CN112101418B - Mammary tumor type identification method, system, medium and equipment - Google Patents

Abstract

The invention discloses a breast tumor type identification method, system, medium and equipment. The method comprises the following steps: inputting biological characteristic data of a sample; constructing a dynamic convergence differential neural network with power excitation; Each connection weight component between and remains unchanged, and each connection weight component between the hidden layer and the output layer is randomly initialized; using a variety of different types of mapping functions to train a variety of different power-driven dynamic convergence difference neural network models; Put multiple power-inspired dynamic convergence differential neural networks that have been trained into the integration framework and perform model integration processing, perform type prediction on the biometric data of input samples, synthesize multiple types of prediction results, and adopt voting based on minority obedience to the majority According to the principle, the final breast tumor type identification result is obtained. The invention solves the problem of mammary gland tumor type diagnosis and achieves the effects of fast diagnosis speed, high diagnosis accuracy rate and reliable diagnosis result.

Description

A breast tumor type identification method, system, medium and equipment

technical field

The invention relates to the technical field of artificial intelligence identification, in particular to a breast tumor type identification method, system, medium and equipment.

Background technique

Statistical learning methods have always played a vital role in the intelligent diagnosis and evaluation of breast tumors. Among all statistical learning models, neural networks are suitable for wide application because of their excellent ability to learn and fit samples. Most traditional neural network training methods use gradient descent to optimize the weights between layers. As the number of neural network layers increases, the training process may fall into a local optimal solution, gradient disappearance or explosion; and in the case of breast tumor types Most of the diagnosis process only uses a single neural network model, and the reliability of the diagnosis results is difficult to guarantee. These defects are greatly restricted in the intelligent diagnosis process.

Contents of the invention

In order to overcome the defects and deficiencies in the prior art, the present invention provides a breast tumor type identification method. The present invention trains a power-excited dynamic convergence differential neural network model based on a neural dynamics method for breast tumor type diagnosis, and does not involve gradient calculations, etc. Complex calculations greatly improve the efficiency of model training, and the use of integrated models can effectively ensure the reliability of breast tumor diagnosis results.

The second object of the present invention is to provide a breast tumor type identification system.

A third object of the present invention is to provide a storage medium.

A fourth object of the present invention is to provide a computing device.

In order to achieve the above object, the present invention adopts the following technical solutions:

A method for identifying breast tumor types, comprising the steps of:

Enter the biometric data of the sample;

Construct a power-driven dynamic convergence differential neural network, set the number of input neurons to be the number of biological characteristics of a single breast tumor sample, the power-driven dynamic convergence differential neural network includes an input layer, a hidden layer and an output layer, and set the input layer Each connection weight component between the hidden layer and the hidden layer remains unchanged, and each connection weight component between the hidden layer and the output layer is randomly initialized;

Using a variety of different types of mapping functions to train a variety of different power-inspired dynamic convergence difference neural network models;

Put multiple power-inspired dynamic convergence differential neural networks that have been trained into the integration framework and perform model integration processing, perform type prediction on the biometric data of input samples, synthesize multiple types of prediction results, and adopt voting based on minority obedience to the majority According to the principle, the final breast tumor type identification result is obtained.

In order to achieve the above-mentioned second purpose, the present invention adopts the following technical solutions:

A breast tumor type recognition system, comprising: a biometric input module, a neural network building module, a neural network training module, an integrated model building module and a comprehensive judgment module;

The biometric input module is used to input biometric data of samples;

The neural network construction module is used to construct a power-driven dynamic convergence differential neural network, and the number of input neurons is set to be the number of biological characteristics of a single breast tumor sample;

The power excitation dynamic convergence difference neural network includes an input layer, a hidden layer and an output layer, each connection weight component between the input layer and the hidden layer is set and remains unchanged, and each connection weight component between the hidden layer and the output layer The connection weight components are initialized randomly;

The neural network training module is used to train a variety of different power-driven dynamic convergence differential neural network models using multiple different types of mapping functions;

The integrated model construction module is used to put multiple power-excited dynamic convergence differential neural networks that have completed the training into the integrated framework and perform model integration processing to construct an integrated model, and the integrated model is used for biometric data of input samples make type predictions;

The comprehensive judging module is used for synthesizing the prediction results of multiple types, and adopts the voting principle based on the minority obeying the majority to obtain the final breast tumor type identification result.

