input the new real case data into the trained feature extraction model to obtain new real case data after feature extraction, input the new real case data after feature extraction into the trained classifier to obtain the benign tumor stage corresponding to the new real case data.

[0165] like Figure 2 As shown, the present invention also provides a system for monitoring and evaluating the canceration of benign breast tumors, which uses the above-mentioned method for monitoring and evaluating the canceration of benign breast tumors, including:

[0166] Breast tumor case data preparation unit, used to collect and store multiple real case data;

[0167] A breast tumor case data expansion unit, used to generate multiple pieces of generated case data through a trained data expansion model; the data expansion model is trained using a generative adversarial network algorithm based on symmetry regulation;

[0168] A breast tumor case data feature processing unit, used to extract features from the real case data or the generated case data through a trained feature extraction model; the feature extraction model is trained using a neural network algorithm guided by quantum fluctuations;

[0169] The breast tumor case data identification unit is used to identify the benign tumor stage corresponding to the real case data or the generated case data after feature extraction through a trained classifier, and the classifier is trained using a fractional order neural network algorithm based on error adjustment.

Claims

1. A method for monitoring and evaluating the canceration of benign breast tumors, characterized in that: include: S1. Collect multiple pieces of real case data, manually annotate the benign tumor stage corresponding to each piece of the real case data, and then store them; Each piece of real case data includes the patient's age, gender, physical monitoring data, family medical history, tumor location, tumor size, imaging features, cytological grade, treatment method, follow-up time and follow-up results; The benign tumor staging includes benign early stage, benign middle stage and benign late stage; S2, using the plurality of real case data to train the data expansion model to obtain a trained data expansion model, using the trained data expansion model to generate a plurality of generated case data, and using the plurality of real case data and the plurality of generated case data as case training data sets; The training of the data expansion model adopts a generative adversarial network algorithm based on symmetry adjustment; S3, inputting the case training data set into the feature extraction model, training the feature extraction model, and obtaining the trained feature extraction model and the case training data set after feature extraction; The training of the feature extraction model adopts a neural network algorithm guided by quantum fluctuations; S4, inputting the case training data set after the feature extraction into the classifier, training the classifier, and obtaining the trained classifier; The training of the classifier adopts a fractional order neural network algorithm based on error adjustment; S5. Collect new real case data, input the new real case data into the trained feature extraction model to obtain the new real case data after feature extraction, input the new real case data after feature extraction into the trained classifier to obtain the benign tumor stage corresponding to the new real case data.

2. A method for monitoring and evaluating benign breast tumor canceration according to claim 1, characterized in that: The training of the data expansion model in S2 includes: S201, initializing the parameters of the generator and discriminator of the generative adversarial network; S202, in the first training stage, adaptively adjusting the weights of the generator and the discriminator, and adjusting the direction and amplitude of the weight perturbation according to the generation effect of the initial batch of generated case data; S203, in the second training stage, each cycle includes processing multiple batches of case data, in each batch, the generator generates generated case data that conforms to the target distribution, and the discriminator distinguishes between real case data and generated case data; S204, in the third training stage, according to the classification accuracy of the discriminator for each type of case data, the weights of each type of case data in the loss function are adjusted, and for case data categories with low classification accuracy, their weights in the loss function are increased to encourage the generator to pay more attention to these categories; S205. In the fourth training stage, the symmetry of the case data distribution during the entire training process of the generative adversarial network is monitored. If it is found that the generated case data deviates from the distribution of the real case data, the symmetry of the case data distribution is restored by adjusting the parameters of the non-uniform weight perturbation to ensure the diversity and authenticity of the generated case data. S206, repeating iterations S202 to S205 until a preset stop iteration condition is met, indicating that the data expansion model training is completed.

3. A method for monitoring and evaluating benign breast tumor canceration according to claim 1, characterized in that: The training of the feature extraction model in S3 includes: S301, initializing the parameters of the neural network; S302, the input training data is processed by each layer of the neural network, the activation of each neuron is determined by the quantum dynamic equation, and the activation value is calculated based on the input of the current neuron; S303, through a dynamically adjusted feedback loop control mechanism, during the training process of the neural network, the feature activity and the response of the neural network in the forward propagation process are monitored in real time, and the activation function parameters of each layer of the neural network are dynamically adjusted; S304, adjusting the weights through a back-propagation algorithm according to the error between the actual output of the breast tumor case training data and the output of the neural network, wherein the update of each weight depends not only on the gradient but also on the fluctuation of the weight in the quantum field; S305, dynamically adjust the fluctuation guidance coefficient according to the complexity of the output features of each layer, the first Fluctuation guidance coefficient of the layer The adjustment strategy is as follows: ， In the formula, is the adjustment coefficient of the fluctuation guidance coefficient, is the feature sensitivity adjustment factor, is the entropy measurement function of the feature, For the neural network The input feature set of the layer; S306. During the iteration process, the training progress of the neural network is continuously monitored. Once the feature extraction model reaches the preset performance threshold, the training is stopped. The convergence of the neural network is tested by the following quantum stability indicators: ， In the formula, is the quantum stability index of the neural network, is the number of layers of the neural network; represents the natural exponential function; is a correction factor, which is adjusted based on the performance of the feature extraction model on the training set or the stability in past iterations; represents the L2 norm; For the neural network The weight of the layer, For the neural network The weight of the layer; It is a stability adjustment factor, which is used to evaluate the smoothness of weight changes to determine whether to terminate training.

4. A method for monitoring and evaluating benign breast tumor canceration according to claim 3, characterized in that: In the iterative process of S306, the neural network Layer weights The update method is expressed as: ， In the formula, is the parameter update operation, is the learning rate of the neural network, is the loss function of the neural network with respect to the Layer weight gradients; is the hyperbolic tangent function; is the cosine function; To adjust the intensity of the foundation; is the sensitivity coefficient of the neural network, is the change in the neural network weights, The neural network The fluctuation guidance coefficient of the layer.

5. A method for monitoring and evaluating benign breast tumor canceration according to claim 1, characterized in that: The training of the classifier in S4 comprises: S401, initializing the parameters of the fractional-order neural network, including the weights and biases of the fractional-order neural network; using fractional-order perturbations for initialization, and applying perturbations to the weights during the initialization process of the fractional-order neural network, the magnitude of which is determined by the fractional-order adjustment factor; S402, in each training iteration, the case data after feature extraction is forward propagated through the fractional-order neural network, and finally the classification result is obtained through the output layer; S403, calculating the gradient of the fractional-order neural network using a back-propagation algorithm; S404, during the training process of the fractional-order neural network, dynamically adjusting the learning rate according to the changing trend of the loss function; S405, repeating S402-S404 until a preset stop iteration condition is met, which means that the classifier training is completed.

6. A method for monitoring and evaluating benign breast tumor canceration according to claim 5, characterized in that: The forward propagation method in S402 is expressed as: ， In the formula, is the fractional order neural network The input of the layer, The fractional order neural network The weight matrix of the layer, The fractional order neural network The output of the layer, The fractional-order neural network The bias term of the layer, is the fractional order neural network The weight coefficients associated with each feature of the layer, is the fractional order neural network The input feature matrix of the layer, is the fractional order neural network The output of the layer, is the Sigmoid activation function.

7. A system for monitoring and evaluating benign breast tumor canceration, characterized in that: A method for monitoring and evaluating the canceration of benign breast tumors according to any one of claims 1 to 6, comprising: Breast tumor case data preparation unit, used to collect and store multiple real case data; A breast tumor case data expansion unit, used to generate multiple pieces of generated case data through a trained data expansion model; the data expansion model is trained using a generative adversarial network algorithm based on symmetry regulation; A breast tumor case data feature processing unit, used to extract features from the real case data or the generated case data through a trained feature extraction model; the feature extraction model is trained using a neural network algorithm guided by quantum fluctuations; The breast tumor case data identification unit is used to identify the benign tumor stage corresponding to the real case data or the generated case data after feature extraction through a trained classifier, and the classifier is trained using a fractional order neural network algorithm based on error adjustment.

Images