As a preferred technical solution, the mapping function adopts three types: linear mapping function, Sin mapping function and Sinh mapping function.

As a preferred technical solution, the expression used in the specific training method of the neural network training module is:

E(k+1)-E(k)＝-αΦ(E(k)), α＞0

Among them, where E(k) represents the result of subtracting the expected value from the output of the dynamic convergence difference neural network with power excitation after the kth learning sample, α represents the neural dynamic coefficient, and Φ represents the mapping function.

As a preferred technical solution, the activation function of each neuron in the hidden layer forms a power function sequence, and the activation function of each neuron in the output layer adopts a Softsign function.

As a preferred technical solution, the specific execution steps of the integrated model include:

Let the input breast tumor sample be the matrix X, the weight matrix between the hidden layer and the output layer after the kth learning sample is W(k), and the power excitation dynamic convergence difference neural network output Y(k):

Y(k)=g(H(XI)W(k))

in:

Let the actual attribution type of breast tumor samples be the expected value matrix Then />

Calculate the training error ε(k) after the k-th learning sample, power excitation dynamic convergence differential neural network judgment result output Y(k) passes through the Softmax layer to obtain the probability matrix P(k), and obtains the probability matrix P(k) for each breast tumor sample for each The degree of attribution of each type, and take the type corresponding to the maximum degree as the result of prediction and judgment.

As a preferred technical solution, the neural network training module uses the cross-entropy loss function formula to obtain the training error ε (k), and the specific calculation is:

Among them, L _sr and P _sr represent the elements in row s and column r in L and P(k) respectively;

Set the training error threshold to ε', if ε(k)<ε', stop iteratively solving the weight matrix W between the hidden layer and the output layer;

Solving the weight matrix between the hidden layer and the output layer is W(k+1), expressed as:

Among them, (H(XI)) ⁺ represents the Moore-Penrose pseudo-inverse of matrix H(XI), and the function Represents the inverse function of the function g.

In order to achieve the above-mentioned third purpose, the present invention adopts the following technical solutions:

A storage medium stores a program, and when the program is executed by a processor, the above-mentioned breast tumor type identification method is realized.

In order to achieve the above-mentioned fourth purpose, the present invention adopts the following technical solutions:

A computing device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the above method for identifying breast tumor types is realized.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

(1) The present invention adopts the technical solution of the integrated power excitation dynamic convergence differential neural network, which solves the technical problems of breast tumor type diagnosis, and achieves the technical effects of fast diagnosis speed, high diagnosis accuracy and reliable diagnosis results.

Description of drawings

Fig. 1 is a flow chart of the breast tumor type identification method of the present invention;

Fig. 2 is a structural schematic diagram of the breast tumor type recognition system of the present invention;

Fig. 3 is the schematic diagram of dynamic convergence difference neural network of power excitation of the present invention;

Fig. 4 is a schematic diagram of the integrated model of the present invention.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

Example

As shown in Figure 1, the present embodiment provides a method for identifying breast tumor types, including the following establishment steps:

S1: After inputting the biometric data of the breast tumor sample, the construction steps of the dynamic convergence differential neural network with power excitation are as follows:

S11: When constructing a dynamic convergence differential neural network with power excitation, set the number of input neurons to be the number of biological characteristics of a single breast tumor sample, the number of output neurons to be the total number of types of breast tumors, and the excitation of each neuron in the hidden layer The function constitutes a power function sequence, and the number needs to be adjusted and tested manually. The excitation function of each neuron in the output layer is a Softsign function;

S12: Set the weight components of each connection between the input layer and the hidden layer to be 1 and remain unchanged, and the weight components of each connection between the hidden layer and the output layer are randomly initialized and need to be updated for training;

S2: Power excitation dynamic convergence difference neural network training steps are as follows:

S21: Construct the expression of the neural dynamics training method, and set the mapping function type to three types: linear mapping function, Sin mapping function and Sinh mapping function;

The expression of the neurodynamic training method in this embodiment is: E(k+1)-E(k)=-αΦ(E(k)), α>0, wherein E(k) indicates that the intelligent doctor is at the After the second learning sample, the dynamic convergence differential neural network output minus the expected value result is equivalent to the deviation between the predicted diagnosis result of the breast tumor sample and the actual attribution type; α>0 indicates the neural dynamic coefficient, which is equivalent to In the process, the intelligent doctor learns and recognizes the input breast tumor sample; Φ represents the mapping function, which is equivalent to the learning method;

S22: According to different types of mapping functions, the neurodynamic training method is applied to the training power excitation dynamic convergence difference neural network, and iteratively updates each connection weight component between the hidden layer and the output layer, according to using a type of mapping The function trains an incentive dynamic convergence difference neural network, and trains three at the same time;

S3: Introduce the trained dynamic convergence differential neural network into the integrated framework to form an integrated model. The specific steps include:

S31: put the trained dynamic convergence differential neural network into the integrated framework for the prediction and judgment of unknown breast tumor sample types;

S32: Perform model integration processing on the three network models in the integration framework, and set a voting rule based on the minority obeying the majority, which is used for comprehensive evaluation of unknown breast tumor sample types;

S4: The specific steps of comprehensive evaluation include:

S41: For the same unknown breast tumor sample, each power in the integration framework stimulates the dynamic convergence difference neural network to make predictions and judgments respectively, and the three network models obtain three results;

S42: Use the voting rule based on the minority to obey the majority to make a comprehensive judgment on these prediction and judgment results, and obtain the final type of the sample;

S43: For multiple unknown breast tumor samples, repeat S41 and S42 one by one to predict the type;

S5: Tumor type output results.

The tumor biological characteristics input in this embodiment include: breast tumor mass thickness, cell size uniformity, cell shape uniformity, cell marginal adhesion, epithelial cell size, nuclei exposure, chromatin sparseness, nucleolus normality and The stage of nuclear division, etc., the output is the type of breast tumor, that is, benign or malignant;

As shown in Figure 2, the present embodiment also provides a breast tumor type identification system, including: a biometric input module, a neural network building module, a neural network training module, an integrated model building module and a comprehensive judgment module;

In this embodiment, the biometric input module is used to input the biometric data of the sample;

In this embodiment, the neural network construction module is used to construct a power-driven dynamic convergence differential neural network, and the number of input neurons is set to be the number of biological characteristics of a single breast tumor sample;

In this embodiment, the dynamic convergence difference neural network with power excitation includes an input layer, a hidden layer and an output layer, and each connection weight component between the input layer and the hidden layer is set and kept unchanged, and the hidden layer and the output layer Each connection weight component between is randomly initialized;

In this embodiment, the neural network training module is used to train a variety of different power-driven dynamic convergence difference neural network models using a variety of different types of mapping functions;

In this embodiment, the integrated model building module is used to put multiple power-excited dynamic convergence differential neural networks that have completed training into the integrated framework and perform model integration processing to construct an integrated model, which is used to input samples Type prediction of biometric data;

In this embodiment, the comprehensive judgment module is used to synthesize the prediction results of multiple types, and adopt the voting principle to obtain the final breast tumor type recognition result.

The expression of the neurodynamic training method in this embodiment is: E(k+1)-E(k)=-αΦ(E(k)), α>0, wherein E(k) indicates that the intelligent doctor is at the After the second learning sample, the power excitation dynamic convergence differential neural network output minus the expected value is equivalent to the deviation between the predicted diagnosis result of the breast tumor sample and the actual attribution type; α>0 indicates the neural dynamic coefficient, which is equivalent to In the process, the intelligent doctor learns and recognizes the input breast tumor sample; Φ represents the mapping function, which is equivalent to the learning method.

This embodiment uses three different types of mapping functions to train three different power-driven dynamic convergence differential neural network models, namely linear mapping functions, Sin mapping functions, and Sinh mapping functions, thereby forming three intelligent doctors. Through different learning The cognition samples are formed by means of methods, so as to analyze and judge the tumor types for the input samples of unknown types.

As shown in Figure 3, for the intelligent doctor involved in the neural network building block and neural network training module, that is, the dynamic convergence differential neural network model with power excitation, the network contains three layers: input layer, hidden layer, and output layer. The input layer and The weight components of each connection between the hidden layers are 1 and remain unchanged, that is, the weight matrix between the input layer and the hidden layer is the identity matrix I, so as to summarize the biometric data of the input breast tumor samples; The activation function of each neuron in the hidden layer constitutes a power function sequence. From top to bottom, the activation function of the neurons in the first hidden layer is h ₀ (z)=1, the second is h 1 (z)=z, and the third is h ₁ (z)=z. One is h ₂ (z)=z ² , ..., and so on, the nth one is h _n (z)=z ^n-1 , where z is an independent variable, so as to carry out preliminary cognition learning on samples; implicit The weight matrix W between the output layer and the output layer is randomly initialized and waits for training, which is equivalent to a specific analysis of the input sample data; the activation function g of each neuron in the output layer is a Softsign function, and the analysis results are converted by nonlinear mapping into prediction results. The expression of the Softsign function is as follows:

Suppose the input breast tumor sample is a matrix X, the weight matrix between the hidden layer and the output layer is W(k) after learning the sample for the kth time, and the dynamic convergence differential neural network output Y(k) is powered by the power, that is, the intelligent doctor’s input to the sample The prediction and judgment results of can be deduced by the following formula:

Y(k)=g(H(XI)W(k)) (2)

in:

Let the actual attribution type of breast tumor samples be the expected value matrix Then /> Firstly, the training error ε(k) after the kth learning sample is calculated, and the training error is equivalent to the total deviation between the prediction and judgment results of multiple breast tumor samples and the actual attribution types of these samples.

Power excitation dynamic convergence difference neural network judgment result output Y(k) can get probability matrix P(k) through Softmax layer, and obtain the degree of attribution of each breast tumor sample to each type, and take the type corresponding to the maximum degree as the prediction critical result.

Assuming that the scales of matrix Y(k) and P(k) are both l×q, for the th The prediction and judgment result y _s corresponding to a breast tumor sample = [y _s1 y _s2 ... y _sq ], the row vector P _s of the s-th row in P(k) means that the corresponding s-th sample corresponds to each type of The degree of attribution is obtained by the following formula:

At the same time, as the actual class label of the sample, the expected value matrix The scale is set to l×q, and the elements inside are composed of -1 and 1. Therefore, it is necessary to change the value of -1 to 0 through the coding layer, and the rest of the elements remain unchanged, thus obtaining the coding matrix L, which is actually a coding method for the actual attribution type of the breast tumor sample. Therefore, the training error ε(k) can be obtained through the cross-entropy loss function formula:

Among them, L _sr and P _sr represent the s-th row in L and P(k) respectively. column elements.

Set the training error threshold to ε', so as to check the degree of intelligent doctors learning breast tumor samples. If ε(k)<ε', stop iteratively solving the weight matrix W between the hidden layer and the output layer, and end the learning process of the intelligent doctor on the samples. Otherwise, E(k+1)=E(k)-αΦ(E(k)) can be deduced from the expression of the neural dynamics method, that is, the dynamic convergence of power excitation after the intelligent doctor's (k+1)th learning sample is obtained The result of subtracting the expected value from the output of the differential neural network. On this basis, it is necessary to obtain the weight matrix between the hidden layer and the output layer as W(k+1). Analogy formula (2) and It can be obtained that E(k+1) and Y(k+1) satisfy the following equations at the same time:

Y(k+1)=g(H(XI)W(k+1)) (7)

The solution expression of W(k+1) can be obtained from formula (6) and formula (7):

Further, the relationship between W(k+1) and W(k) after the (k+1)th learning sample can be obtained. W(k+1) iterative solution expression is:

In formula (8) and formula (9), (H(XI)) ⁺ represents the Moore-Penrose pseudoinverse of matrix H(XI); the function In theory, it should represent the inverse function of the function g. However, g here is a Softsign function (as shown in formula (1)), and does not have an inverse function. Therefore, the formula (1) is segmented according to the value range of the independent variable z, and its inverse function is calculated for each segment, and then the inverse functions of each value range are spliced together and approximated.

therefore The expression is:

As shown in Figure 4, for the integrated model involved in the integrated model building block and comprehensive judgment module in Figure 2, the trained three-power-inspired dynamic convergence difference neural network is put into the integrated framework and the model integration process is performed, and the setting is based on The majority voting rule is used for comprehensive judgment of unknown breast tumor sample types. That is, multiple intelligent doctors who use different methods to study breast tumor samples gather together to analyze and judge unknown types of breast tumor samples, and the prediction results obtained will be used for comprehensive analysis. In the process of comprehensive analysis, the three smart doctors first predict the type of breast tumor samples. After all the smart doctors complete the predictions, the integrated framework adopts a voting principle based on the minority obeying the majority for these prediction results, and comprehensively judges the final type of the breast tumor sample. For example, for an unknown type of breast tumor sample in Figure 4, if the first intelligent doctor predicts a malignant tumor and the other two intelligent doctors predict a benign tumor, the integrated model finally judges that the test sample belongs to a benign tumor.

This embodiment also provides a storage medium, which can be a storage medium such as ROM, RAM, magnetic disk, optical disk, etc., and the storage medium stores one or more programs, and when the programs are executed by the processor, the above-mentioned mammary gland tumor types can be realized. recognition methods.

This embodiment also provides a computing device, and the computing device may be a desktop computer, a notebook computer, a smart phone, a PDA handheld terminal, a tablet computer or other terminal devices with a display function, and the computing device includes a processing One or more programs are stored in the memory, and when the processor executes the programs stored in the memory, the above-mentioned breast tumor type identification method is realized.

The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above-mentioned embodiment, and any other changes, modifications, substitutions, combinations, Simplifications should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims (7)

1. A method for identifying a breast tumor type, comprising the steps of:

inputting biometric data of the sample;

constructing a power excitation dynamic convergence differential neural network, setting the number of input neurons as the biological characteristic number of a single breast tumor sample, wherein the power excitation dynamic convergence differential neural network comprises an input layer, an hidden layer and an output layer, setting all connection weight components between the input layer and the hidden layer and keeping unchanged, and randomly initializing all connection weight components between the hidden layer and the output layer;

training a plurality of different power excitation dynamic convergence differential neural network models by adopting a plurality of different types of mapping functions;

the specific training method adopts the expression:

E(k+1)-E(k)＝-αΦ(E(k)),α＞0

wherein E (k) represents the result of subtracting the expected value from the output of the power-excited dynamic convergence differential neural network after the kth learning sample, alpha represents the thermodynamic coefficient, and phi represents the mapping function;

placing the trained multiple power excitation dynamic convergence differential neural network into an integration framework, carrying out model integration treatment, carrying out type prediction on biological characteristic data of an input sample, synthesizing multiple type prediction results, and obtaining a final breast tumor type recognition result by adopting a voting principle based on minority compliance;

the specific implementation steps of the model integration processing include:

let the input breast tumor sample be matrix X, the weight matrix between the hidden layer and the output layer after the kth learning sample be W (k), power excitation dynamic convergence differential neural network output Y (k):

Y(k)＝g(H(XI)W(k))

wherein:

setting the actual attribution type of the breast tumor sample into an expected value matrixThen->

The power excitation dynamic convergence difference neural network judgment result output Y (k) is subjected to a Softmax layer to obtain a probability matrix P (k), the attribution degree of each breast tumor sample to each type is obtained, and the type corresponding to the maximum degree is taken as a prediction judgment result.

2. A breast tumor type recognition system, comprising: the system comprises a biological characteristic input module, a neural network construction module, a neural network training module, an integrated model construction module and a comprehensive judgment module;

the biological characteristic input module is used for inputting biological characteristic data of a sample;

the neural network construction module is used for constructing a power excitation dynamic convergence differential neural network, and setting the number of input neurons as the number of biological characteristics of a single breast tumor sample;

the power excitation dynamic convergence differential neural network comprises an input layer, an implicit layer and an output layer, wherein each connection weight component between the input layer and the implicit layer is set and kept unchanged, and each connection weight component between the implicit layer and the output layer is randomly initialized;

the neural network training module is used for training a plurality of different power excitation dynamic convergence differential neural network models by adopting a plurality of different types of mapping functions;

the specific training method of the neural network training module adopts the following expression:

E(k+1)-E(k)＝-αΦ(E(k)),α＞0

wherein E (k) represents the result of subtracting the expected value from the output of the power-excited dynamic convergence differential neural network after the kth learning sample, alpha represents the thermodynamic coefficient, and phi represents the mapping function;

the integrated model construction module is used for putting the trained multiple power excitation dynamic convergence difference neural networks into an integrated framework and carrying out model integration treatment to construct an integrated model, and the integrated model is used for carrying out type prediction on biological characteristic data of an input sample;

the integrated model specifically executing steps comprise:

let the input breast tumor sample be matrix X, the weight matrix between the hidden layer and the output layer after the kth learning sample be W (k), power excitation dynamic convergence differential neural network output Y (k):

Y(k)＝g(H(XI)W(k))

wherein:

H(z)＝{h ₀ (z)；h ₁ (z)；h ₂ (z)…h _n-1 (z)}

＝{1；z；z ² …z ^n-1 }

setting the actual attribution type of the breast tumor sample into an expected value matrixThen->

The power excitation dynamic convergence difference neural network judgment result output Y (k) obtains a probability matrix P (k) through a Softmax layer, obtains the attribution degree of each breast tumor sample for each type, and takes the type corresponding to the maximum degree as a prediction judgment result;

the comprehensive judgment module is used for integrating a plurality of type prediction results and obtaining a final breast tumor type recognition result by adopting a voting principle based on a minority compliance majority.

3. The breast tumor type recognition system of claim 2, wherein the mapping function employs three of a linear mapping function, a Sin mapping function, and a Sinh mapping function.

4. The breast tumor type recognition system of claim 2, wherein the excitation functions of the neurons of the hidden layer form a power function sequence and the excitation functions of the neurons of the output layer employ Softsign functions.

5. The breast tumor type recognition system of claim 2, wherein the neural network training module derives a training error epsilon (k) using a cross entropy loss function formula, specifically calculated as:

wherein L is _sr And P _sr Respectively representing the r elements of the s th row and the r th column in L and P (k);

setting a training error threshold value as epsilon ', and stopping iterative solution of a weight matrix W between an implicit layer and an output layer if epsilon (k) is less than epsilon';

solving a weight matrix between the hidden layer and the output layer to be W (k+1), wherein the weight matrix is expressed as:

wherein (H (XI)) ⁺ Moore-Penrose pseudo-inverse, function representing matrix H (XI)Representing the inverse of function g.

6. A storage medium storing a program which, when executed by a processor, implements the breast tumor type recognition method according to claim 1.

7. A computing device comprising a processor and a memory for storing a program executable by the processor, wherein the processor, when executing the program stored in the memory, implements the breast tumor type identification method of claim 1.